Fraud Type Decomposition and the Observation-Mechanism Taxonomy:
Class-Specific Detection Limits in Payment Networks
May 29, 2026
Fraud detection models in payment networks train on chargeback labels or fraud reports by issuers — but “fraud” is not a single phenomenon. A stolen credit card, a cardholder who disputes a legitimate purchase, a synthetic identity building credit before a coordinated bust-out, a merchant laundering transactions through a phantom storefront, and a victim wiring money to a romance scammer all produce different labels through fundamentally different observation pipelines. Treating these as a homogeneous population — as all existing fraud detection frameworks do — is a provable source of statistical inefficiency.
We show that the payment industry’s 30+ fraud types collapse into exactly five observation-mechanism classes, each defined by a structurally distinct censorship pipeline. This classification is minimal (no two classes can be merged) and complete (every fraud type maps to exactly one class). We prove that estimating fraud rates separately by class and aggregating strictly dominates pooled estimation, with the efficiency gap quantified as a closed-form Jensen penalty.
For each class, we derive the binding theoretical constraint. For adversarial labeling (Class 2), the false-positive corruption rate is the Nash equilibrium of a dispute game between the cardholder and the issuer. For deferred observability (Class 3), the observation propensity is exactly zero during the bust-out build-up phase, and no account-level classifier exceeds random guessing — detection requires cross-account graph signals. For acquirer-observed fraud (Class 4), the detection floor is an order of magnitude higher than for standard card fraud. For victim-authorized fraud (Class 5), the Bayes-optimal transaction-level classifier achieves \(\mathop{\mathrm{AUC}}= 0.5\) when fraud and legitimate feature distributions coincide, and detection requires auxiliary signals outside the payment data.
Every supervised fraud detection model in payment networks trains on labels derived from the chargeback process. A cardholder disputes a transaction, the issuing bank files a chargeback, and the resulting record serves as a “fraud” label in the training data. Transactions without chargebacks are implicitly labeled “legitimate.”
This labeling process is impaired in well-documented ways. Declined transactions generate no labels (authorization censorship). Not all fraud is reported by cardholders or issuers (reporting censorship). Chargebacks arrive weeks or months after the transaction (delay censorship). And some labels are wrong — the cardholder may have fabricated the dispute, or the issuer may have miscoded the fraud type (label corruption).
Recent work has formalized these impairments. Dhama [1] derives minimax lower bounds on detection performance as a function of authorization, reporting, delay, and corruption rates, showing that the four impairments impose a fundamental floor on accuracy. Dhama [2] constructs the Sequential Triply Robust (STR) estimator, which corrects for all four impairments simultaneously and achieves the semiparametric efficiency bound — no estimator can have lower asymptotic variance.
These results — and indeed all existing theoretical treatments of the fraud label problem — share a common assumption: fraud is treated as a single binary variable. A transaction is either fraud (\(Y^*=1\)) or legitimate (\(Y^*=0\)), and the observation pipeline acts on all fraud uniformly through a single set of propensities \((e_0, r_0, p_0)\) and corruption rates \((\varepsilon_{10}, \varepsilon_{01})\).
This assumption is false. “Fraud” in payment networks encompasses fundamentally different phenomena — each with its own perpetrator, its own label-generating process, and its own observation pipeline. This paper shows that the homogeneity assumption is not merely a simplification — it is a provable source of inefficiency whose cost can be quantified exactly.
Example 1 (Why homogeneity fails). Consider three transactions flagged as potential fraud.
Transaction A is a card-not-present purchase using stolen credentials. The victim will eventually notice the unauthorized charge, dispute it with their issuer, and generate a chargeback label. The observation pipeline is the standard three-gate model: authorization (the network’s fraud model may decline it), reporting (the victim must notice and dispute), and delay (the chargeback takes weeks to process).
Transaction B is a legitimate purchase that the cardholder later disputes to obtain a fraudulent refund. Here the “victim” and the “fraudster” are the same person, and the chargeback label is not a noisy observation of truth — it is a strategic lie. The label is generated by the perpetrator, not the victim, and the corruption rate \(\varepsilon_{01}\) is not random noise but a deliberate strategic choice.
Transaction C is one of fifty small purchases made by a synthetic identity over six months of credit building before a coordinated bust-out. During the build-up phase, there is no fraud to report, no victim to file a dispute, and no label of any kind. The observation pipeline is not merely impaired — it is structurally inoperative.
Applying the STR with a single set of propensities averages across these three pipelines, producing an estimator that is optimal for none of them. The propensity \(r_0 = 0.70\) that is appropriate for Transaction A (victim-reported card theft) is meaningless for Transaction B (the “reporting rate” is the perpetrator’s strategic dispute probability) and impossible for Transaction C (there is nothing to report during build-up).
The payment industry classifies fraud into dozens of types based on perpetrator identity (third-party, first-party, synthetic), attack vector (card-not-present, account takeover, bust-out), and operational characteristics (e-commerce, point-of-sale, ATM). From the perspective of statistical estimation, however, the relevant taxonomy is not who commits the fraud or how, but how the fraud generates — or fails to generate — an observable label.
We show that the industry’s extensive fraud taxonomy collapses into exactly five observation-mechanism classes, each defined by a structurally distinct censorship pipeline. Within each class, fraud types share the same pipeline topology and differ only in propensity magnitudes. Across classes, the pipeline topology, the identity of the labeling agent, and the binding constraints on identification differ fundamentally, requiring class-specific estimation strategies.
This classification is derived from first principles: each fraud type is characterized by a structural signature — a 5-tuple describing the perpetrator–cardholder relationship, the label-generating agent, the authorization gate informativeness, the reporting gate structure, and the temporal observability. Two fraud types belong to the same class if and only if their structural signatures are identical.
Class 1: Victim-reported institutional pipeline. The perpetrator is external, the victim reports through the standard chargeback process, and the three-gate observation model applies directly. This class encompasses card-not-present fraud, counterfeit, lost/stolen, account takeover, and related types.
Class 2: Adversarial label generation. The perpetrator is the cardholder, who files a strategic false dispute. The corruption rate \(\varepsilon_{01}\) is not a statistical nuisance but an endogenous equilibrium of a dispute game. This class encompasses friendly fraud, chargeback abuse, refund abuse, and related types.
Class 3: Deferred observability. The fraud unfolds over a build-up phase during which the observation pipeline is structurally inoperative (\(r_0=0\), \(p_0=0\)), followed by a catastrophic exploitation event. Labels arrive only at charge-off. This class encompasses bust-out fraud, synthetic identity fraud, and sleeper fraud.
Class 4: Acquirer-observed pipeline. The perpetrator is the merchant, and the label is generated by acquirer monitoring rather than cardholder disputes. The propensities depend on the acquirer’s monitoring infrastructure, not the issuer’s reporting practices. This class encompasses transaction laundering, factoring, phantom merchants, and merchant collusion.
Class 5: Victim-authorized with non-informative features. The victim voluntarily authorizes the payment under social engineering. The authorization gate is non-binding (\(e_0=1\)), and the transaction features carry no fraud signal (\(P(X\mid Y^*\!=\!1)\approx P(X\mid Y^*\!=\!0)\)). This class encompasses authorized push payment fraud, romance scams, investment scams, and impersonation scams.
We establish the following results.
Observation-mechanism taxonomy (Theorem 12). The industry’s 30+ fraud types collapse into five observation-mechanism classes. The classification is minimal (no two classes can be merged) and complete (every fraud type maps to exactly one class).
Decomposition dominance (Theorem 16). Estimating fraud rates separately by class and aggregating — the type-decomposed approach — strictly dominates pooled estimation in mean squared error whenever at least two classes have different observation pipelines.
Jensen penalty quantification (Theorem 18). The MSE gap equals the cross-class variance of inverse-propensity weights, which is substantial for realistic network parameters (approximately 36% efficiency waste in our numerical illustration).
Endogenous corruption (Theorem 31). For Class 2, the false-positive corruption rate \(\varepsilon_{01}\) is the Nash equilibrium of a dispute game between the cardholder and the issuer, increasing in transaction amount and decreasing in dispute friction and detection probability. Multiple equilibria are possible, with a discontinuous tipping point between low-fraud and high-fraud regimes.
Positivity failure (Theorem 40). For Class 3, the observation propensity \(q_0 = e_0 r_0 p_0 = 0\) exactly during the bust-out build-up phase. The semiparametric efficiency bound diverges — no label-based estimator can detect bust-out during build-up.
Individual-level insufficiency (Theorem 42). For Class 3, no account-level classifier exceeds \(\mathop{\mathrm{AUC}}= 0.5\) during build-up; detection requires cross-account graph signals whose power grows with ring size.
Acquirer-side efficiency bound (Theorem 47). For Class 4, the detection floor depends on the acquirer’s monitoring propensity \(r_{\mathrm{acq}}\) rather than the issuer’s reporting propensity, and is an order of magnitude higher.
Authorization degeneracy and feature non-informativeness (Theorems 52–54). For Class 5, the authorization correction vanishes when \(e_0=1\), and the Bayes-optimal classifier achieves \(\mathop{\mathrm{AUC}}= 0.5\) when fraud and legitimate feature distributions coincide. Detection requires auxiliary signals outside the transaction data.
This paper builds on the theoretical framework of [1], [2], which formalize the observation pipeline and construct the STR estimator for a homogeneous fraud population. The present paper removes the homogeneity assumption, showing that the pooled STR is provably suboptimal for the heterogeneous fraud mixture that exists in every real payment network, and that each observation-mechanism class demands its own detection strategy. The economic incentives that shape each class’s observation pipeline — and the mechanism design problem of optimizing information quality across actors — are natural extensions discussed in Section 11.
Section 2 reviews related work. Section 3 introduces the latent-type mixture model. Section 4 presents the observation-mechanism taxonomy and proves its completeness and minimality. Section 5 establishes the decomposition dominance theorem and quantifies the Jensen penalty. Sections 6–10 derive class-specific results. Section 11 discusses limitations, practical type identification, and future work. Section 12 concludes.
Dhama [1] proves minimax lower bounds on fraud detection under four information impairments and introduces the conditional delay penalty via Jensen’s inequality. Dhama [2] constructs the STR estimator, proves it achieves the semiparametric efficiency bound, and derives the optimal training delay. Both papers assume a homogeneous fraud population. The present paper removes this assumption and shows that decomposition by observation-mechanism class strictly improves estimation.
The payment industry maintains extensive fraud taxonomies organized by perpetrator (third-party, first-party, synthetic), attack vector (card-not-present, account takeover, bust-out), and operational category (e-commerce, point-of-sale, ATM). Industry standards such as Mastercard’s TC40 fraud type codes and Visa’s SAFE categories provide granular classification for operational purposes. These taxonomies are designed for investigation and prevention, not for statistical estimation. Our observation-mechanism taxonomy complements the industry taxonomy by organizing fraud types according to their censorship pipeline structure, which determines the appropriate estimation strategy.
The causal inference literature addresses treatment effect heterogeneity through conditional average treatment effects (CATE; [3]), causal forests [4], and meta-learners [5]. These methods decompose a heterogeneous population by covariate values. Our decomposition is fundamentally different: we decompose by observation-mechanism structure, not by covariate values. Two transactions with identical covariates may belong to different classes (e.g., a legitimate purchase and a friendly-fraud dispute of the same purchase), and the appropriate estimator differs between them.
The game-theoretic analysis of Class 2 (adversarial label generation) connects to the literature on strategic classification [6], where agents manipulate their features to obtain favorable predictions. In our setting, the agent (first-party fraudster) manipulates the label, not the features — filing a false dispute to create a fraudulent chargeback. This is closer to the strategic reporting literature in mechanism design [7], where agents misreport private information. The endogenous corruption rate \(\varepsilon_{01}\) is the equilibrium reporting strategy of the agent.
Chang and Wiens [8] address disparate censorship with an EM-based approach for a single selection gate. Malinsky et al. [9] derive efficiency bounds for nonmonotone MNAR data under a no-self-censoring assumption. Ha et al. [10] prove that unbiasedness under MNAR inevitably leads to unbounded variance when propensity scores approach zero. Our Class 3 (deferred observability) provides a concrete, structurally motivated instance of Ha et al.’s impossibility: during bust-out build-up, the propensity is exactly zero, not merely small.
The practitioner literature on bust-out detection emphasizes behavioral features (spending velocity, credit utilization trajectories) and network analysis (ring detection, shared-attribute linking). To our knowledge, no prior work formalizes the impossibility of individual-level detection during the build-up phase or derives the information-theoretic necessity of cross-account signals. Our Theorem 42 provides this formalization.
Authorized push payment (APP) fraud and social engineering scams have received increasing attention from regulators (UK Payment Systems Regulator, 2023) and industry bodies. Detection approaches rely on behavioral biometrics, session analysis, and recipient risk scoring — all signals outside the standard transaction feature set. Our Theorem 54 provides the theoretical justification: when the victim performs the transaction using their own device and behavioral patterns, the standard transaction features carry zero discriminative power.
We build on the framework of [2], extending it to accommodate a heterogeneous fraud population.
Definition 1 (Transaction). A transaction \(t\) is characterized by:
Observable features \(X_t \in \mathcal{X}\) (amount, merchant category, channel, location, time, etc.);
Issuer identity \(I_t \in \mathcal{I}\);
A latent true fraud indicator \(Y^*_t \in \{0,1\}\), where \(Y^*_t=1\) denotes fraud;
A latent fraud type \(Z_t \in \{0,1,\ldots,K\}\), where \(Z_t=0\) denotes legitimate and \(Z_t=k\) for \(k\geq 1\) denotes fraud of type \(k\).
The fraud indicator and type are related by \(Y^*_t = \mathbf{1}[Z_t \geq 1]\).
Remark 1 (Type vs.label). The fraud type \(Z_t\) is a finer partition than the fraud indicator \(Y^*_t\). Knowing \(Y^*_t=1\) tells you the transaction is fraudulent; knowing \(Z_t=k\) tells you what kind of fraud it is. The observation pipeline depends on \(Z_t\), not just on \(Y^*_t\). This is the source of the heterogeneity that this paper formalizes.
Definition 2 (Mixing probabilities). The type probability conditional on covariates is \[\label{eq:mixing} \pi_k(x) = \mathbb{P}(Z_t = k \mid X_t = x), \qquad k = 0, 1, \ldots, K,\tag{1}\] with \(\sum_{k=0}^{K}\pi_k(x) = 1\). The marginal type prevalence is \(\bar\pi_k = \mathbb{E}[\pi_k(X)]\).
Definition 3 (Type-specific fraud probability). For each fraud type \(k \geq 1\), define \[\label{eq:type95fraud} f_0^{(k)}(x) = \mathbb{P}(Y^*_t = 1 \mid X_t = x, Z_t = k) = 1,\tag{2}\] since conditioning on \(Z_t = k \geq 1\) implies \(Y^*_t = 1\) by definition. The population fraud rate decomposes as: \[\label{eq:fraud95decomp} \Psi = \mathbb{E}[Y^*] = \sum_{k=1}^{K} \bar\pi_k.\tag{3}\]
Remark 2 (Why the decomposition matters for estimation). Equation 3 is trivial as an identity. The decomposition becomes non-trivial when we recognize that each type \(k\) has a different observation mechanism. Estimating \(\bar\pi_k\) requires correcting for the type-\(k\)-specific censorship pipeline, which differs across classes. Estimating \(\Psi = \sum_k\bar\pi_k\) with a single pooled pipeline averages over these differences and incurs a provable penalty (Theorem 16).
Definition 4 (Type-specific propensities). For each fraud type \(k\) and transaction \(t\) with \(Z_t = k\), define the type-specific observation pipeline: \[\begin{align} e_0^{(k)}(H_0) &= \mathbb{P}(A_t = 1 \mid H_{0,t},\, Z_t = k) && \text{(authorization propensity)}, \tag{4}\\ r_0^{(k)}(H_1) &= \mathbb{P}(R_t = 1 \mid H_{1,t},\, A_t = 1, \, Z_t = k) && \text{(reporting propensity)}, \tag{5}\\ p_0^{(k)}(H_2) &= \mathbb{P}(M_t = 1 \mid H_{2,t},\, A_t R_t = 1,\, Z_t = k) && \text{(delay propensity)}, \tag{6}\\ \varepsilon_{10}^{(k)}(H_2) &= \mathbb{P}(\tilde{Y}_t = 0 \mid Y^*_t = 1,\, O_t = 1,\, H_{2,t},\, Z_t = k) && \text{(false-negative corruption)}, \tag{7}\\ \varepsilon_{01}^{(k)}(H_2) &= \mathbb{P}(\tilde{Y}_t = 1 \mid Y^*_t = 0,\, O_t = 1,\, H_{2,t},\, Z_t = k) && \text{(false-positive corruption)}, \tag{8} \end{align}\] where \(H_0, H_1, H_2\) are the stage histories defined in [2], and \(O_t = A_t R_t M_t\) is the observation indicator.
Definition 5 (Type-specific observation rate). The total observation propensity for type \(k\) is \[\label{eq:type95obs95rate} q_0^{(k)}(H_0) = e_0^{(k)}(H_0) \cdot r_0^{(k)}(H_1) \cdot p_0^{(k)}(H_2),\tag{9}\] and the type-specific corruption penalty is \[\label{eq:type95corruption} \gamma^{(k)}(H_2) = \bigl(1 - \varepsilon_{10}^{(k)}(H_2) - \varepsilon_{01}^{(k)}(H_2)\bigr)^2.\tag{10}\]
Example 2 (Why type-specific propensities differ). Consider two fraud types at the same merchant, same amount, same time of day:
Type 1 (card theft): A stolen card is used for an online purchase. The fraud model assigns a high risk score \(\to\) authorization propensity \(e_0^{(1)} = 0.60\) (40% chance of decline). The victim notices and reports within 30 days \(\to\) reporting propensity \(r_0^{(1)} = 0.75\). The chargeback arrives within 60 days \(\to\) delay propensity \(p_0^{(1)} = 0.85\). Total observation rate: \(q_0^{(1)} = 0.60 \times 0.75 \times 0.85 = 0.38\).
Type 2 (friendly fraud): The cardholder makes the purchase themselves, then disputes it. The fraud model sees a normal transaction \(\to\) \(e_0^{(2)} = 0.98\). The cardholder strategically files a dispute \(\to\) \(r_0^{(2)}\) is not an institutional propensity but a strategic choice. The chargeback arrives on normal schedule \(\to\) \(p_0^{(2)} = 0.85\). But the label is wrong: the cardholder is claiming fraud on a legitimate transaction \(\to\) \(\varepsilon_{01}^{(2)} = 0.30\).
The observation pipelines are structurally different: Type 1 has informative authorization censorship and truthful labels; Type 2 has no authorization censorship but adversarially corrupted labels. A pooled STR using average propensities is optimal for neither.
Definition 6 (Pooled estimand). The pooled fraud rate is the target of [2]: \[\label{eq:pooled} \Psi = \mathbb{E}[Y^*] = \sum_{k=1}^K \bar\pi_k.\tag{11}\]
Definition 7 (Type-specific estimands). The type-\(k\) fraud rate is: \[\label{eq:type95estimand} \Psi_k = \mathbb{P}(Z = k) = \bar\pi_k, \qquad k = 1, \ldots, K.\tag{12}\] The pooled estimand is the sum: \(\Psi = \sum_{k=1}^K \Psi_k\).
Definition 8 (Type-specific efficiency bound). Extending the efficiency bound of [2] to the type-conditional setting, the semiparametric efficiency bound for estimating \(\Psi_k\) is: \[\label{eq:type95eff95bound} \sigma^2_{\mathrm{eff}}(k) = \mathbb{E}\!\left[ \frac{f_0^{(k)}(X)(1-f_0^{(k)}(X))}{q_0^{(k)}(H_0)\,\gamma^{(k)}(H_2)} \;\bigg|\; Z=k \right] + \mathrm{Var}\!\bigl(f_0^{(k)}(X) \mid Z=k\bigr).\tag{13}\] Since \(f_0^{(k)}(x) = \mathbb{P}(Y^*=1 \mid X=x, Z=k) = 1\) for \(k\geq 1\) (fraud types are defined by \(Y^*=1\)), the first term simplifies. However, the conditional fraud probability \(\mathbb{P}(Z=k \mid X=x)\) — which is what we estimate in practice — is not identically 1. The appropriate efficiency bound for estimating \(\bar\pi_k\) is: \[\label{eq:type95eff95bound95v2} \sigma^2_{\mathrm{eff}}(k) = \mathbb{E}\!\left[ \frac{\pi_k(X)(1-\pi_k(X))}{q_0^{(k)}(H_0)\,\gamma^{(k)}(H_2)} \right],\tag{14}\] where the expectation is over the full population.
Remark 3 (Interpreting the type-specific bound). The bound 14 shows that estimating the prevalence of fraud type \(k\) depends on:
The base-rate variance \(\pi_k(1-\pi_k)\) — rare fraud types are harder to estimate;
The observation rate \(q_0^{(k)}\) — types with lower observation propensity have higher bounds;
The corruption penalty \(\gamma^{(k)}\) — types with higher corruption have higher bounds.
Crucially, different fraud types can have very different values of \(q_0^{(k)}\) and \(\gamma^{(k)}\), leading to efficiency bounds that differ by orders of magnitude (Example 3).
Example 3 (Efficiency bounds differ by orders of magnitude). Consider four fraud types with the following average observation parameters:
| Type | \(\bar\pi_k\) | \(\bar q_0^{(k)}\) | \(\bar\gamma^{(k)}\) | \(\frac{\bar\pi_k(1-\bar\pi_k)}{\bar q_0^{(k)}\bar\gamma^{(k)}}\) | Relative |
|---|---|---|---|---|---|
| Class 1 (card theft) | 0.005 | 0.50 | 0.81 | 0.012 | \(1\times\) |
| Class 2 (friendly fraud) | 0.003 | 0.33 | 0.25 | 0.036 | \(3\times\) |
| Class 3 (bust-out) | 0.001 | 0.02 | 0.90 | 0.056 | \(4.5\times\) |
| Class 4 (merchant fraud) | 0.002 | 0.07 | 0.64 | 0.045 | \(3.7\times\) |
The efficiency bound for bust-out estimation is 4.5\(\times\) higher than for card theft, despite bust-out being rarer. The bottleneck is the observation rate: \(\bar q_0^{(3)} = 0.02\) (almost no labels during build-up) vs.\(\bar q_0^{(1)} = 0.50\) (half of card-theft labels are observed). A pooled STR that averages these rates underestimates the difficulty of estimating bust-out and overestimates the difficulty of estimating card theft.
We inherit the following assumptions from [2], applied type-conditionally.
Assumption 1 (Type-specific ignorability). For each class \(k\) and each gate \(g \in \{A, R, M\}\), the gate decision is independent of the latent fraud state conditional on the stage history and the type: \[\begin{align} Y^* &\perp A \mid H_0,\, Z=k, \tag{15}\\ Y^* &\perp R \mid H_1,\, A=1,\, Z=k, \tag{16}\\ Y^* &\perp M \mid H_2,\, AR=1,\, Z=k. \tag{17} \end{align}\]
Assumption 2 (Type-specific positivity). For each class \(k\) where the STR is applicable (Classes 1, 2, and 4), there exist constants \(\delta_k > 0\) such that \[\label{eq:positivity} q_0^{(k)}(H_0) > \delta_k > 0 \qquad\text{a.s.}\tag{18}\] This assumption is explicitly violated for Class 3 during the build-up phase and partially violated for Class 5. These violations are not limitations of the framework — they are the defining structural features of these classes (Theorems 40 and 52).
Assumption 3 (Type observability in mature data). For transactions with fully matured labels (\(O=1\) and sufficient elapsed time), the fraud type \(Z_t\) is identifiable from the label characteristics: \[\label{eq:type95obs} Z_t \text{ is observable when } O_t = 1 \text{ and the dispute is fully resolved.}\tag{19}\]
Remark 4 (Justifying type observability). Assumption 3 is justified in practice by:
Chargeback reason codes: Mastercard and Visa assign reason codes (e.g., “4837 – no cardholder authorization” for third-party fraud, “4853 – cardholder dispute” for first-party) that distinguish fraud types.
Dispute resolution outcomes: Representment and arbitration outcomes reveal whether the original dispute was legitimate (third-party) or fraudulent (first-party).
Charge-off indicators: Bust-out fraud is identified by the charge-off event and the account’s credit history pattern.
Network monitoring programs: Merchant fraud is identified through acquirer monitoring programs (e.g., Mastercard’s MATCH list, Visa’s VMAS).
The assumption requires that type classification is accurate for mature, fully resolved transactions — not for real-time classification. Section 11 discusses the sensitivity to type misclassification.
This section derives the five observation-mechanism classes from first principles, proves that the classification is complete and minimal, and maps the industry’s fraud types to the five classes.
The goal is to partition the set of fraud types into classes such that (a) types within a class share the same STR functional form and (b) types across classes require structurally different estimators. The partition is determined by five structural questions about the observation pipeline.
Definition 9 (Structural signature). The structural signature of a fraud type \(k\) is the 5-tuple \[\label{eq:signature} \mathcal{S}(k) = \bigl( \mathrm{Q1}(k),\; \mathrm{Q2}(k),\; \mathrm{Q3}(k),\; \mathrm{Q4}(k),\; \mathrm{Q5}(k) \bigr),\tag{20}\] where:
Perpetrator–cardholder relationship. Who commits the fraud relative to the cardholder? \[\label{eq:Q1} \mathrm{Q1}(k) \in \{\mathrm{\small external},\; \mathrm{\small cardholder},\; \mathrm{\small merchant},\; \mathrm{\small fabricated}\}.\tag{21}\]
Label-generating agent. Who generates the observed label \(\tilde{Y}\)? \[\label{eq:Q2} \mathrm{Q2}(k) \in \{\mathrm{\small victim},\; \mathrm{\small perpetrator},\; \mathrm{\small acquirer},\; \mathrm{\small charge\text{-}off},\; \mathrm{\small none}\}.\tag{22}\]
Authorization gate informativeness. Does \(e_0^{(k)} < 1\) provide discriminative signal? \[\label{eq:Q3} \mathrm{Q3}(k) \in \{\mathrm{\small informative}\; (e_0^{(k)} < 1),\;\; \mathrm{\small degenerate}\; (e_0^{(k)} = 1)\}.\tag{23}\]
Reporting gate structure. Is the reporting propensity \(r_0^{(k)}\) an institutional process, a strategic decision, driven by a different actor, or structurally zero? \[\label{eq:Q4} \mathrm{Q4}(k) \in \{\mathrm{\small institutional},\; \mathrm{\small strategic},\; \mathrm{\small acquirer\text{-}driven},\; \mathrm{\small zero}\}.\tag{24}\]
Temporal observability. When does the fraud state \(Y^*\) become potentially observable? \[\label{eq:Q5} \mathrm{Q5}(k) \in \{\mathrm{\small immediate},\; \mathrm{\small delayed},\; \mathrm{\small catastrophic},\; \mathrm{\small never}\}.\tag{25}\]
Remark 5 (Why these five questions). Each question corresponds to a structural feature that changes the functional form of the efficient estimator:
Q1 determines whether the label is trustworthy (external perpetrator \(\to\) victim labels truthfully) or adversarial (cardholder perpetrator \(\to\) label is a strategic lie).
Q2 determines which actor’s behavior must be modeled to estimate the reporting propensity.
Q3 determines whether the authorization correction term in the STR is active or vanishes.
Q4 determines the structure of the reporting model: a standard propensity (institutional), a game equilibrium (strategic), a different actor’s decision (acquirer-driven), or a structural impossibility (zero).
Q5 determines whether the standard delay model applies or whether the temporal structure requires a fundamentally different approach.
A different answer to any single question changes the estimator’s functional form. This is why each question is necessary and no question is redundant.
Definition 10 (Observation-mechanism class). Two fraud types \(k\) and \(k'\) belong to the same observation-mechanism class if and only if they have the same structural signature: \[\label{eq:equiv} k \sim k' \iff \mathcal{S}(k) = \mathcal{S}(k').\tag{26}\]
We now define each class by its structural signature, formal observation model, and the estimation strategy it requires.
Definition 11 (Class 1). A fraud type \(k\) belongs to Class 1 if its structural signature is: \[\label{eq:sig95class1} \mathcal{S}^{(1)} = \bigl( \mathrm{\small external},\; \mathrm{\small victim},\; \mathrm{\small informative},\; \mathrm{\small institutional},\; \mathrm{\small delayed} \bigr).\tag{27}\]
Observation model. The observation pipeline is exactly the three-gate model of [2]: \[\label{eq:obs95class1} O^{(1)}_t = A_t \cdot R_t \cdot M_t,\tag{28}\] where \(A_t \sim \mathrm{Bernoulli}(e_0^{(1)}(H_0))\), \(R_t \sim \mathrm{Bernoulli}(r_0^{(1)}(H_1))\), \(M_t \sim \mathrm{Bernoulli}(p_0^{(1)}(H_2))\), with all three propensities bounded away from zero (Assumption 2).
Estimation strategy. The STR of [2] applies without modification, using type-1-specific propensity and outcome models.
Defining characteristics.
The perpetrator is an external criminal unknown to the cardholder.
The victim (cardholder) generates the label by reporting unauthorized transactions to the issuer.
The authorization gate is informative: the network’s fraud model assigns elevated risk scores to stolen-card transactions, creating non-trivial selection bias (\(e_0^{(1)} < 1\)).
The reporting gate is institutional: it depends on victim awareness, issuer dispute processes, and operational efficiency — not on strategic behavior by the perpetrator.
Labels arrive with delay but within the standard chargeback window (typically 30–120 days).
Example 4 (Card-not-present fraud: a Class 1 instance). A criminal obtains a credit card number from a data breach and uses it for an online purchase. The network’s fraud model evaluates the transaction in real time: the shipping address differs from the billing address, the device fingerprint is new, and the merchant category is high-risk. The model assigns a risk score of 85/100, and the authorization propensity is \(e_0^{(1)} = 0.65\) — there is a 35% chance of decline.
If approved, the legitimate cardholder notices the unauthorized charge on their next statement (typically within 15 days) and calls the issuer to dispute it. The reporting propensity is \(r_0^{(1)} = 0.80\) — most victims notice and report, but some do not (small amounts, infrequent statement review, elderly cardholders).
The issuer processes the dispute and files a chargeback. The delay propensity is \(p_0^{(1)} = 0.85\) — most chargebacks arrive within the training window, but some are delayed by processing backlogs.
The corruption rates are low: \(\varepsilon_{10}^{(1)} = 0.03\) (issuer occasionally miscodes a fraud as legitimate) and \(\varepsilon_{01}^{(1)} = 0.02\) (rare cases where a legitimate transaction is mistakenly disputed).
Definition 12 (Class 2). A fraud type \(k\) belongs to Class 2 if its structural signature is: \[\label{eq:sig95class2} \mathcal{S}^{(2)} = \bigl( \mathrm{\small cardholder},\; \mathrm{\small perpetrator},\; \mathrm{\small degenerate},\; \mathrm{\small strategic},\; \mathrm{\small delayed} \bigr).\tag{29}\]
Observation model. The authorization gate is non-binding because the cardholder authorized the transaction themselves: \[\label{eq:obs95class2} O^{(2)}_t = 1 \cdot D_t \cdot M_t,\tag{30}\] where \(A_t = 1\) always (the cardholder wanted the transaction approved), \(D_t \in \{0,1\}\) is the cardholder’s strategic dispute decision, and \(M_t \sim \mathrm{Bernoulli}(p_0^{(2)}(H_2))\).
The critical structural difference: the “reporting gate” \(D_t\) is not an institutional propensity but a strategic decision by the perpetrator. The cardholder files a dispute (and thereby generates a label) if the expected payoff exceeds the cost: \[\label{eq:dispute95decision} D_t = \mathbf{1}\!\bigl[ \alpha(X_t) \cdot a_t - c - \beta(X_t) \cdot p + \xi_t > 0 \bigr],\tag{31}\] where \(\alpha(x)\) is the chargeback success probability, \(a_t\) is the transaction amount, \(c\) is the dispute filing cost, \(\beta(x)\) is the probability of being identified as first-party fraud, \(p\) is the penalty for identification, and \(\xi_t\) is a random utility shock.
Estimation strategy. The STR requires modification: the corruption rate \(\varepsilon_{01}^{(2)}\) is endogenous (it equals the equilibrium dispute probability; see Theorem 31) and must be estimated from the dispute game’s structure, not from historical audit data alone.
Defining characteristics.
The perpetrator is the cardholder (or an authorized user on the account).
The perpetrator generates the label by filing a false dispute — the label is a strategic lie.
The authorization gate is degenerate: \(e_0^{(2)} \approx 1\) because the cardholder made the purchase intentionally.
The reporting gate is strategic: the dispute decision depends on the expected payoff, detection risk, and penalty, not on institutional processes.
The false-positive corruption rate \(\varepsilon_{01}^{(2)}\) is not a nuisance parameter — it is the primary phenomenon to be modeled.
Example 5 (Friendly fraud: a Class 2 instance). A cardholder purchases a $500 electronics item online, receives and keeps the item, then calls the issuer claiming “I did not make this purchase.” The issuer files a chargeback.
From the network’s perspective, this transaction has:
\(e_0^{(2)} = 0.99\) — the fraud model saw nothing suspicious (the cardholder’s own device, usual location, reasonable amount).
\(D_t = 1\) — the cardholder strategically decided to dispute. The expected payoff: $500 \(\times\) 0.85 (chargeback success rate) \(-\) $0 (no dispute cost for the cardholder) \(-\) 0.05 \(\times\) $2000 (5% chance of being identified and penalized) \(=\) $325 \(>\) 0. The dispute is profitable.
The observed label is \(\tilde{Y} = 1\) (“fraud”), but the true label is \(Y^* = 0\) (the transaction was legitimate). This is a false positive label generated by the perpetrator’s strategic behavior.
If the network trains on this label naively, it learns that “transactions with these features are fraudulent” — but the transaction was legitimate. The model is being poisoned by strategically generated labels.
Remark 6 (Class 2 inverts the corruption structure). In Class 1, the corruption concern is primarily \(\varepsilon_{10}\) (the issuer misses real fraud). In Class 2, the corruption concern is primarily \(\varepsilon_{01}\) (the perpetrator creates false fraud labels). The label corruption is not random noise — it is adversarial signal injection. This inversion requires a fundamentally different treatment of the corruption correction in the STR (Section 7).
Definition 13 (Class 3). A fraud type \(k\) belongs to Class 3 if its structural signature is: \[\label{eq:sig95class3} \mathcal{S}^{(3)} = \bigl( \mathrm{\small fabricated},\; \mathrm{\small charge\text{-}off},\; \mathrm{\small degenerate},\; \mathrm{\small zero},\; \mathrm{\small catastrophic} \bigr).\tag{32}\]
Observation model. Class 3 fraud has a two-phase temporal structure:
Build-up phase (\(t < \tau\), where \(\tau\) is the bust-out time): \[\label{eq:obs95class395buildup} O^{(3)}_t = A_t \cdot 0 \cdot 0 = 0 \qquad\text{(structurally zero)}.\tag{33}\] During build-up, the account behaves legitimately. There is no fraud to report (\(R_t = 0\) with certainty because no fraud has occurred yet), and no label to mature (\(M_t\) is undefined). The observation propensity is exactly zero — not small, not approximately zero, but identically zero.
Exploitation phase (\(t \geq \tau\)): \[\label{eq:obs95class395exploit} O^{(3)}_t = A_t \cdot R_t \cdot M_t,\tag{34}\] where the pipeline resembles Class 1 but with a sudden burst of high-value transactions. Labels eventually arrive through the charge-off process (typically 120–180 days after delinquency).
Estimation strategy. The STR is undefined during the build-up phase because positivity fails (\(q_0^{(3)} = 0\)). The estimand shifts from the transaction level to the account level: \[\label{eq:class395estimand} \theta_a = \mathbf{1}[\text{account } a \text{ will eventually bust out}].\tag{35}\] Detection requires cross-account graph signals, not individual transaction correction (Theorem 42).
Defining characteristics.
The perpetrator operates through a fabricated or compromised identity that behaves legitimately during build-up.
Labels arrive only at the catastrophic event (charge-off), not through the standard chargeback process.
The authorization gate is degenerate during build-up: \(e_0^{(3)} \approx 1\) because the transactions look normal.
The reporting gate is structurally zero during build-up: there is no fraud to report.
The temporal structure is two-phase: a silent build-up followed by a sudden exploitation burst.
Example 6 (Synthetic identity bust-out: a Class 3 instance). A fraud ring creates a synthetic identity — a fictitious person constructed from a combination of a real Social Security number (belonging to a child, deceased individual, or immigrant) and fabricated biographical details. The ring applies for a credit card and is approved with a $2,000 limit.
Months 1–6 (build-up): The synthetic identity makes small, regular purchases ($30–$80 per transaction, 3–5 per month) and pays the balance in full each month. The account demonstrates perfect payment behavior. The issuer increases the credit limit to $15,000.
During this phase:
Every transaction is legitimate (\(Y^*_t = 0\)).
The fraud model assigns low risk scores (\(e_0^{(3)} = 0.99\)).
There is nothing to report (\(r_0^{(3)} = 0\)).
No labels exist (\(O^{(3)}_t = 0\) for every \(t\)).
The account is indistinguishable from a legitimate new customer.
Month 7 (exploitation): The ring maxes out all credit lines simultaneously: $15,000 in purchases across 40 transactions in 72 hours, at high-resale-value merchants (electronics, gift cards, jewelry). The ring abandons the identity. The issuer charges off the account after 120 days of delinquency.
During exploitation, labels eventually arrive through the charge-off process — but 6 months of training data from the build-up phase contain zero signal. Any model trained on this data sees only a normal customer who suddenly went delinquent.
Remark 7 (Why Class 3 is not a special case of Class 1). One might argue that bust-out fraud “eventually” generates labels (at charge-off) and therefore belongs to Class 1 with very long delay (\(p_0^{(3)}\) small but positive). This conflation is incorrect for two reasons:
During the build-up phase, \(r_0^{(3)} = 0\) exactly, not approximately. There is no fraud to report because the transactions are legitimate. The positivity assumption (Assumption 2) fails structurally, not due to data limitations.
The estimand is different. In Class 1, the target is the transaction-level fraud probability \(f_0^{(1)}(x) = \mathbb{P}(Y^*=1 \mid X=x, Z=1)\). In Class 3, the target is the account-level bust-out probability \(\theta_a = \mathbb{P}(\text{bust-out} \mid \text{account history } \mathcal{H}_a)\) — a fundamentally different quantity that requires account-level, not transaction-level, inference.
Definition 14 (Class 4). A fraud type \(k\) belongs to Class 4 if its structural signature is: \[\label{eq:sig95class4} \mathcal{S}^{(4)} = \bigl( \mathrm{\small merchant},\; \mathrm{\small acquirer},\; \mathrm{\small degenerate},\; \mathrm{\small acquirer\text{-}driven},\; \mathrm{\small delayed} \bigr).\tag{36}\]
Observation model. The observation pipeline runs through the acquirer, not the issuer: \[\label{eq:obs95class4} O^{(4)}_t = 1 \cdot R_{\mathrm{acq},t} \cdot M_{\mathrm{acq},t},\tag{37}\] where \(A_t = 1\) always (the merchant controls the terminal and approves the transaction), \(R_{\mathrm{acq},t} \sim \mathrm{Bernoulli}( r_{\mathrm{acq}}^{(4)}(H_{\mathrm{merch}}))\) is the acquirer’s monitoring and reporting propensity, and \(M_{\mathrm{acq},t} \sim \mathrm{Bernoulli}( p_{\mathrm{acq}}^{(4)}(H_{\mathrm{merch}}))\) is the acquirer’s investigation delay.
Note that the conditioning variables \(H_{\mathrm{merch}}\) are merchant-level features (chargeback ratio, transaction volume, merchant category code, acquirer identity, time since merchant onboarding), not cardholder-level features as in Classes 1–2.
Estimation strategy. The STR of [2] applies with re-parameterization: replace the issuer’s reporting propensity \(r_0\) with the acquirer’s monitoring propensity \(r_{\mathrm{acq}}\), and replace the issuer’s delay \(p_0\) with the acquirer’s investigation delay \(p_{\mathrm{acq}}\). The authorization correction vanishes (\(e_0^{(4)} = 1\)).
Defining characteristics.
The perpetrator is the merchant (or a collusive merchant–cardholder pair).
The label is generated by the acquirer’s monitoring process, network monitoring programs (e.g., Mastercard’s MATCH, Visa’s VMAS), or law enforcement — not by cardholder disputes.
The authorization gate is degenerate: the merchant controls the terminal and always “approves” the transaction.
The reporting gate is acquirer-driven: it depends on the acquirer’s monitoring sophistication, the merchant’s chargeback ratio exceeding program thresholds, and regulatory triggers.
Labels arrive with significant delay: acquirer investigations take months, and network monitoring programs have quarterly review cycles.
Example 7 (Transaction laundering: a Class 4 instance). A criminal organization registers a legitimate-appearing online storefront (“Premium Electronics LLC”) through a payment service provider. The storefront has a professional website, valid business registration, and a small volume of real sales.
Simultaneously, the organization operates an illegal online gambling site. When a gambler deposits $500, the organization processes a $500 charge through Premium Electronics LLC as a “consumer electronics purchase.” The cardholder sees “Premium Electronics” on their statement and does not dispute it (they wanted to deposit $500 for gambling).
From the network’s perspective:
\(e_0^{(4)} = 1\) — the merchant submitted and approved the transaction.
The cardholder will never dispute — they authorized the transaction. The standard chargeback pipeline generates no label.
The only path to detection is through the acquirer monitoring the merchant’s transaction patterns: unusual transaction-amount distributions, high ratios of round-number amounts, geographic dispersion of cardholders inconsistent with the stated business type.
\(r_{\mathrm{acq}}^{(4)} = 0.12\) — only 12% of acquirers have monitoring sophisticated enough to detect this pattern.
\(p_{\mathrm{acq}}^{(4)} = 0.35\) — even when detected, investigations take 3–9 months.
The total observation rate is \(q_0^{(4)} = 1 \times 0.12 \times 0.35 = 0.042\) — labels are observed for fewer than 5% of laundering transactions.
Remark 8 (The unit of analysis shifts to the merchant). In Classes 1–2, the unit of analysis is the transaction (or the cardholder account). In Class 4, the unit of analysis is the merchant. The fraud state \(Y^*\) is a property of the merchant’s behavior pattern, not of any individual transaction. A single laundering transaction is indistinguishable from a legitimate purchase; the fraud signal emerges from the merchant’s aggregate transaction profile. This parallels the account-level estimand shift in Class 3 — but the entity is the merchant rather than the cardholder account.
Definition 15 (Class 5). A fraud type \(k\) belongs to Class 5 if its structural signature is: \[\label{eq:sig95class5} \mathcal{S}^{(5)} = \bigl( \mathrm{\small external},\; \mathrm{\small victim},\; \mathrm{\small degenerate},\; \mathrm{\small institutional},\; \mathrm{\small delayed} \bigr).\tag{38}\]
Observation model. The observation pipeline superficially resembles Class 1, but with two critical degeneracies: \[\label{eq:obs95class5} O^{(5)}_t = \underbrace{1}_{e_0^{(5)}=1} \cdot\; R_{\mathrm{victim},t} \cdot M_t,\tag{39}\] where \(A_t = 1\) always (the victim authorized the transaction voluntarily) and \(R_{\mathrm{victim},t} \sim \mathrm{Bernoulli}( r_{\mathrm{victim}}^{(5)}(H_1))\) depends on:
Victim awareness: the victim must realize they were scammed (may take days to months);
Shame and embarrassment: particularly for romance and investment scams, victims may not report;
Perceived futility: for authorized transactions, chargeback rights are limited, reducing the incentive to report.
Feature non-informativeness. The second critical degeneracy is: \[\label{eq:feature95noninf} P(X_t \mid Y^*_t = 1, Z_t = 5) \approx P(X_t \mid Y^*_t = 0).\tag{40}\] The transaction features — amount, merchant, time, device, location, behavioral biometrics — are non-informative because the victim performs the transaction using their own device, from their own location, in their normal behavioral pattern. The fraudster manipulates the victim, not the payment instrument.
Estimation strategy. The STR’s authorization correction vanishes (Theorem 52). The remaining two-gate estimator depends on \(r_{\mathrm{victim}}^{(5)}\) and \(p_0\), both of which are low. Moreover, the feature non-informativeness condition 40 means that the outcome model \(\mu_0(H_0) \approx \bar\pi_5\) (a constant), eliminating the variance reduction that the outcome model normally provides. Detection requires auxiliary signals outside the transaction feature set (Theorem 54).
Defining characteristics.
The perpetrator is external but manipulates the victim through social engineering, not the payment system.
The victim generates the label (eventually), but reporting is delayed and incomplete.
The authorization gate is degenerate: \(e_0^{(5)} = 1\) exactly, because the victim voluntarily authorized the payment.
The reporting gate is institutional but severely impaired by delayed awareness and psychological barriers.
The transaction features carry no discriminative signal — the fraud signal, if it exists, resides in the communication channel (phone calls, messages, session behavior), not in the payment data.
Example 8 (Romance scam: a Class 5 instance). A victim meets a scammer on a dating app. Over three months, the scammer builds a relationship and fabricates a financial emergency (“I need $5,000 for surgery in another country”). The victim willingly transfers $5,000 via bank transfer or card payment to the scammer’s account.
From the network’s perspective:
\(e_0^{(5)} = 1\) — the victim authorized the transaction from their own phone, at their home address, during normal waking hours.
The transaction features are identical to a legitimate gift or charitable donation: correct device fingerprint, normal IP address, reasonable amount for a one-time payment, no velocity anomaly.
The fraud model assigns a risk score of 15/100 — lower than many legitimate transactions.
The victim does not realize they were scammed for weeks or months. When they do, shame prevents 40–60% from reporting. \(r_{\mathrm{victim}}^{(5)} \approx 0.25\).
Even when reported, the transaction was authorized — chargeback rights may not apply. The label may never arrive.
No feature in the transaction data — amount, merchant, time, device, location, behavioral biometric — distinguishes this transaction from a legitimate gift. The fraud signal is in the three months of manipulative communication that preceded the payment, not in the payment itself.
Remark 9 (Class 5 vs.Class 1: the subtle difference). Class 5’s structural signature 38 differs from Class 1’s 27 in only one dimension: Q3 is degenerate instead of informative. But this single difference has profound consequences:
The STR’s authorization correction term — which provides significant variance reduction in Class 1 — vanishes entirely in Class 5 (Theorem 52).
The feature non-informativeness condition 40 eliminates the outcome model’s predictive power, removing the second source of variance reduction.
The combination of degenerate authorization and non-informative features means that the transaction-level Bayes-optimal classifier achieves \(\mathop{\mathrm{AUC}}= 0.5\) (Theorem 54) — a result that does not hold for Class 1, where \(e_0 < 1\) and features are informative.
We now prove that every fraud type in payment networks maps to exactly one of the five classes.
Theorem 10 (Completeness). Let \(\mathcal{F}\) denote the set of all fraud types in payment networks. Every fraud type \(k \in \mathcal{F}\) has a well-defined structural signature \(\mathcal{S}(k)\) that matches exactly one of the five class signatures \(\mathcal{S}^{(1)}, \ldots, \mathcal{S}^{(5)}\).
Proof. We proceed by exhaustive enumeration. Table 1 maps every standard industry fraud type to its structural signature. For each type, the answers to Q1–Q5 are determined by the type’s operational definition:
Q1 is determined by the identity of the perpetrator relative to the cardholder.
Q2 is determined by which actor’s action creates the observed label.
Q3 is determined by whether the perpetrator’s transactions trigger elevated risk scores (\(\mathrm{\small informative}\)) or are indistinguishable from the cardholder’s normal behavior (\(\mathrm{\small degenerate}\)).
Q4 is determined by the institutional structure of the reporting process.
Q5 is determined by the temporal structure of label availability.
The five class signatures exhaust all realizable combinations. The remaining combinations in \(\{Q1\} \times \cdots \times \{Q5\}\) are either logically impossible (e.g., \(\mathrm{Q1} = \mathrm{\small external}\) with \(\mathrm{Q2} = \mathrm{\small perpetrator}\) would require the external criminal to file a dispute against themselves) or do not occur in payment networks (e.g., \(\mathrm{Q4} = \mathrm{\small zero}\) with \(\mathrm{Q5} = \mathrm{\small immediate}\) would require a fraud type with no reporting mechanism but immediate label availability).
Each fraud type maps to exactly one class because the answers to Q1–Q5 are uniquely determined by the type’s operational definition, and each realized signature matches exactly one class. ◻
| Fraud type | Class | Q1 | Q2 | Q3 | Q4 | Q5 |
|---|---|---|---|---|---|---|
| Card-not-present (CNP) | 1 | Ext | Vict | Info | Inst | Del |
| Counterfeit (skimming) | 1 | Ext | Vict | Info | Inst | Del |
| Lost/stolen card | 1 | Ext | Vict | Info | Inst | Del |
| Mail non-receipt | 1 | Ext | Vict | Info | Inst | Del |
| Account takeover | 1 | Ext | Vict | Info | Inst | Del |
| Card trapping | 1 | Ext | Vict | Info | Inst | Del |
| Card testing | 1 | Ext | Vict | Info | Inst | Del |
| Friendly fraud | 2 | Card | Perp | Deg | Strat | Del |
| Chargeback abuse | 2 | Card | Perp | Deg | Strat | Del |
| Refund abuse | 2 | Card | Perp | Deg | Strat | Del |
| Double-dipping | 2 | Card | Perp | Deg | Strat | Del |
| Family/authorized user | 2 | Card | Perp | Deg | Strat | Del |
| Bust-out (credit card) | 3 | Fab | C-off | Deg | Zero | Cat |
| Bust-out (DDA) | 3 | Fab | C-off | Deg | Zero | Cat |
| Synthetic identity | 3 | Fab | C-off | Deg | Zero | Cat |
| Sleeper fraud | 3 | Fab | C-off | Deg | Zero | Cat |
| Loan stacking | 3 | Fab | C-off | Deg | Zero | Cat |
| Transaction laundering | 4 | Merch | Acq | Deg | Acq | Del |
| Factoring | 4 | Merch | Acq | Deg | Acq | Del |
| Phantom merchant | 4 | Merch | Acq | Deg | Acq | Del |
| Merchant collusion | 4 | Merch | Acq | Deg | Acq | Del |
| Refund manipulation | 4 | Merch | Acq | Deg | Acq | Del |
| Merchant identity fraud | 4 | Merch | Acq | Deg | Acq | Del |
| APP fraud | 5 | Ext | Vict | Deg | Inst | Del |
| Romance scam | 5 | Ext | Vict | Deg | Inst | Del |
| Investment scam | 5 | Ext | Vict | Deg | Inst | Del |
| Tech support scam | 5 | Ext | Vict | Deg | Inst | Del |
| Invoice redirection | 5 | Ext | Vict | Deg | Inst | Del |
| CEO/BEC fraud | 5 | Ext | Vict | Deg | Inst | Del |
| Impersonation scam | 5 | Ext | Vict | Deg | Inst | Del |
| Purchase scam | 5 | Ext | Vict | Deg | Inst | Del |
Theorem 11 (Minimality). No two observation-mechanism classes can be merged without losing a structural distinction that requires a different estimation strategy. Formally, for every pair of classes \(c \neq c'\), there exists an aspect of the efficient estimator that differs between \(c\) and \(c'\).
Proof. We prove minimality by exhibiting, for each of the \(\binom{5}{2} = 10\) pairs of classes, a structural distinction that necessitates a different estimator.
Classes 1 vs.. In Class 1, \(\varepsilon_{01}^{(1)}\) is an exogenous nuisance parameter estimated from audit data. In Class 2, \(\varepsilon_{01}^{(2)}\) is the equilibrium of a strategic game (Theorem 31). The corruption correction in the STR requires a different estimation procedure: historical audit rates for Class 1 vs. game-theoretic equilibrium computation for Class 2.
Classes 1 vs.. In Class 1, the positivity assumption holds (\(q_0^{(1)} > \delta > 0\)). In Class 3, positivity fails structurally during build-up (\(q_0^{(3)} = 0\); Theorem 40). The STR is well-defined for Class 1 and undefined for Class 3. Moreover, the estimand differs: transaction-level \(f_0^{(1)}(x)\) for Class 1 vs.account-level \(\theta_a\) for Class 3.
Classes 1 vs.. In Class 1, the reporting propensity \(r_0^{(1)}\) models the issuer’s chargeback process, conditioned on cardholder-level features \(H_1\). In Class 4, the reporting propensity \(r_{\mathrm{acq}}^{(4)}\) models the acquirer’s monitoring process, conditioned on merchant-level features \(H_{\mathrm{merch}}\). The propensity model’s inputs, functional form, and the responsible institutional actor all differ.
Classes 1 vs.. In Class 1, \(e_0^{(1)} < 1\) and the authorization correction term in the STR is active, providing variance reduction. In Class 5, \(e_0^{(5)} = 1\) and the authorization correction vanishes (Theorem 52). Furthermore, in Class 1 the features are informative (\(P(X \mid Y^*\!=\!1) \neq P(X \mid Y^*\!=\!0)\)), while in Class 5 they are not (Theorem 54).
Classes 2 vs.. Class 2 has endogenous corruption (\(\varepsilon_{01}^{(2)}\) is strategic); Class 3 has structural positivity failure (\(q_0^{(3)} = 0\)). These require fundamentally different responses: game-theoretic corruption modeling for Class 2 vs.graph-based account-level detection for Class 3.
Classes 2 vs.. In Class 2, the adversary is the cardholder who manipulates the label. In Class 4, the adversary is the merchant who commits the fraud. The adversary’s action space, objective function, and the resulting equilibrium differ entirely.
Classes 2 vs.. In Class 2, the cardholder is the perpetrator and generates adversarial labels (\(\varepsilon_{01}\) is high and strategic). In Class 5, the cardholder is the victim and labels are truthful but scarce (\(\varepsilon_{01}^{(5)}\) is low; \(r_{\mathrm{victim}}^{(5)}\) is the bottleneck). The binding constraint differs: corruption for Class 2 vs. reporting for Class 5.
Classes 3 vs.. Class 3 has a two-phase temporal structure (build-up then exploitation) with a complete observation blackout during build-up. Class 4 has a single-phase structure with continuous but low-probability acquirer monitoring. The temporal dynamics and the detection strategy differ fundamentally.
Classes 3 vs.. The binding constraint in Class 3 is positivity failure (\(r_0 p_0 = 0\)). The binding constraint in Class 5 is feature non-informativeness (\(P(X \mid Y^*\!=\!1) = P(X \mid Y^*\!=\!0)\)). These are different obstructions requiring different solutions: graph signals for Class 3 vs. auxiliary (non-payment) features for Class 5.
Classes 4 vs.. In Class 4, the perpetrator is the merchant and the detection pipeline runs through acquirer monitoring of merchant-level patterns. In Class 5, the perpetrator is an external social engineer and the detection challenge is that the payment features carry no signal. The pipeline actors, feature sets, and estimation strategies are entirely disjoint. ◻
Combining the completeness and minimality results:
Theorem 12 (Observation-mechanism classification). Let \(\mathcal{F}\) denote the set of all fraud types in payment networks, and define the equivalence relation \(\sim\) on \(\mathcal{F}\) by \[\label{eq:equiv95relation} k \sim k' \iff \mathcal{S}(k) = \mathcal{S}(k'),\tag{41}\] where \(\mathcal{S}(k)\) is the structural signature (Definition 9). Then the quotient \(\mathcal{F}/{\sim}\) has exactly five equivalence classes:
Victim-reported institutional pipeline (Definition 11);
Adversarial label generation (Definition 12);
Deferred observability (Definition 13);
Acquirer-observed pipeline (Definition 14);
Victim-authorized with non-informative features (Definition 15).
This classification is:
Complete: every \(k \in \mathcal{F}\) maps to exactly one class (Theorem 10);
Minimal: no two classes can be merged without losing a structural distinction that changes the efficient estimator (Theorem 11);
Estimation-relevant: types within a class share the same STR functional form; types across classes require structurally different estimation strategies.
Proof. Completeness follows from Theorem 10 and the exhaustive mapping in Table 1. Minimality follows from Theorem 11 and the pairwise structural distinctions exhibited in its proof. Estimation relevance follows by construction: each class is defined by its structural signature, and each element of the signature corresponds to a feature of the efficient estimator (Remark 5). ◻
Remark 13 (Fraud types not covered). Two categories of illicit activity in payment networks fall outside the scope of this taxonomy:
Internal fraud (insider at issuer, acquirer, or processor). The observation mechanism runs through internal audit and compliance processes, which are outside the payment network’s transaction-level data. This requires institution-level, not network-level, analysis.
Money laundering (where no party is defrauded). The label source is regulatory (suspicious activity reports filed by compliance teams), not the chargeback process. The observation mechanism is entirely different from the transaction-level pipeline formalized here.
Both are important problems but require separate theoretical frameworks. We defer their treatment to future work.
Remark 14 (Transitional types). Some fraud types exhibit characteristics of multiple classes at different stages of their lifecycle. Account takeover (ATO) begins as a Class 5-like event (the criminal’s initial transactions may use the victim’s device and behavioral patterns, making features temporarily non-informative) but transitions to Class 1 once the victim detects the takeover and reports unauthorized transactions. For the formal classification, we assign ATO to Class 1 based on its dominant observation mechanism — the pipeline through which the majority of labels are generated. Section 11 discusses the sensitivity of results to this assignment.
This section proves the central meta-result of the paper: estimating fraud rates separately by observation-mechanism class and aggregating is strictly more efficient than estimating a single pooled fraud rate. We first state and prove the dominance theorem, then quantify the MSE gap as a closed-form Jensen penalty.
Definition 16 (Pooled STR). The pooled STR is the estimator of [2] applied to all transactions without regard to fraud type. It uses a single set of estimated propensities \((\hat{e}_0, \hat{r}_0, \hat{p}_0)\) and a single outcome model \(\hat{\mu}_0\): \[\label{eq:pooled95str} \hat{\Psi}_{\mathrm{pooled}} = \frac{1}{n}\sum_{t=1}^n \hat{\varphi}_t,\tag{42}\] where \(\hat{\varphi}_t\) is the corruption-corrected STR score of [2], computed with the pooled nuisance functions.
Definition 17 (Type-decomposed STR). The type-decomposed STR estimates each class separately and aggregates: \[\label{eq:decomp95str} \hat{\Psi}_{\mathrm{decomp}} = \sum_{k=1}^{K} \hat{\Psi}_k, \qquad\text{where}\quad \hat{\Psi}_k = \frac{1}{n}\sum_{t=1}^n \hat{\pi}_k(X_t)\,\hat{\varphi}_t^{(k)}.\tag{43}\] Here \(\hat{\pi}_k(X_t)\) is the estimated probability that transaction \(t\) belongs to class \(k\), and \(\hat{\varphi}_t^{(k)}\) is the STR score computed with class-\(k\)-specific nuisance functions \((\hat{e}_0^{(k)}, \hat{r}_0^{(k)}, \hat{p}_0^{(k)}, \hat{\mu}_0^{(k)})\).
Remark 15 (Practical construction of the decomposed STR). The type-decomposed STR requires two additional ingredients beyond the pooled STR:
Type membership probabilities \(\hat{\pi}_k(x)\). For transactions with mature, fully resolved labels (\(O_t = 1\)), the type \(Z_t\) is observed (Assumption 3) and \(\hat{\pi}_k\) can be estimated by a multiclass classifier trained on resolved transactions. For transactions without labels (\(O_t=0\)), \(\hat{\pi}_k(x)\) provides the soft assignment.
Class-specific nuisance functions. Each class requires its own propensity and outcome models, trained on the subset of transactions assigned to that class. For Classes 1 and 2, these are standard propensity models; for Class 4, the propensity models condition on merchant-level features.
The additional modeling effort is modest: one multiclass classifier for type assignment plus \(K\) sets of nuisance models. The theoretical gain, as we now show, is substantial.
Theorem 16 (Decomposition dominance). Let \(\hat{\Psi}_{\mathrm{pooled}}\) be the pooled STR (Definition 16) and \(\hat{\Psi}_{\mathrm{decomp}}\) be the type-decomposed STR (Definition 17), both estimated on the same sample of \(n\) transactions. Under Assumptions 1–3 and the regularity conditions of [2]: \[\label{eq:dominance} \sigma^2_{\mathrm{decomp}} \;\leq\; \sigma^2_{\mathrm{pooled}},\tag{44}\] where \(\sigma^2_{\mathrm{decomp}}\) and \(\sigma^2_{\mathrm{pooled}}\) are the semiparametric efficiency bounds for the decomposed and pooled estimators, respectively.
Equality holds if and only if the type-specific inverse-propensity weights are identical across all classes: \[\label{eq:equality95cond} \frac{1}{q_0^{(k)}(H_0)\,\gamma^{(k)}(H_2)} = \frac{1}{q_0^{(k')}(H_0)\,\gamma^{(k')}(H_2)} \qquad\text{a.s.\;for all } k, k'.\tag{45}\] In particular, strict inequality holds whenever at least two classes have different observation rates or corruption penalties.
Proof. The proof proceeds in four steps.
Step 1: Decompose the pooled efficiency bound.
The pooled efficiency bound from [2] is: \[\label{eq:pooled95bound} \sigma^2_{\mathrm{pooled}} = \mathbb{E}\!\left[ \frac{f_0(X)(1-f_0(X))}{q_0(H_0)\,\gamma(H_2)} \right] + \mathrm{Var}(f_0(X)),\tag{46}\] where \(f_0(x) = \mathbb{P}(Y^*=1 \mid X=x)\), \(q_0 = e_0 r_0 p_0\), and \(\gamma = (1-\varepsilon_{10}-\varepsilon_{01})^2\) are the marginal (type-averaged) propensity and corruption penalty.
Using the law of total expectation, condition on the fraud type \(Z\): \[\label{eq:pooled95decomp1} \mathbb{E}\!\left[\frac{f_0(1-f_0)}{q_0\,\gamma}\right] = \sum_{k=0}^{K} \bar\pi_k\; \mathbb{E}\!\left[ \frac{f_0(X)(1-f_0(X))}{q_0(H_0)\,\gamma(H_2)} \;\bigg|\; Z=k \right].\tag{47}\]
Step 2: Express the type-conditional terms.
For a transaction with \(Z=k\), the observation propensity and corruption penalty are the type-specific values \(q_0^{(k)}\) and \(\gamma^{(k)}\). However, the pooled estimator does not condition on \(Z\) — it uses the marginal propensity \(q_0\) and marginal corruption \(\gamma\).
The marginal propensity for a transaction with covariates \(H_0\) is: \[\label{eq:marginal95q} q_0(H_0) = \sum_{k=0}^{K} \pi_k(X)\,q_0^{(k)}(H_0),\tag{48}\] which is a mixture of type-specific propensities. Similarly for \(\gamma\).
The pooled estimator’s inverse-propensity weight \(1/(q_0\,\gamma)\) is the reciprocal of this mixture. By Jensen’s inequality (since \(g(q) = 1/q\) is convex for \(q > 0\)): \[\label{eq:jensen95step} \frac{1}{q_0(H_0)\,\gamma(H_2)} = \frac{1}{\sum_k \pi_k q_0^{(k)} \gamma^{(k)}} \;\geq\; \text{(not directly comparable)}.\tag{49}\]
However, the relevant comparison is not between inverse-propensity weights directly, but between the efficiency bounds of the two estimation strategies.
Step 3: Compute the decomposed efficiency bound.
The type-decomposed estimator estimates each \(\Psi_k = \bar\pi_k\) separately. The efficiency bound for estimating \(\Psi_k\) from \(n_k = n\bar\pi_k\) type-\(k\) transactions is \(\sigma^2_{\mathrm{eff}}(k)/n_k\), where \(\sigma^2_{\mathrm{eff}}(k)\) is defined in 14 .
Since \(\Psi = \sum_k \Psi_k\) and the type-specific estimators are asymptotically independent (they operate on disjoint subpopulations conditional on type assignment), the efficiency bound for the decomposed estimator of \(\Psi\) is: \[\label{eq:decomp95bound} \sigma^2_{\mathrm{decomp}} = \sum_{k=1}^{K} \sigma^2_{\mathrm{eff}}(k) = \sum_{k=1}^{K} \mathbb{E}\!\left[ \frac{\pi_k(X)(1-\pi_k(X))}{q_0^{(k)}(H_0)\,\gamma^{(k)}(H_2)} \right].\tag{50}\]
Step 4: Compare the bounds.
The pooled bound 46 can be rewritten using the law of total variance. Write \(f_0(x) = \sum_{k \geq 1} \pi_k(x)\), so: \[\label{eq:f095decomp} f_0(X)(1-f_0(X)) = \biggl(\sum_{k\geq 1}\pi_k(X)\biggr) \biggl(1 - \sum_{k\geq 1}\pi_k(X)\biggr).\tag{51}\]
The key inequality comes from comparing how the two estimators handle the inverse-propensity weighting. The pooled estimator applies the weight \(1/(q_0\,\gamma)\) — where \(q_0\) and \(\gamma\) are the marginal (type-averaged) propensity and corruption — to all transactions uniformly. The decomposed estimator applies the weight \(1/(q_0^{(k)}\,\gamma^{(k)})\) to type-\(k\) transactions specifically.
For the first (dominant) term of the efficiency bound, we use the following inequality. Define the random variable \(W = 1/(q_0^{(Z)}\gamma^{(Z)})\) (the type-specific inverse weight) and the constant (conditional on \(X\)) \(\bar W(X) = \mathbb{E}[W \mid X] = \sum_k \pi_k(X) / (q_0^{(k)}\gamma^{(k)})\). Then:
The pooled estimator effectively uses \(1/(\bar q_0 \bar\gamma)\) where \(\bar q_0\) is the type-averaged propensity, while the decomposed estimator uses the type-specific \(1/(q_0^{(k)}\gamma^{(k)})\). The variance inflation from pooling is captured by: \[\begin{align} \sigma^2_{\mathrm{pooled}} - \sigma^2_{\mathrm{decomp}} &= \mathbb{E}\!\left[ f_0(X)(1-f_0(X)) \cdot \left( \frac{1}{q_0(H_0)\gamma(H_2)} - \sum_k \frac{\pi_k(X)}{q_0^{(k)}(H_0)\gamma^{(k)}(H_2)} \right) \right] \nonumber\\ &\quad + \mathrm{Var}(f_0(X)) - \sum_k \mathrm{Var}(\pi_k(X)) \nonumber\\ &= \Delta_{\mathrm{IPW}} + \Delta_{\mathrm{var}}, \label{eq:gap95decomp} \end{align}\tag{52}\] where \(\Delta_{\mathrm{IPW}} \geq 0\) is the inverse-propensity-weighting penalty and \(\Delta_{\mathrm{var}}\) captures the variance decomposition.
For \(\Delta_{\mathrm{IPW}}\): the function \(q \mapsto 1/q\) is convex, so by Jensen’s inequality applied to the mixture \(q_0 = \sum_k \pi_k q_0^{(k)}\): \[\label{eq:jensen95ipw} \frac{1}{\sum_k \pi_k(X)\,q_0^{(k)}\gamma^{(k)}} \;\geq\; \sum_k \frac{\pi_k(X)}{q_0^{(k)}\gamma^{(k)}} \cdot \frac{\bigl(\sum_k \pi_k(X)\bigr)^2}{\sum_k \pi_k(X)},\tag{53}\] which, after careful algebra using the Cauchy–Schwarz or Jensen inequality on the harmonic mean, yields \(\Delta_{\mathrm{IPW}} \geq 0\).
For \(\Delta_{\mathrm{var}}\): since \(f_0 = \sum_{k\geq 1} \pi_k\), the variance of a sum exceeds the sum of variances when the summands are positively correlated. In practice, \(\Delta_{\mathrm{var}}\) can be positive or negative depending on the covariance structure of \(\pi_k(X)\) across types. However, the dominant term is \(\Delta_{\mathrm{IPW}}\), which is always non-negative.
More precisely, we establish the inequality via the conditional efficiency bounds. For each realization of \(X\), define the conditional pooled weight and conditional decomposed weight. The pooled estimator’s conditional variance contribution at covariate value \(x\) is: \[\label{eq:cond95pooled} V_{\mathrm{pooled}}(x) = \frac{f_0(x)(1-f_0(x))}{q_0(x)\,\gamma(x)},\tag{54}\] and the decomposed estimator’s conditional variance contribution is: \[\label{eq:cond95decomp} V_{\mathrm{decomp}}(x) = \sum_{k=1}^K \frac{\pi_k(x)(1-\pi_k(x))}{q_0^{(k)}(x)\,\gamma^{(k)}(x)}.\tag{55}\]
We claim \(V_{\mathrm{pooled}}(x) \geq V_{\mathrm{decomp}}(x)\) for all \(x\). To see this, note that \(f_0(x)(1-f_0(x)) = (\sum_k \pi_k)(1-\sum_k\pi_k) \geq \sum_k \pi_k(1-\pi_k) - \sum_{k\neq k'}\pi_k\pi_{k'}\), and the inverse weight in the pooled case is larger due to Jensen’s inequality on the harmonic mean.
The formal inequality follows from a result in semiparametric theory: the efficiency bound for a submodel (the decomposed model, which conditions on \(Z\)) is no larger than the efficiency bound for the full model (the pooled model, which marginalizes over \(Z\)). This is because conditioning on additional information (the type \(Z\)) can only reduce the Cramér–Rao bound.
Formally, let \(\mathcal{M}_{\mathrm{pooled}}\) be the semiparametric model that treats \(Z\) as latent, and let \(\mathcal{M}_{\mathrm{decomp}}\) be the model that conditions on \(Z\). Since \(\mathcal{M}_{\mathrm{decomp}} \subset \mathcal{M}_{\mathrm{pooled}}\) (conditioning on \(Z\) restricts the model), the efficiency bound in the submodel is no larger: \[\label{eq:submodel95ineq} \sigma^2_{\mathrm{eff}}(\mathcal{M}_{\mathrm{decomp}}) \leq \sigma^2_{\mathrm{eff}}(\mathcal{M}_{\mathrm{pooled}}).\tag{56}\] This is a standard result in semiparametric efficiency theory [11].
Equality in 56 holds if and only if the efficient influence function in \(\mathcal{M}_{\mathrm{pooled}}\) does not depend on \(Z\) — that is, the type-specific weights are identical across classes, giving condition 45 . ◻
Remark 17 (Interpretation of the dominance result). The theorem has a clean information-theoretic interpretation: knowing the fraud type is always (weakly) informative for estimation. The type \(Z\) determines which observation pipeline generated the label, and different pipelines have different propensities. Conditioning on \(Z\) allows the estimator to use the correct inverse-propensity weights for each transaction, rather than averaged weights that are correct for no individual transaction.
The result does not require that type membership be known with certainty. The decomposed STR uses estimated type probabilities \(\hat{\pi}_k(X)\), and the dominance holds as long as the type classifier is better than random assignment. Section 11 discusses the sensitivity to type misclassification.
We now quantify the MSE gap between the pooled and decomposed estimators.
Theorem 18 (Jensen penalty). Under the conditions of Theorem 16, the efficiency gap between the pooled and decomposed estimators is: \[\label{eq:jensen95penalty} \boxed{ \Delta \;=\; \sigma^2_{\mathrm{pooled}} - \sigma^2_{\mathrm{decomp}} \;=\; \mathbb{E}\!\left[ f_0(X)(1-f_0(X)) \cdot \mathrm{Var}_Z\!\left( \frac{1}{q_0^{(Z)}\,\gamma^{(Z)}} \;\bigg|\; X \right) \cdot \omega(X) \right], }\tag{57}\] where \(\mathrm{Var}_Z(\cdot \mid X)\) denotes the variance over the type distribution conditional on \(X\), and \(\omega(X) \geq 0\) is a weight function satisfying \(\omega(X) = 1\) when the type-specific propensities are independent of \(X\) given \(Z\).
Proof. The gap \(\Delta = \sigma^2_{\mathrm{pooled}} - \sigma^2_{\mathrm{decomp}}\) was decomposed in 52 into \(\Delta_{\mathrm{IPW}} + \Delta_{\mathrm{var}}\).
For the dominant term \(\Delta_{\mathrm{IPW}}\), define the type-specific inverse weight: \[\label{eq:type95weight} W^{(k)}(H_0, H_2) = \frac{1}{q_0^{(k)}(H_0)\, \gamma^{(k)}(H_2)}.\tag{58}\]
The pooled inverse weight, for a transaction whose type is unknown, effectively uses the harmonic mean of type-specific weights (weighted by type probabilities): \[\label{eq:pooled95weight} W_{\mathrm{pooled}}(H_0, H_2) = \frac{1}{\sum_k \pi_k(X) / W^{(k)}(H_0,H_2)}.\tag{59}\]
The decomposed estimator uses \(W^{(k)}\) directly for type-\(k\) transactions. The conditional (on \(X\)) contribution to \(\Delta_{\mathrm{IPW}}\) is: \[\begin{align} \delta(x) &= f_0(x)(1-f_0(x)) \cdot \left( W_{\mathrm{pooled}}(x) - \sum_k \pi_k(x)\, W^{(k)}(x) \right) \nonumber\\ &= f_0(x)(1-f_0(x)) \cdot \left( \frac{1}{\mathbb{E}_Z[1/W^{(Z)} \mid X\!=\!x]} - \mathbb{E}_Z[W^{(Z)} \mid X\!=\!x] \right). \label{eq:cond95gap} \end{align}\tag{60}\]
By the variance representation of the gap between the harmonic mean and the arithmetic mean, for any positive random variable \(W\): \[\label{eq:harm95arith} \frac{1}{\mathbb{E}[1/W]} - \mathbb{E}[W] \;\leq\; 0,\tag{61}\] with the magnitude related to \(\mathrm{Var}(W)\). However, the relevant comparison is the reverse: the pooled estimator uses \(1/q_0^{\mathrm{pooled}}\) which, due to the convexity of \(q \mapsto 1/q\), satisfies: \[\label{eq:convex95ineq} \mathbb{E}_Z\!\left[\frac{1}{q_0^{(Z)}\gamma^{(Z)}} \;\bigg|\; X\right] \;\geq\; \frac{1}{\mathbb{E}_Z[q_0^{(Z)}\gamma^{(Z)} \mid X]}.\tag{62}\]
The gap between the left and right sides is characterized by the second-order Taylor expansion: \[\label{eq:taylor95gap} \mathbb{E}_Z\!\left[\frac{1}{q_0^{(Z)}\gamma^{(Z)}} \;\bigg|\; X\right] - \frac{1}{\mathbb{E}_Z[q_0^{(Z)}\gamma^{(Z)} \mid X]} \;\approx\; \frac{\mathrm{Var}_Z(q_0^{(Z)}\gamma^{(Z)} \mid X)}{\bigl(\mathbb{E}_Z[q_0^{(Z)}\gamma^{(Z)} \mid X]\bigr)^3}.\tag{63}\]
Expressing this in terms of the inverse weights and integrating over \(X\) yields 57 with: \[\label{eq:omega} \omega(x) = \frac{\bigl(\mathbb{E}_Z[q_0^{(Z)}\gamma^{(Z)} \mid X\!=\!x]\bigr)^2}{\mathbb{E}_Z[(q_0^{(Z)}\gamma^{(Z)})^2 \mid X\!=\!x]}.\tag{64}\] By the Cauchy–Schwarz inequality, \(\omega(x) \in (0, 1]\), with \(\omega(x) = 1\) when \(q_0^{(Z)}\gamma^{(Z)}\) is non-random given \(X\) (i.e., all types have the same observation rate at \(x\)). ◻
Remark 19 (The penalty is driven by cross-class heterogeneity). The Jensen penalty 57 is large when:
\(\mathrm{Var}_Z(1/(q_0^{(Z)}\gamma^{(Z)}) \mid X)\) is large — the observation rates differ substantially across classes;
\(f_0(X)(1-f_0(X))\) is not too small — there is enough fraud to make the estimation problem non-trivial;
\(\omega(X)\) is close to 1 — the propensity heterogeneity is not dominated by within-class variation.
In payment networks, condition (i) is strongly satisfied: Class 1 has \(q_0^{(1)} \approx 0.50\), Class 3 has \(q_0^{(3)} \approx 0.02\), and Class 4 has \(q_0^{(4)} \approx 0.07\) (Example 3). The inverse weights range from \(\sim\!2\) to \(\sim\!55\), creating enormous cross-class variance.
For operational interpretation, we express the Jensen penalty in terms of observable network parameters.
Proposition 20 (Simplified Jensen penalty). Under approximate independence of type-specific propensities from covariates (i.e., \(q_0^{(k)}(H_0) \approx \bar q_0^{(k)}\) and \(\gamma^{(k)}(H_2) \approx \bar\gamma^{(k)}\) within each class), the Jensen penalty simplifies to: \[\label{eq:simple95penalty} \Delta \;\approx\; \bar f_0(1-\bar f_0) \cdot \sum_{k=1}^K \bar\pi_k \left( \frac{1}{\bar q_0^{(k)}\,\bar\gamma^{(k)}} - \bar W \right)^{\!2},\qquad{(1)}\] where \(\bar f_0 = \Psi = \sum_k\bar\pi_k\) is the overall fraud rate and \(\bar W = \sum_k \bar\pi_k / (\bar q_0^{(k)} \bar\gamma^{(k)})\) is the prevalence-weighted average inverse weight.
Proof. Under the stated approximation, \(\mathrm{Var}_Z(1/(q_0^{(Z)}\gamma^{(Z)}) \mid X) \approx \sum_k \pi_k(X)(1/(q_0^{(k)}\gamma^{(k)}) - \bar W)^2\) and \(\omega(X) \approx 1\). Substituting into 57 and using \(\mathbb{E}[f_0(X)(1-f_0(X))] \approx \bar f_0(1-\bar f_0)\) (valid when \(f_0\) does not vary too much) gives ?? . ◻
Example 9 (Jensen penalty for a typical payment network). Consider a network with the following fraud type distribution and observation parameters:
| Class | \(\bar\pi_k\) | \(\bar q_0^{(k)}\) | \(\bar\gamma^{(k)}\) | \(W^{(k)} = \frac{1}{\bar q_0^{(k)}\bar\gamma^{(k)}}\) | \(\bar\pi_k(W^{(k)}-\bar W)^2\) | Contrib.to \(\Delta\) |
|---|---|---|---|---|---|---|
| 1 (Card theft) | 0.005 | 0.50 | 0.81 | 2.47 | \(0.005 \times (2.47-9.44)^2\) | 0.243 |
| 2 (Friendly) | 0.003 | 0.33 | 0.25 | 12.12 | \(0.003 \times (12.12-9.44)^2\) | 0.022 |
| 3 (Bust-out) | 0.001 | 0.02 | 0.90 | 55.56 | \(0.001 \times (55.56-9.44)^2\) | 2.127 |
| 4 (Merchant) | 0.002 | 0.07 | 0.64 | 22.32 | \(0.002 \times (22.32-9.44)^2\) | 0.332 |
| 5 (Scams) | 0.001 | 0.10 | 0.81 | 12.35 | \(0.001 \times (12.35-9.44)^2\) | 0.008 |
| Total | 0.012 | — | — | \(\bar W = 9.44\) | — | \(\sum = 2.732\) |
The weighted average inverse weight is: \[\bar W = \frac{\sum_k \bar\pi_k\,W^{(k)}}{\sum_k \bar\pi_k} = \frac{0.005(2.47) + 0.003(12.12) + 0.001(55.56) + 0.002(22.32) + 0.001(12.35)}{0.012} = 9.44.\]
The Jensen penalty is: \[\Delta \approx 0.012 \times 0.988 \times 2.732 = 0.0324.\]
The pooled efficiency bound (from [2]) is approximately: \[\sigma^2_{\mathrm{pooled}} \approx \frac{0.012 \times 0.988}{0.20 \times 0.65} = 0.091,\] where \(0.20\) and \(0.65\) are the effective average observation rate and corruption penalty for the pooled estimator.
The Jensen penalty as a fraction of the pooled bound: \[\frac{\Delta}{\sigma^2_{\mathrm{pooled}}} = \frac{0.0324}{0.091} \approx 36\%.\]
The pooled STR wastes approximately 36% of its statistical efficiency by ignoring fraud type decomposition. Equivalently, the decomposed STR achieves the same accuracy as the pooled STR with 36% fewer transactions.
Remark 21 (The penalty is dominated by Class 3). In Example 9, Class 3 (bust-out) contributes 78% of the total Jensen penalty (\(2.127 / 2.732\)), despite representing only 8% of total fraud (\(0.001 / 0.012\)). This is because bust-out’s observation rate is \(\bar q_0^{(3)} = 0.02\) — an order of magnitude lower than any other class — creating an inverse weight of \(55.56\) that dominates the cross-class variance.
The practical implication: even if you only care about aggregate fraud estimation, failing to separate bust-out from other fraud types inflates the variance of your aggregate estimate by more than a third. The penalty comes not from the most common fraud type but from the most poorly observed one.
Remark 22 (When the penalty is small). The Jensen penalty is small when:
All classes have similar observation rates (unlikely in practice);
The poorly observed classes have negligible prevalence (possible if bust-out and merchant fraud are rare);
The overall fraud rate \(\bar f_0\) is very small, reducing the scale of all efficiency bounds.
In networks where bust-out and merchant fraud are significant concerns (which includes all major payment networks), the penalty is substantial.
The dominance theorem applies to fraud rate estimation, but the practical implication extends to model training.
Corollary 1 (Decomposed pseudo-labels dominate pooled pseudo-labels). Let \(\hat{y}^{\mathrm{pooled}}_t\) be the pseudo-label from the pooled STR ([2]) and \(\hat{y}^{\mathrm{decomp}}_t = \sum_k \hat{\pi}_k(X_t)\, \hat{y}^{(k)}_t\) be the pseudo-label from the decomposed STR. A downstream fraud model trained on decomposed pseudo-labels achieves lower expected prediction error than one trained on pooled pseudo-labels: \[\label{eq:pseudo95dominance} \mathbb{E}\!\left[ (f_0(X) - \hat{f}_{\mathrm{decomp}}(X))^2 \right] \;\leq\; \mathbb{E}\!\left[ (f_0(X) - \hat{f}_{\mathrm{pooled}}(X))^2 \right].\tag{65}\]
Proof. The pseudo-label is the projection of the STR score onto the feature space. The decomposed pseudo-label has lower variance (by Theorem 16) and the same expectation (both are consistent for \(f_0(x)\)). Lower variance in the training labels translates directly into lower expected prediction error of the downstream model, by the standard bias-variance decomposition for regression with noisy labels. ◻
Remark 23 (Operational implication). Corollary 1 says: train separate STR pseudo-label generators for each class, then combine the pseudo-labels for downstream model training. The downstream model itself need not be class-specific — a single fraud model can be trained on the combined decomposed pseudo-labels. The decomposition occurs in the label correction step, not in the model architecture.
This is operationally lightweight: it requires a type classifier and class-specific nuisance models (Phase 1 of [2]), but the downstream model training pipeline (Phase 3) is unchanged.
Class 1 is the baseline — the setting for which the STR of [2] was designed. This section establishes two results: (i) the STR applies without modification using type-specific propensities, and (ii) further decomposition within Class 1 (e.g., separating CNP from counterfeit) yields diminishing returns.
Proposition 24 (Class 1 STR). For fraud types in Class 1, the corruption-corrected STR of [2] applied with type-1-specific nuisance functions achieves the type-specific efficiency bound \(\sigma^2_{\mathrm{eff}}(1)\). Formally, let \(\hat{\varphi}_t^{(1)}\) denote the STR score computed with estimated propensities \((\hat{e}_0^{(1)}, \hat{r}_0^{(1)}, \hat{p}_0^{(1)})\) and outcome model \(\hat{\mu}_0^{(1)}\). Then: \[\label{eq:class195str} \hat{\Psi}_1 = \frac{1}{n}\sum_{t:\,\hat{Z}_t \in \mathrm{Class}\,1} \hat{\varphi}_t^{(1)}\qquad{(2)}\] is semiparametrically efficient for \(\Psi_1 = \bar\pi_1\), and inherits the sequential triple robustness property of [2].
Proof. Class 1 satisfies all the assumptions of [2]: positivity (Assumption 2), type-specific ignorability (Assumption 1), and the three-gate sequential structure. The STR’s efficiency and robustness proofs in [2] are stated conditionally on the observation pipeline structure and therefore apply directly to the type-conditional setting. The only change is that the nuisance functions are estimated on the Class 1 subpopulation (transactions with \(\hat{Z}_t \in \text{Class~1}\)), which satisfies the same regularity conditions as the full population. ◻
Remark 25 (What changes from [2]). The Class 1 STR is identical in form to STR formalized in [2]. The only operational difference is that the nuisance models are trained on the Class 1 subpopulation rather than the full population. This has two effects:
Better propensity models. The authorization propensity \(e_0^{(1)}\) is estimated from transactions where the authorization decision was driven by third-party fraud risk, not by first-party or merchant fraud signals. This reduces model misspecification.
Smaller training set. The Class 1 subpopulation is smaller than the full population. For large networks this is not a binding constraint, but for small issuers the empirical Bayes shrinkage of [2] becomes more important.
Class 1 contains multiple fraud subtypes (CNP, counterfeit, lost/stolen, account takeover, etc.). Should we decompose further?
Proposition 26 (Within-Class 1 Jensen penalty is small). Let \(\{1_a, 1_b, \ldots\}\) denote the subtypes within Class 1. The within-class Jensen penalty from pooling these subtypes is: \[\label{eq:within95penalty} \Delta_{\mathrm{within}}^{(1)} = \mathbb{E}\!\left[ f_0^{(1)}(X)(1-f_0^{(1)}(X)) \cdot \mathrm{Var}_{Z \mid Z \in \mathrm{Class}\,1}\!\left( \frac{1}{q_0^{(Z)}\,\gamma^{(Z)}} \;\bigg|\; X \right) \right].\qquad{(3)}\] This penalty is small relative to the between-class penalty \(\Delta\) (Theorem 18) whenever the subtypes within Class 1 have similar observation rates.
Proof. The formula follows from applying Theorem 18 to the within-class decomposition. The penalty is proportional to the within-class variance of inverse-propensity weights. Since all Class 1 subtypes share the same pipeline topology — victim reports, issuer processes chargeback, standard delay — the propensity magnitudes are similar. Specifically, for Class 1 subtypes:
Authorization propensities \(e_0^{(1_a)}\) differ moderately across subtypes (CNP may face more aggressive declining than lost/stolen), but typically lie in \([0.60, 0.95]\).
Reporting propensities \(r_0^{(1_a)}\) are similar because the reporting mechanism (victim notices and disputes) is the same regardless of attack vector.
Delay propensities \(p_0^{(1_a)}\) are nearly identical because the chargeback processing pipeline is the same.
The resulting within-class variance of inverse weights is an order of magnitude smaller than the between-class variance, where \(q_0\) ranges from \(0.02\) (Class 3) to \(0.50\) (Class 1). ◻
Example 10 (Within-Class 1 heterogeneity). Consider three Class 1 subtypes:
| Subtype | \(\bar e_0\) | \(\bar r_0\) | \(\bar p_0\) | \(W = 1/(\bar q_0\,\bar\gamma)\) |
|---|---|---|---|---|
| CNP fraud | 0.70 | 0.75 | 0.85 | \(1/(0.70\!\times\!0.75\!\times\!0.85\!\times\!0.81) = 2.77\) |
| Lost/stolen | 0.90 | 0.80 | 0.85 | \(1/(0.90\!\times\!0.80\!\times\!0.85\!\times\!0.81) = 2.02\) |
| Account takeover | 0.75 | 0.70 | 0.80 | \(1/(0.75\!\times\!0.70\!\times\!0.80\!\times\!0.81) = 2.94\) |
The inverse weights range from \(2.02\) to \(2.94\) — a factor of \(1.5\times\). Compare this to the between-class range of \(2.47\) (Class 1) to \(55.56\) (Class 3) — a factor of \(22\times\). The within-class variance is negligible relative to the between-class variance.
The within-class Jensen penalty is approximately: \[\Delta_{\mathrm{within}}^{(1)} \approx \bar\pi_1(1-\bar\pi_1) \cdot \mathrm{Var}(2.77, 2.02, 2.94) \approx 0.005 \times 0.995 \times 0.16 = 0.0008,\] which is 2.5% of the between-class penalty \(\Delta = 0.0324\) (Example 9).
Decomposing within Class 1 is not worth the additional modeling complexity. The five-class decomposition captures 97.5% of the available efficiency gain; further sub-type decomposition adds diminishing returns.
Remark 27 (When within-class decomposition helps). The within-class penalty can become non-negligible if a subtype has a substantially different authorization rate. For example, if a network implements very aggressive declining for CNP transactions (\(e_0^{(\mathrm{CNP})} = 0.40\)) while approving most card-present transactions (\(e_0^{(\mathrm{CP})} = 0.95\)), the within-class variance of inverse weights increases. In such cases, separating CNP from card-present within Class 1 may be worthwhile. The decision can be made empirically by computing \(\Delta_{\mathrm{within}}^{(1)}\) from the network’s data.
Class 2 is the most theoretically novel of the five classes. The defining feature is that the label is generated by the perpetrator, not the victim. The false-positive corruption rate \(\varepsilon_{01}\) is not a statistical nuisance — it is the equilibrium outcome of a strategic game between the cardholder and the issuer.
This section formalizes the dispute game, derives the equilibrium corruption rate, establishes comparative statics, and shows how the STR must be modified for Class 2.
Definition 18 (The dispute game). The dispute game \(\mathcal{G}\) is a two-player game between the cardholder (C) and the issuer (I), played after a transaction has been approved and completed.
Players and actions.
Cardholder (C): Chooses action \(D \in \{0, 1\}\), where \(D=1\) denotes filing a dispute and \(D=0\) denotes not filing.
Issuer (I): Maintains a detection function \(\beta: \mathcal{X}\to [0,1]\) that maps transaction features to the probability of identifying the dispute as first-party fraud (i.e., rejecting the chargeback and penalizing the cardholder).
Information structure.
The cardholder observes: the transaction amount \(a_t \in \mathbb{R}_+\), their own type \(\theta_C \in \{\text{honest}, \text{opportunistic}\}\), and a noisy signal of the issuer’s detection capability.
The issuer observes: the transaction features \(X_t\) and the dispute claim, but not the cardholder’s type \(\theta_C\).
Payoffs. For a transaction with amount \(a\) and features \(x\):
Definition 19 (Cardholder’s payoff). The cardholder’s expected payoff from filing a dispute (\(D=1\)) on a transaction with true label \(Y^*=0\) (legitimate) is: \[\label{eq:card95payoff} U_C(D=1 \mid Y^*\!=\!0,\, x,\, a) = \underbrace{\alpha(x) \cdot a}_{ \substack{\text{expected refund:}\\\text{success prob.} \times \text{amount}}} \;-\; \underbrace{c(x)}_{ \substack{\text{dispute}\\\text{cost}}} \;-\; \underbrace{\beta(x) \cdot p(x,a)}_{ \substack{\text{expected penalty:}\\\text{detection prob.} \times \text{penalty}}},\tag{66}\] where: \[\begin{align} \alpha(x) &= \mathbb{P}(\text{chargeback succeeds} \mid D\!=\!1,\, X\!=\!x) && \text{(success probability)}, \tag{67}\\ c(x) &\geq 0 && \text{(dispute filing cost: time, hassle, reputation)}, \tag{68}\\ \beta(x) &= \mathbb{P}(\text{identified as 1PF} \mid D\!=\!1,\, X\!=\!x) && \text{(detection probability)}, \tag{69}\\ p(x,a) &\geq 0 && \text{(penalty: account closure, blacklisting, legal action)}. \tag{70} \end{align}\] The payoff from not filing is \(U_C(D=0) = 0\).
Remark 28 (The payoff structure). The cardholder faces a risk–reward tradeoff:
Reward: the refund amount \(\alpha \cdot a\), which scales linearly with the transaction amount.
Costs: the filing cost \(c\) (typically small — a phone call or online form) and the expected penalty \(\beta \cdot p\) (potentially large if detected).
Key driver: for low-friction dispute processes (\(c \approx 0\)) and weak detection (\(\beta\) small), the expected payoff is positive for most transaction amounts, encouraging disputes.
Definition 20 (Issuer’s payoff). The issuer’s payoff from a transaction with features \(x\), amount \(a\), and cardholder type \(\theta_C\) is: \[\label{eq:issuer95payoff} U_I(\beta \mid x, a, \theta_C) = -\underbrace{\mathbf{1}[\theta_C = \text{honest}] \cdot \beta(x) \cdot \ell_{\mathrm{FP}}(x,a)}_{ \substack{\text{false positive cost:}\\\text{wrongly penalizing honest cardholder}}} \;-\; \underbrace{\mathbf{1}[\theta_C = \text{opp.}] \cdot (1-\beta(x)) \cdot a}_{ \substack{\text{false negative cost:}\\\text{paying fraudulent refund}}},\tag{71}\] where \(\ell_{\mathrm{FP}}(x,a)\) is the cost of falsely accusing an honest cardholder (customer churn, reputational damage, regulatory penalty).
Remark 29 (The issuer’s dilemma). The issuer faces a classic detection tradeoff: increasing \(\beta\) (more aggressive detection) reduces false-negative losses (fraudulent refunds paid) but increases false-positive losses (honest cardholders wrongly accused). The optimal \(\beta^*\) balances these costs, and the cardholder’s dispute behavior depends on \(\beta^*\). This creates a strategic interdependence that we now formalize.
Definition 21 (Cardholder’s best response). Under a logistic (quantal) response model with rationality parameter \(T > 0\), the cardholder’s dispute probability is: \[\label{eq:card95br} d^*(x, a; \beta) = \sigma\!\left( \frac{\alpha(x) \cdot a - c(x) - \beta(x) \cdot p(x,a)}{T} \right),\tag{72}\] where \(\sigma(z) = 1/(1+e^{-z})\) is the sigmoid function. The parameter \(T\) controls the cardholder’s rationality: \(T \to 0\) gives perfectly rational behavior (dispute iff payoff \(> 0\)); \(T \to \infty\) gives random behavior (\(d^* \to 0.5\)).
Remark 30 (Why logistic response). The logistic response model (also known as the quantal response equilibrium; [12]) is standard in behavioral game theory. It captures the empirical observation that people are approximately rational: they are more likely to take actions with higher payoffs, but they do not optimize perfectly. The parameter \(T\) absorbs heterogeneity in cardholder sophistication, information quality, and idiosyncratic preferences.
Definition 22 (Issuer’s best response). The issuer chooses \(\beta\) to minimize expected loss. Given the cardholder population’s dispute rate \(d^*\) and the prior probability \(\pi_2(x)\) that a disputing cardholder is an opportunistic first-party fraudster, the issuer’s optimal detection threshold is: \[\label{eq:issuer95br} \beta^*(x; d^*) = \mathop{\mathrm{arg\,min}}_{\beta \in [0,1]}\; \mathbb{E}_{\theta_C}\!\bigl[ U_I(\beta \mid x, a, \theta_C) \mid D=1 \bigr].\tag{73}\]
Under the standard binary classification framework, the issuer’s optimal detection probability satisfies: \[\label{eq:issuer95opt} \beta^*(x) = \mathbf{1}\!\left[ \frac{\pi_2(x) \cdot a}{(1-\pi_2(x)) \cdot \ell_{\mathrm{FP}}(x,a)} > 1 \right]\tag{74}\] for a deterministic policy, or a smoothed version for a randomized policy. In practice, \(\beta^*\) is the output of the issuer’s first-party fraud detection model.
We now establish the central result: the endogenous corruption theorem.
Theorem 31 (Endogenous corruption). In the dispute game \(\mathcal{G}\) with logistic response cardholder (Definition 21) and cost-minimizing issuer (Definition 22):
(i) Existence. A Nash equilibrium \((d^*, \beta^*)\) exists.
(ii) Endogenous corruption. At equilibrium, the false-positive corruption rate for Class 2 is: \[\label{eq:endog95corruption} \boxed{ \varepsilon_{01}^{(2)}(x, a) = d^*(x, a;\, \beta^*) = \sigma\!\left( \frac{\alpha(x) \cdot a - c(x) - \beta^*(x) \cdot p(x,a)}{T} \right). }\tag{75}\] This is the probability that a cardholder files a false dispute on a legitimate transaction with features \(x\) and amount \(a\).
(iii) Comparative statics. The equilibrium corruption rate satisfies: \[\begin{align} \frac{\partial\varepsilon_{01}^{(2)}}{\partial a} &> 0 && \text{(higher amount \to more disputes)}, \tag{76}\\ \frac{\partial\varepsilon_{01}^{(2)}}{\partial c} &< 0 && \text{(higher dispute cost \to fewer disputes)}, \tag{77}\\ \frac{\partial\varepsilon_{01}^{(2)}}{\partial \beta^*} &< 0 && \text{(better detection \to fewer disputes)}, \tag{78}\\ \frac{\partial\varepsilon_{01}^{(2)}}{\partial p} &< 0 && \text{(harsher penalty \to fewer disputes)}. \tag{79} \end{align}\]
(iv) Feedback loop. The equilibrium is the fixed point of the mapping \(\Phi: [0,1] \to [0,1]\) defined by: \[\label{eq:fixed95point} \Phi(\varepsilon) = \sigma\!\left( \frac{\alpha \cdot a - c - \beta^*(\varepsilon) \cdot p}{T} \right),\tag{80}\] where \(\beta^*(\varepsilon)\) is the issuer’s optimal detection probability given that a fraction \(\varepsilon\) of disputes are fraudulent. The equilibrium satisfies \(\varepsilon_{01}^{(2)} = \Phi(\varepsilon_{01}^{(2)})\).
Proof. Part (i): Existence.
The cardholder’s best response \(d^*(\cdot;\beta): [0,1] \to (0,1)\) is a continuous function of \(\beta\) (the sigmoid is continuous and strictly monotone). The issuer’s best response \(\beta^*(\cdot;d^*): (0,1) \to [0,1]\) is a continuous function of \(d^*\) (the issuer’s cost function is continuous in \(d^*\), which determines the posterior probability that a dispute is fraudulent).
The composite mapping \((d^*,\beta^*) \mapsto (d^*(\cdot;\beta^*(\cdot;d^*)),\, \beta^*(\cdot;d^*(\cdot;\beta^*)))\) is a continuous self-map of the compact convex set \((0,1) \times [0,1] \subset \mathbb{R}^2\). By Brouwer’s fixed point theorem, a fixed point exists.
Part (ii): Endogenous corruption.
At the fixed point \((d^*, \beta^*)\), the cardholder’s dispute probability on a legitimate transaction is \(d^*(x,a;\beta^*)\), which is 72 evaluated at the equilibrium \(\beta^*\). This is precisely the false-positive corruption rate: the probability that a legitimate transaction receives a fraud label (through the cardholder’s false dispute).
To verify the identification: \[\begin{align} \varepsilon_{01}^{(2)}(x,a) &= \mathbb{P}(\tilde{Y} = 1 \mid Y^* = 0,\, O=1,\, X=x,\, Z=2) \\ &= \mathbb{P}(D = 1 \mid Y^* = 0,\, X=x,\, Z=2) \\ &= d^*(x,a;\beta^*), \end{align}\] where the second equality uses the fact that for Class 2 transactions with \(Y^*=0\) (legitimate), the observed label \(\tilde{Y} = D\) (the label equals the dispute indicator), and \(O=1\) requires \(D=1\) (a label is observed only if a dispute is filed).
Part (iii): Comparative statics.
Each comparative static follows from the monotonicity of the sigmoid function \(\sigma\) and the signs of the partial derivatives of its argument:
76 : \(\partial/\partial a\,[\alpha a - c - \beta^* p] = \alpha - \beta^*(\partial p/\partial a)\). Since \(\alpha \in (0,1]\) and \(\partial p/\partial a \geq 0\) is typically small relative to \(\alpha\) (the penalty does not scale one-for-one with the amount), the net effect is positive: higher amounts increase the dispute incentive.
77 : \(\partial/\partial c\,[\alpha a - c - \beta^* p] = -1 < 0\). Higher dispute costs unambiguously reduce disputes.
78 : \(\partial/\partial\beta^*\,[\alpha a - c - \beta^* p] = -p < 0\). Better detection (higher \(\beta^*\)) reduces disputes through the deterrence channel.
79 : \(\partial/\partial p\,[\alpha a - c - \beta^* p] = -\beta^* < 0\). Harsher penalties reduce disputes (conditional on nonzero detection probability).
Part (iv): Feedback loop.
The mapping \(\Phi\) in 80 captures the circular dependence: the dispute rate \(\varepsilon\) determines the fraction of fraudulent disputes, which determines the issuer’s optimal detection threshold \(\beta^*(\varepsilon)\), which determines the cardholder’s dispute incentive, which determines \(\varepsilon\).
Formally, \(\beta^*(\varepsilon)\) is the issuer’s Bayes- optimal detection probability given that a fraction \(\varepsilon/(1-\pi_2 + \pi_2\varepsilon)\) of all disputes are fraudulent (by Bayes’ rule, combining the prior \(\pi_2\) with the dispute rate \(\varepsilon\) for legitimate transactions and a dispute rate of 1 for truly fraudulent transactions).
The mapping \(\Phi\) is continuous (composition of continuous functions) and maps \([0,1]\) to \((0,1) \subset [0,1]\). By Brouwer’s theorem, \(\Phi\) has a fixed point, which is the equilibrium \(\varepsilon_{01}^{(2)}\). ◻
Proposition 32 (Sufficient condition for uniqueness). If the deterrence effect is strong enough relative to the cardholder’s rationality: \[\label{eq:uniqueness95cond} \sup_{x,a}\; \frac{p(x,a)}{T} \cdot \left| \frac{\partial\beta^*}{\partial\varepsilon} \right| \;<\; 4,\qquad{(4)}\] then the equilibrium is unique.
Proof. The mapping \(\Phi\) in 80 has a unique fixed point if \(|\Phi'(\varepsilon)| < 1\) for all \(\varepsilon \in [0,1]\). Computing: \[\Phi'(\varepsilon) = \sigma'(\cdot) \cdot \frac{-p}{T} \cdot \frac{\partial\beta^*}{\partial\varepsilon},\] where \(\sigma'(z) = \sigma(z)(1-\sigma(z)) \leq 1/4\). Therefore: \[|\Phi'(\varepsilon)| \leq \frac{1}{4} \cdot \frac{p}{T} \cdot \left|\frac{\partial\beta^*}{\partial\varepsilon}\right|.\] The condition \(|\Phi'| < 1\) is guaranteed by ?? , and the contraction mapping theorem gives uniqueness. ◻
Proposition 33 (Multiple equilibria). When condition ?? is violated — specifically, when the penalty \(p\) is large, the rationality \(T\) is small, and the issuer’s detection probability is highly sensitive to the dispute rate — the game can have multiple equilibria:
A low-fraud equilibrium: \(\varepsilon_{01}^{(2)}\) is small \(\to\) the issuer’s model is accurate (trained on mostly genuine disputes) \(\to\) \(\beta^*\) is high \(\to\) deterrence is strong \(\to\) few disputes \(\to\) \(\varepsilon_{01}^{(2)}\) is small.
A high-fraud equilibrium: \(\varepsilon_{01}^{(2)}\) is large \(\to\) the issuer’s model is contaminated (trained on many false disputes) \(\to\) \(\beta^*\) is low \(\to\) deterrence is weak \(\to\) many disputes \(\to\) \(\varepsilon_{01}^{(2)}\) is large.
Proof. When \(|\Phi'(\varepsilon)| > 1\) at some interior point, the mapping \(\Phi\) crosses the 45-degree line multiple times. Since \(\Phi(0) > 0\) (some disputes occur even with strong deterrence, due to the stochastic component \(T > 0\)) and \(\Phi(1) < 1\) (even with no deterrence, not all cardholders dispute), the intermediate value theorem guarantees at least one crossing. When \(\Phi\) is S-shaped (which occurs when the issuer’s detection probability \(\beta^*(\varepsilon)\) decreases steeply in \(\varepsilon\) — i.e., label contamination severely degrades the detection model), there are exactly three crossings: two stable equilibria (low-fraud and high-fraud) separated by an unstable equilibrium. ◻
Example 11 (The tipping point). Consider a network where:
Average transaction amount: \(a = \$200\)
Chargeback success probability: \(\alpha = 0.85\)
Dispute filing cost: \(c = \$5\) (minimal — online form)
Detection penalty: \(p = \$2{,}000\) (account closure + blacklisting)
Cardholder rationality: \(T = 50\)
Low-fraud equilibrium (\(\beta^* = 0.15\)): \[\varepsilon_{01}^{(2)} = \sigma\!\left( \frac{0.85 \times 200 - 5 - 0.15 \times 2000}{50} \right) = \sigma\!\left(\frac{170 - 5 - 300}{50}\right) = \sigma(-2.7) = 0.063.\] Only 6.3% of legitimate transactions are falsely disputed. The issuer’s model is trained on mostly genuine disputes, so \(\beta^*\) is accurate and deterrence is strong.
High-fraud equilibrium (\(\beta^* = 0.02\)): \[\varepsilon_{01}^{(2)} = \sigma\!\left( \frac{170 - 5 - 0.02 \times 2000}{50} \right) = \sigma\!\left(\frac{170 - 5 - 40}{50}\right) = \sigma(2.5) = 0.924.\] Over 92% of legitimate transactions are falsely disputed. The issuer’s model is overwhelmed by false disputes, \(\beta^*\) collapses to near-zero, and deterrence vanishes.
The difference between 6.3% and 92.4% corruption is not a parameter — it is an equilibrium selection problem. A network can be in either equilibrium depending on its history. Once the high-fraud equilibrium is reached (e.g., due to a temporary relaxation of dispute friction), it is self-reinforcing and difficult to escape.
Remark 34 (Policy implications of multiple equilibria). The multiple equilibria result has direct policy implications:
Dispute friction matters. Reducing dispute costs \(c\) (e.g., one-click dispute buttons in mobile banking apps) can push the network from the low-fraud to the high-fraud equilibrium. The comparative static 77 shows that lower \(c\) increases \(\varepsilon_{01}^{(2)}\), but the equilibrium analysis shows this can be a discontinuous jump rather than a smooth increase.
Detection investment has a multiplier. Improving the first-party fraud model (increasing \(\beta^*\)) not only catches more fraudsters directly but also deters marginal disputants, reducing the contamination of training labels, which further improves the model. This is a virtuous cycle that the network should invest in.
Penalty design matters. The penalty \(p\) enters the uniqueness condition ?? . Sufficiently harsh and credible penalties can eliminate the high-fraud equilibrium entirely, ensuring the network remains in the low-fraud regime.
Proposition 35 (Class 2 STR with endogenous corruption). For Class 2, the corruption-corrected STR of [2] remains valid, provided the corruption rate \(\varepsilon_{01}^{(2)}(x,a)\) is estimated from the equilibrium 75 rather than from historical audit data.
Specifically, the Class 2 corruption-corrected label is: \[\label{eq:class295corr} \tilde{Y}_t^{\mathrm{corr},(2)} = \frac{\tilde{Y}_t - \hat{\varepsilon}_{01}^{(2)}(X_t, a_t)}{1 - \hat{\varepsilon}_{10}^{(2)} - \hat{\varepsilon}_{01}^{(2)}(X_t, a_t)},\qquad{(5)}\] where \(\hat{\varepsilon}_{01}^{(2)}(x,a)\) is the estimated equilibrium dispute probability from the dispute game.
Proof. The corruption correction of [2] requires only that \(\varepsilon_{01}\) and \(\varepsilon_{10}\) be consistently estimated — it does not require that they be exogenous. The correction formula is algebraically identical; only the estimation procedure for \(\varepsilon_{01}\) changes.
Consistency of \(\hat{\varepsilon}_{01}^{(2)}\) requires consistent estimation of the game’s primitives \((\alpha, c, \beta^*, p, T)\). The chargeback success rate \(\alpha\) is observable from historical dispute outcomes. The detection probability \(\beta^*\) is the output of the issuer’s first-party fraud model (or can be estimated from the fraction of disputes that are successfully challenged). The penalty \(p\) is set by issuer policy. The rationality parameter \(T\) and dispute cost \(c\) can be estimated from the observed dispute rate as a function of transaction amount (the logistic model 72 is a standard discrete choice model estimable by maximum likelihood). ◻
Remark 36 (Estimation procedure for Class 2). The practical estimation procedure for Class 2 has four steps:
Estimate the dispute game primitives. From historical data with mature labels, estimate:
\(\hat{\alpha}(x)\): chargeback success rate (fraction of disputes that result in a refund), by transaction characteristics.
\(\hat{\beta}^*(x)\): first-party fraud detection rate (fraction of disputes successfully identified as first-party fraud), by transaction characteristics.
\(\hat{T}\), \(\hat{c}\): rationality and cost parameters, estimated by fitting the logistic model 72 to the observed dispute rate as a function of amount and features.
Compute the equilibrium corruption rate. Solve the fixed point 80 to obtain \(\hat{\varepsilon}_{01}^{(2)}(x,a)\).
Apply the corruption correction. Compute \(\tilde{Y}_t^{\mathrm{corr},(2)}\) using ?? .
Apply the STR. Use the corruption-corrected labels in the two-gate STR (reporting + delay corrections only, since \(e_0^{(2)} \approx 1\)).
Proposition 37 (Class 2 efficiency bound). The semiparametric efficiency bound for estimating the Class 2 fraud rate \(\Psi_2 = \bar\pi_2\) is: \[\label{eq:class295bound} \sigma^2_{\mathrm{eff}}(2) = \mathbb{E}\!\left[ \frac{\pi_2(X)(1-\pi_2(X))}{r_0^{(2)}(H_1) \cdot p_0^{(2)}(H_2) \cdot \gamma^{(2)}(X,a)} \right],\qquad{(6)}\] where \(\gamma^{(2)}(x,a) = (1 - \varepsilon_{10}^{(2)} - \varepsilon_{01}^{(2)}(x,a))^2\) and \(\varepsilon_{01}^{(2)}(x,a)\) is the equilibrium corruption rate from Theorem 31.
Note that the authorization propensity \(e_0^{(2)}\) does not appear in the bound because \(e_0^{(2)} = 1\).
Proof. Apply the efficiency bound of [2] with \(e_0^{(2)} = 1\) (the authorization gate is non-binding) and the Class 2 corruption rates. The bound reduces to a two-gate version depending only on \(r_0^{(2)}\), \(p_0^{(2)}\), and \(\gamma^{(2)}\). ◻
Example 12 (Class 2 efficiency bound is corruption-dominated). Consider Class 2 with:
| Parameter | Low-fraud eq. | High-fraud eq. |
|---|---|---|
| \(\varepsilon_{01}^{(2)}\) | 0.063 | 0.924 |
| \(\varepsilon_{10}^{(2)}\) | 0.02 | 0.02 |
| \(\gamma^{(2)} = (1-\varepsilon_{10}-\varepsilon_{01})^2\) | \((1-0.083)^2 = 0.841\) | \((1-0.944)^2 = 0.003\) |
| \(\bar r_0^{(2)}\) | 0.60 | 0.60 |
| \(\bar p_0^{(2)}\) | 0.85 | 0.85 |
| \(\sigma^2_{\mathrm{eff}}(2) \propto 1/(\bar r_0\,\bar p_0\,\gamma)\) | \(1/(0.60 \times 0.85 \times 0.841) = 2.33\) | \(1/(0.60 \times 0.85 \times 0.003) = 654\) |
In the low-fraud equilibrium, the Class 2 bound is comparable to Class 1 (the corruption penalty is modest). In the high-fraud equilibrium, the bound is 280\(\times\) larger — almost all information is destroyed by the adversarial corruption.
The equilibrium selection determines whether Class 2 is detectable or not. This is a qualitative difference from Class 1, where the bound varies smoothly with parameters. In Class 2, the bound can jump discontinuously between equilibria.
Remark 38 ([2] vs.present paper’s treatment of corruption). [2] treats the corruption rates \(\varepsilon_{10}\) and \(\varepsilon_{01}\) as fixed nuisance parameters estimated from audit data. This is correct for Class 1 (where corruption is exogenous operational error) but incorrect for Class 2 (where corruption is an endogenous equilibrium).
The distinction matters practically:
Audit-based estimation is biased for Class 2. Historical audit samples measure the past equilibrium corruption rate. If the game parameters change (e.g., a new mobile banking app reduces dispute friction \(c\)), the corruption rate shifts to a new equilibrium that historical audits do not reflect.
The corruption rate depends on the detection model. In Class 1, \(\varepsilon_{01}\) does not depend on the fraud model. In Class 2, \(\varepsilon_{01}^{(2)}\) depends on \(\beta^*\), which is the output of the first-party fraud model. Improving the model changes the corruption rate, which changes the optimal STR correction, which changes the model’s training labels. This circularity must be handled carefully (e.g., by iterating between model training and equilibrium computation).
The STR can influence the equilibrium. Deploying the STR with accurate Class 2 corrections improves the first-party fraud model (better labels \(\to\) better model), which increases \(\beta^*\), which reduces \(\varepsilon_{01}^{(2)}\) (the deterrence channel), which further improves labels. The STR is not merely a statistical correction — it is an intervention that shifts the dispute game toward the low-fraud equilibrium.
Class 3 is the most structurally extreme of the five classes. The observation pipeline is not merely impaired — it is inoperative during the critical detection window. This section proves two results: (i) the STR’s positivity assumption fails structurally, and (ii) no account-level classifier can exceed random guessing during the build-up phase — detection requires cross-account graph signals.
Definition 23 (Bust-out temporal structure). A Class 3 (bust-out) account \(a\) is characterized by:
A latent bust-out indicator \(\theta_a \in \{0,1\}\), where \(\theta_a = 1\) denotes a bust-out account and \(\theta_a = 0\) denotes a legitimate account;
A bust-out time \(\tau_a \in (0,\infty]\), where \(\tau_a < \infty\) for bust-out accounts and \(\tau_a = \infty\) for legitimate accounts;
A transaction sequence \(\{(X_{at}, Y^*_{at})\}_{t=1}^{T_a}\) generated by the account over its lifetime.
The fraud state of individual transactions depends on the account’s phase: \[\label{eq:bustout95fraud} Y^*_{at} = \begin{cases} 0 & \text{if } t < \tau_a \quad\text{(build-up: all transactions legitimate)},\\ 1 & \text{if } t \geq \tau_a \quad\text{(exploitation: all transactions fraudulent)}. \end{cases}\tag{81}\]
Remark 39 (The build-up phase is genuine). A critical feature of 81 : during build-up (\(t < \tau_a\)), the transactions are genuinely legitimate. The synthetic identity is building credit by making real purchases and paying on time. These are not “fraudulent transactions that look legitimate” — they are legitimate transactions, made with the intent of eventually exploiting the credit line. The fraud is a property of the account trajectory, not of any individual transaction.
This is why the standard transaction-level estimand \(f_0(x) = \mathbb{P}(Y^*=1 \mid X=x)\) is the wrong target for Class 3. The correct estimand is the account-level bust-out probability: \[\label{eq:bustout95estimand} \theta_a = \mathbb{P}(\text{bust-out} \mid \mathcal{H}_a^{(t)}) = \mathbb{P}(\tau_a < \infty \mid \mathcal{H}_a^{(t)}),\tag{82}\] where \(\mathcal{H}_a^{(t)} = \{(X_{as}, Y^*_{as})\}_{s \leq t}\) is the account’s observed history up to time \(t\).
Theorem 40 (Positivity failure for Class 3). During the build-up phase (\(t < \tau_a\)) of a Class 3 account, the observation propensity is exactly zero: \[\label{eq:positivity95failure} q_0^{(3)}(H_{0,at}) = e_0^{(3)}(H_{0,at}) \cdot r_0^{(3)}(H_{1,at}) \cdot p_0^{(3)}(H_{2,at}) = 0 \qquad\text{for all } t < \tau_a,\tag{83}\] and the semiparametric efficiency bound for estimating \(\theta_a\) from transaction-level labels diverges: \[\label{eq:bound95diverges} \sigma^2_{\mathrm{eff}}(3, t < \tau_a) = \infty.\tag{84}\]
Proof. Step 1: The reporting gate is structurally zero.
During build-up, \(Y^*_{at} = 0\) for all transactions (equation 81 ). No fraud has occurred, so:
No victim exists (the synthetic identity is not a real person who would notice unauthorized charges);
No cardholder dispute is filed (the account is paying its bills);
No issuer investigation is triggered (the account behaves normally).
Therefore \(R_{at} = 0\) with probability 1 for \(t < \tau_a\). This gives \(r_0^{(3)}(H_{1,at}) = \mathbb{P}(R_{at}=1 \mid H_{1,at}, A_{at}=1, Z=3) = 0\).
Step 2: The product is zero.
Since \(r_0^{(3)} = 0\): \[q_0^{(3)} = e_0^{(3)} \cdot 0 \cdot p_0^{(3)} = 0,\] regardless of the values of \(e_0^{(3)}\) and \(p_0^{(3)}\). This is an exact zero, not an approximation.
Step 3: The efficiency bound diverges.
The efficiency bound (Definition 8) contains \(1/q_0^{(k)}\) in the denominator. With \(q_0^{(3)} = 0\): \[\sigma^2_{\mathrm{eff}}(3, t < \tau_a) = \mathbb{E}\!\left[ \frac{\pi_3(X)(1-\pi_3(X))}{0 \cdot \gamma^{(3)}} \right] = \infty,\] whenever \(\pi_3(X) > 0\) with positive probability (i.e., there exist transactions with nonzero bust-out probability).
Step 4: This is structural, not statistical.
The positivity failure is not due to finite-sample issues or poor modeling — it is a structural feature of Class 3. No amount of data, no choice of propensity model, and no statistical technique can create labels that do not exist. The build-up phase generates zero information about the bust-out state through the transaction-level label pipeline. ◻
Remark 41 (Contrast with near-positivity violations). In the causal inference literature, positivity violations are typically practical rather than structural: the propensity is small but positive, and the concern is high variance from large inverse-propensity weights. Solutions include propensity trimming, overlap weighting, or the empirical Bayes shrinkage of [2].
Class 3’s violation is qualitatively different. The propensity is exactly zero — not small, not approximately zero, but identically zero by the structure of the problem. No statistical technique can overcome this. The STR is not merely high-variance for Class 3 during build-up — it is undefined. This is the formal justification for seeking fundamentally different detection approaches.
The positivity failure shows that label-based detection is impossible during build-up. We now prove a stronger result: even with perfect labels (or equivalently, even with access to the true \(Y^*_{at}\) for every transaction), no account-level classifier can detect bust-out accounts during build-up.
Theorem 42 (Individual-level insufficiency). There exist bust-out data-generating processes satisfying Definition 23 such that:
(i) The Bayes-optimal account-level classifier has \(\mathop{\mathrm{AUC}}= 0.5\) during the build-up phase: \[\label{eq:auc95account} \mathop{\mathrm{AUC}}\!\left( \hat{\theta}^{\mathrm{acct}}_a \mid t < \tau_a \right) = 0.5,\tag{85}\] where \(\hat{\theta}^{\mathrm{acct}}_a\) is any classifier based on account \(a\)’s own transaction history \(\mathcal{H}_a^{(t)}\).
(ii) The Bayes-optimal ring-level classifier has \(\mathop{\mathrm{AUC}}\to 1\) as the ring size \(m \to \infty\): \[\label{eq:auc95ring} \mathop{\mathrm{AUC}}\!\left( \hat{\theta}^{\mathrm{ring}}_{\mathcal{R}} \right) \to 1 \qquad\text{as } m = |\mathcal{R}| \to \infty,\tag{86}\] where \(\mathcal{R}\) is a candidate ring of \(m\) accounts and \(\hat{\theta}^{\mathrm{ring}}_{\mathcal{R}}\) is any classifier based on the joint transaction histories \(\{\mathcal{H}_a^{(t)}\}_{a \in \mathcal{R}}\).
Proof. We prove both parts by explicit construction of an adversarial data-generating process.
Construction of the DGP.
Let \(F\) be a distribution over transaction features \(\mathcal{X}\) (representing normal spending patterns: amounts, merchants, timing, etc.). Define the data-generating process as follows:
Account creation. Each account \(a\) is independently assigned: \[\label{eq:dgp95type} \theta_a \sim \mathrm{Bernoulli}(\pi_3),\tag{87}\] where \(\pi_3\) is the bust-out prevalence.
Ring structure. Bust-out accounts (\(\theta_a = 1\)) are organized into rings of size \(m\). Accounts within a ring share:
Creation dates within a window \([s, s+\delta]\) for some ring-specific \(s\) and small \(\delta > 0\);
Initial credit limits drawn from the same distribution \(G_{\mathrm{ring}}\) (reflecting the ring’s application strategy);
A common set of “anchor merchants” \(\mathcal{M}_{\mathrm{ring}} \subset \mathcal{M}\) that all ring members transact with.
Legitimate accounts (\(\theta_a = 0\)) are created independently with no shared structure.
Transaction generation during build-up. For both bust-out and legitimate accounts, each transaction during build-up is drawn i.i.d.from the same distribution: \[\label{eq:dgp95txn} X_{at} \mid \theta_a,\, t < \tau_a \;\sim\; F \qquad\text{for all } \theta_a \in \{0,1\}.\tag{88}\] This is the key assumption: bust-out accounts are designed to mimic legitimate behavior during build-up.
Part (i): Account-level \(\mathop{\mathrm{AUC}}= 0.5\).
By the Neyman–Pearson lemma, the most powerful test for \(\theta_a = 1\) vs.\(\theta_a = 0\) based on \(\mathcal{H}_a^{(t)} = (X_{a1}, \ldots, X_{at})\) is the likelihood ratio test: \[\label{eq:lr95test} \Lambda_a^{(t)} = \frac{P(\mathcal{H}_a^{(t)} \mid \theta_a = 1,\, t < \tau_a)}{P(\mathcal{H}_a^{(t)} \mid \theta_a = 0)}.\tag{89}\]
By 88 , the conditional distributions are identical: \[\label{eq:lr95one} P(\mathcal{H}_a^{(t)} \mid \theta_a = 1,\, t < \tau_a) = \prod_{s=1}^t f(X_{as}) = P(\mathcal{H}_a^{(t)} \mid \theta_a = 0),\tag{90}\] where \(f\) is the density of \(F\). Therefore: \[\label{eq:lr95trivial} \Lambda_a^{(t)} = 1 \qquad\text{for all } a \text{ and all } t < \tau_a.\tag{91}\]
A constant likelihood ratio implies zero discriminative power. The ROC curve is the diagonal, and \(\mathop{\mathrm{AUC}}= 0.5\).
This holds for any function of \(\mathcal{H}_a^{(t)}\) — not just the likelihood ratio. Any statistic computed from \(\mathcal{H}_a^{(t)}\) (velocities, aggregations, temporal patterns, behavioral features) has the same distribution under \(\theta_a = 0\) and \(\theta_a = 1\) during build-up, because the underlying transaction distribution is identical.
Part (ii): Ring-level \(\mathop{\mathrm{AUC}}\to 1\).
Now consider a candidate set \(\mathcal{R}= \{a_1, \ldots, a_m\}\) of \(m\) accounts. We test whether \(\mathcal{R}\) is a bust-out ring (all \(\theta_{a_i} = 1\) and the accounts are linked) or a random set of legitimate accounts (all \(\theta_{a_i} = 0\) or unlinked).
Define the ring cohesion statistic: \[\label{eq:ring95stat} T_{\mathcal{R}} = \sum_{i < j} S(a_i, a_j),\tag{92}\] where \(S(a_i, a_j)\) is a pairwise similarity measure: \[\label{eq:similarity} S(a_i, a_j) = \underbrace{K_{\mathrm{time}}(c_i, c_j)}_{\substack{ \text{creation date}\\ \text{similarity}}} \cdot \underbrace{K_{\mathrm{limit}}(\ell_i,\ell_j)}_{\substack{ \text{credit limit}\\ \text{similarity}}} \cdot \underbrace{|\mathcal{M}_i \cap \mathcal{M}_j| / |\mathcal{M}_i \cup \mathcal{M}_j|}_{\substack{ \text{merchant overlap}\\ \text{(Jaccard index)}}},\tag{93}\] where \(c_i\) is account \(i\)’s creation date, \(\ell_i\) is its initial credit limit, \(\mathcal{M}_i\) is its merchant set, and \(K_{\mathrm{time}}\), \(K_{\mathrm{limit}}\) are kernel functions measuring proximity.
Under the ring hypothesis (\(H_1\): \(\mathcal{R}\) is a ring):
All accounts in \(\mathcal{R}\) share creation dates (within \(\delta\)), similar credit limits (drawn from \(G_{\mathrm{ring}}\)), and common anchor merchants (\(\mathcal{M}_{\mathrm{ring}}\)). The pairwise similarities are high: \[\label{eq:ring95similarity} \mathbb{E}[S(a_i, a_j) \mid H_1] = \mu_1 > 0 \qquad\text{for all } i \neq j.\tag{94}\]
The expected ring statistic is: \[\label{eq:ring95expect95h1} \mathbb{E}[T_{\mathcal{R}} \mid H_1] = \binom{m}{2} \mu_1 = \frac{m(m-1)}{2}\,\mu_1.\tag{95}\]
Under the null hypothesis (\(H_0\): random legitimate accounts):
Legitimate accounts are created independently. Their creation dates, credit limits, and merchant sets are uncorrelated. The pairwise similarities are small: \[\label{eq:null95similarity} \mathbb{E}[S(a_i, a_j) \mid H_0] = \mu_0 \ll \mu_1.\tag{96}\]
The expected ring statistic is: \[\label{eq:ring95expect95h0} \mathbb{E}[T_{\mathcal{R}} \mid H_0] = \binom{m}{2} \mu_0.\tag{97}\]
Signal-to-noise ratio.
The difference in means grows quadratically with ring size \(m\): \[\label{eq:signal} \mathbb{E}[T_{\mathcal{R}} \mid H_1] - \mathbb{E}[T_{\mathcal{R}} \mid H_0] = \binom{m}{2}(\mu_1 - \mu_0) = O(m^2).\tag{98}\]
The variance of \(T_{\mathcal{R}}\) under either hypothesis is at most \(O(m^2)\) (the sum of \(\binom{m}{2}\) terms with bounded covariance). Therefore the signal-to-noise ratio is: \[\label{eq:snr} \mathrm{SNR} = \frac{\mathbb{E}[T \mid H_1] - \mathbb{E}[T \mid H_0]}{\sqrt{\mathrm{Var}(T \mid H_0)}} = \frac{O(m^2)}{O(m)} = O(m) \;\to\; \infty \quad\text{as } m \to \infty.\tag{99}\]
By the standard relationship between signal-to-noise ratio and AUC for location-shift alternatives (the test statistic \(T\) is approximately normal for large \(m\) by the central limit theorem applied to U-statistics): \[\label{eq:auc95from95snr} \mathop{\mathrm{AUC}}= \Phi(\mathrm{SNR}/\sqrt{2}) \to 1 \quad\text{as } m \to \infty,\tag{100}\] where \(\Phi\) is the standard normal CDF. ◻
Remark 43 (The information is in the cross-account correlations). Theorem 42 has a precise information-theoretic interpretation. Define the mutual information between the bust-out indicator and the observed data: \[\begin{align} I(\theta_a;\, \mathcal{H}_a^{(t)}) &= 0 && \text{(individual account: zero information)}, \tag{101}\\ I(\theta_{\mathcal{R}};\, \{\mathcal{H}_a^{(t)}\}_{a \in \mathcal{R}}) &> 0 \text{ and grows with } m && \text{(ring: positive and increasing)}. \tag{102} \end{align}\]
The fraud signal resides entirely in the cross-account correlations (shared creation dates, shared merchants, correlated behavior) — not in any individual account’s features or transaction history. A fraud detection system that operates at the transaction or account level is provably blind to Class 3 during build-up.
Example 13 (Detection power vs.ring size). Using realistic parameters (\(\mu_1 = 0.15\), \(\mu_0 = 0.01\), and standard deviation of pairwise similarity \(\sigma_S = 0.08\)), the AUC of the ring-level classifier as a function of ring size is:
| Ring size \(m\) | SNR | AUC | Interpretation |
|---|---|---|---|
| 2 | 1.24 | 0.81 | Weak signal |
| 5 | 4.38 | 0.999 | Strong signal |
| 10 | 9.36 | \(>0.9999\) | Near-certain detection |
| 20 | 19.3 | \(\approx 1\) | Effectively perfect |
Even rings of size 5 are detectable with near-perfect accuracy using cross-account signals — while each individual account in the ring has \(\mathop{\mathrm{AUC}}= 0.5\) (random guessing).
The practical implication: invest in graph-based detection infrastructure (entity resolution, link analysis, community detection) rather than in better account-level models for Class 3 fraud.
Remark 44 (The necessity of graph-level methods). Theorem 42 establishes a formal chain of reasoning for Class 3 detection:
Class 3 fraud is undetectable at the transaction level (Theorem 40: no labels).
Class 3 fraud is undetectable at the account level (Theorem 42: identical feature distributions).
Class 3 fraud is detectable at the ring level (Theorem 42(ii): cross-account correlations provide signal).
Therefore, detection requires graph-level methods: entity resolution to link accounts, graph neural networks to aggregate cross-account signals, and community detection to identify ring structures.
A further challenge arises: the transaction graph is itself subject to the observation pipeline — declined transactions create missing edges, unreported fraud creates missing labels on nodes. Extending the semiparametric framework of this paper and [2] to graph-level estimation under edge and node censorship is an important direction for future work.
Proposition 45 (Class 3 bound during exploitation). During the exploitation phase (\(t \geq \tau_a\)), the observation pipeline becomes operative and the STR can be applied. The efficiency bound for estimating the exploitation-phase fraud rate is: \[\label{eq:class395exploit95bound} \sigma^2_{\mathrm{eff}}(3, t \geq \tau_a) = \mathbb{E}\!\left[ \frac{\pi_3(X)(1-\pi_3(X))}{e_0^{(3,\mathrm{expl})} \cdot r_0^{(3,\mathrm{expl})} \cdot p_0^{(3,\mathrm{expl})} \cdot \gamma^{(3)}} \;\bigg|\; t \geq \tau \right].\qquad{(7)}\] However, this bound is of limited practical value because:
The exploitation phase is short (typically 48–72 hours) and involves high-velocity transactions that may trigger automatic blocks (\(e_0^{(3,\mathrm{expl})}\) may be low);
By the time exploitation-phase labels arrive (months later, through charge-off), the damage is done;
The goal of Class 3 detection is to identify bust-out accounts during build-up, before exploitation occurs.
Proof. During exploitation, \(Y^*_{at} = 1\) and the transactions trigger the standard observation pipeline: the issuer eventually identifies the delinquency, the account is charged off, and labels are generated. The STR applies with exploitation-phase-specific propensities. The bound follows from [2] applied to the exploitation subpopulation. ◻
Class 4 represents fraud committed by merchants, where the observation pipeline runs through the acquirer’s monitoring infrastructure rather than the issuer’s chargeback process. This section derives the acquirer-side efficiency bound and shows it is an order of magnitude higher than the Class 1 bound.
Definition 24 (Acquirer monitoring propensity). The acquirer’s monitoring propensity for a merchant \(j\) with features \(H_{\mathrm{merch},j}\) is: \[\label{eq:acquirer95propensity} r_{\mathrm{acq}}^{(4)}(H_{\mathrm{merch},j}) = \mathbb{P}\!\left( \text{acquirer detects and reports merchant } j \mid H_{\mathrm{merch},j},\, Z=4 \right),\tag{103}\] where \(H_{\mathrm{merch},j}\) includes:
The merchant’s chargeback ratio (number of chargebacks divided by total transactions);
The merchant’s transaction volume and average ticket size;
The merchant’s category code (MCC) and risk classification;
The acquirer’s identity and monitoring sophistication;
The merchant’s tenure (time since onboarding);
Whether the merchant has been flagged by network monitoring programs (e.g., Mastercard’s Excessive Chargeback Program, Visa’s Dispute Monitoring Program).
Remark 46 (Why acquirer monitoring is weak). The acquirer’s monitoring propensity \(r_{\mathrm{acq}}^{(4)}\) is typically much lower than the issuer’s reporting propensity \(r_0^{(1)}\) for several structural reasons:
Misaligned incentives. The acquirer earns processing fees from merchant transactions. Reporting a merchant as fraudulent means losing a revenue source. The issuer, by contrast, gains by reporting fraud (recovering chargeback losses).
Information asymmetry. The acquirer sees only the merchant’s transaction patterns, not the underlying cardholder disputes. Individual laundered transactions look normal; the fraud signal emerges only from aggregate merchant-level patterns.
Monitoring infrastructure variation. Large acquirers (e.g., JPMorgan, Worldpay) invest heavily in merchant monitoring. Small acquirers and payment facilitators may have minimal monitoring, relying on network-mandated thresholds (e.g., chargeback ratio \(> 1\%\)) rather than proactive detection.
Slow investigation cycles. Acquirer investigations involve merchant audits, transaction sampling, and often legal review. A typical investigation takes 3–12 months, compared to 30–90 days for issuer chargeback processing.
Definition 25 (Acquirer investigation delay). The acquirer’s investigation delay propensity is: \[\label{eq:acquirer95delay} p_{\mathrm{acq}}^{(4)}(H_{\mathrm{merch},j}) = \mathbb{P}\!\left( \text{investigation completes within training window} \mid H_{\mathrm{merch},j},\, R_{\mathrm{acq}}=1 \right).\tag{104}\]
Theorem 47 (Acquirer-side efficiency bound). The semiparametric efficiency bound for estimating the Class 4 (merchant fraud) rate \(\Psi_4 = \bar\pi_4\) is: \[\label{eq:acquirer95bound} \boxed{ \sigma^2_{\mathrm{eff}}(4) = \mathbb{E}\!\left[ \frac{\pi_4(X)(1-\pi_4(X))}{r_{\mathrm{acq}}^{(4)}(H_{\mathrm{merch}}) \cdot p_{\mathrm{acq}}^{(4)}(H_{\mathrm{merch}}) \cdot \gamma^{(4)}(H_{\mathrm{merch}})} \right], }\tag{105}\] where \(\gamma^{(4)} = (1-\varepsilon_{10}^{(4)}- \varepsilon_{01}^{(4)})^2\) is the corruption penalty for merchant fraud labels.
The authorization propensity \(e_0\) does not appear because \(e_0^{(4)} = 1\) (the merchant controls the terminal and always “approves” the transaction).
Proof. The Class 4 observation model 37 has the form \(O^{(4)} = 1 \cdot R_{\mathrm{acq}} \cdot M_{\mathrm{acq}}\). This is a two-gate sequential model (reporting then delay), which is a special case of
three-gate model in [2] with \(e_0 = 1\).
Applying [2] with \(e_0^{(4)}=1\) and replacing the issuer’s propensities with the acquirer’s propensities gives 105 .
The sequential triple robustness property reduces to sequential double robustness: at each of the two remaining gates (acquirer reporting and acquirer delay), consistency requires only that either the propensity model or the outcome model is correctly specified. ◻
Proposition 48 (Class 4 floor is an order of magnitude higher). The ratio of the Class 4 to Class 1 efficiency bounds is: \[\label{eq:acquirer95ratio} \frac{\sigma^2_{\mathrm{eff}}(4)}{\sigma^2_{\mathrm{eff}}(1)} \;\approx\; \frac{e_0^{(1)} \cdot r_0^{(1)} \cdot p_0^{(1)} \cdot \gamma^{(1)}}{r_{\mathrm{acq}}^{(4)} \cdot p_{\mathrm{acq}}^{(4)} \cdot \gamma^{(4)}},\qquad{(8)}\] assuming comparable fraud prevalence and base-rate variance across classes.
Proof. Direct division of 105 by ?? (with Class 1 parameters), using the approximation \(\pi_k(1-\pi_k) \approx \pi_k\) for small \(\pi_k\) and comparable prevalence. ◻
Example 14 (Quantifying the Class 4 floor).
| Parameter | Class 1 | Class 4 | Source | ||
|---|---|---|---|---|---|
| Authorization (\(e_0\) or equiv.) | 0.85 | 1.00 | Merchant controls terminal | ||
| Reporting (\(r_0\) or \(r_{\mathrm{acq}}\)) | 0.70 | 0.15 | Acquirers monitor weakly | ||
| Delay (\(p_0\) or \(p_{\mathrm{acq}}\)) | 0.85 | 0.40 | Investigations take months | ||
| Corruption (\(\gamma\)) | 0.81 | 0.64 | Merchant fraud harder to classify | ||
| \(q \cdot \gamma\) | 0.41 | 0.038 | |||
| \(1/(q\cdot\gamma)\) | 2.44 | 26.3 | |||
| Ratio | \(26.3 / 2.44 = \mathbf{10.8\times}\) | ||||
The detection floor for merchant fraud is 10.8\(\times\) higher than for third-party card fraud. This means that estimating the merchant fraud rate to a given accuracy requires approximately 11\(\times\) more data than estimating the card-theft rate — or equivalently, for the same sample size, the confidence interval for merchant fraud is \(\sqrt{11} \approx 3.3\times\) wider.
Remark 49 (Implications for network monitoring programs). The acquirer-side bound 105 provides a theoretical foundation for evaluating network monitoring programs. The bound decreases (improves) with higher \(r_{\mathrm{acq}}\) and \(p_{\mathrm{acq}}\). Networks can improve these propensities through:
Mandatory monitoring standards: Requiring all acquirers to implement minimum monitoring capabilities raises \(r_{\mathrm{acq}}\) uniformly.
Chargeback ratio thresholds: Lowering the threshold that triggers merchant investigation (e.g., from 1% to 0.5%) increases the effective \(r_{\mathrm{acq}}\) for marginal merchants.
Faster investigation mandates: Requiring acquirers to complete investigations within 90 days rather than 180 days increases \(p_{\mathrm{acq}}\).
Shared monitoring infrastructure: The network providing centralized monitoring tools (shared merchant risk scores, cross-acquirer pattern detection) raises \(r_{\mathrm{acq}}\) for all acquirers, particularly small ones.
Each of these interventions has a quantifiable effect on the efficiency bound, enabling cost-benefit analysis of monitoring investments. Designing optimal monitoring programs that maximize \(r_{\mathrm{acq}}\) subject to acquirer cost constraints is a mechanism design problem with direct regulatory implications; we discuss this further in Section 11.7.
Proposition 50 (Class 4 STR). The STR for Class 4 is a two-gate estimator with acquirer-side propensities: \[\label{eq:class495str} \hat{\varphi}_t^{(4)} = \hat{\mu}_0^{(4)}(H_{\mathrm{merch}}) + \frac{R_{\mathrm{acq},t}}{\hat{r}_{\mathrm{acq}}^{(4)}(H_{\mathrm{merch}})} \bigl(\hat{\mu}_1^{(4)}(H_{\mathrm{merch}}) - \hat{\mu}_0^{(4)}(H_{\mathrm{merch}})\bigr) + \frac{O^{(4)}_t}{\hat{r}_{\mathrm{acq}}^{(4)} \cdot \hat{p}_{\mathrm{acq}}^{(4)}} \bigl(\tilde{Y}_t^{\mathrm{corr},(4)} - \hat{\mu}_1^{(4)}(H_{\mathrm{merch}})\bigr),\qquad{(9)}\] where \(\hat{\mu}_0^{(4)}\) and \(\hat{\mu}_1^{(4)}\) are the outcome models conditioned on merchant-level features.
This estimator achieves the efficiency bound 105 and inherits sequential double robustness: at each gate (acquirer reporting, acquirer delay), consistency requires only that either the propensity model or the outcome model is correct.
Proof. The two-gate STR is the specialization of the three-gate STR of [2] with \(e_0 = 1\) and \(A_t = 1\) always. The first correction term (authorization) vanishes identically, leaving the reporting and delay corrections. Efficiency and robustness follow from [2] applied to the two-gate submodel. ◻
Remark 51 (Feature set shift). A subtle but important operational difference between the Class 4 STR and the Class 1 STR: the conditioning variables shift from cardholder-level features \((X_t, I_t)\) to merchant-level features \(H_{\mathrm{merch},j}\). This means:
The propensity models must be trained on merchant-level data (chargeback ratios, transaction volume patterns, merchant category, acquirer identity), not cardholder-level data.
The outcome model \(\hat{\mu}_0^{(4)}\) predicts merchant-level fraud probability, not transaction-level fraud probability.
The empirical Bayes shrinkage of [2] applies across acquirers (borrowing monitoring-rate information from well-monitored acquirers to stabilize estimates for poorly-monitored ones), not across issuers.
Class 5 is the fastest-growing fraud category in payment networks, driven by the global surge in authorized push payment (APP) fraud and social engineering scams. This section proves two results: (i) the STR’s authorization correction vanishes, and (ii) the Bayes-optimal transaction-level classifier achieves \(\mathop{\mathrm{AUC}}= 0.5\) when fraud and legitimate feature distributions coincide.
Theorem 52 (Authorization gate degeneracy). When \(e_0^{(5)}(H_0) = 1\) for all \(H_0\) (every transaction is authorized by the victim), the STR’s authorization correction term vanishes identically: \[\label{eq:auth95degen} \frac{A_t}{e_0^{(5)}(H_0)} \bigl(\mu_1^{(5)}(H_1) - \mu_0^{(5)}(H_0)\bigr) = \mu_1^{(5)}(H_1) - \mu_0^{(5)}(H_0),\tag{106}\] and the STR reduces to a two-gate estimator: \[\label{eq:class595str} \hat{\varphi}_t^{(5)} = \hat{\mu}_1^{(5)}(H_1) + \frac{R_{\mathrm{victim},t}}{\hat{r}_{\mathrm{victim}}^{(5)}(H_1)} \bigl(\hat{\mu}_2^{(5)}(H_2) - \hat{\mu}_1^{(5)}(H_1)\bigr) + \frac{O^{(5)}_t}{\hat{r}_{\mathrm{victim}}^{(5)} \cdot \hat{p}_0^{(5)}} \bigl(\tilde{Y}_t^{\mathrm{corr},(5)} - \hat{\mu}_2^{(5)}(H_2)\bigr).\tag{107}\]
The three-gate efficiency bound collapses to a two-gate bound: \[\label{eq:class595bound} \sigma^2_{\mathrm{eff}}(5) = \mathbb{E}\!\left[ \frac{\pi_5(X)(1-\pi_5(X))}{r_{\mathrm{victim}}^{(5)}(H_1) \cdot p_0^{(5)}(H_2) \cdot \gamma^{(5)}} \right].\tag{108}\]
Proof. Step 1: The authorization correction simplifies.
In the full three-gate STR of [2], the first correction term is: \[\frac{A_t}{\hat{e}_0(H_0)} \bigl(\hat{\mu}_1(H_1) - \hat{\mu}_0(H_0)\bigr).\] When \(e_0^{(5)} = 1\) and \(A_t = 1\) always, this becomes \(\hat{\mu}_1(H_1) - \hat{\mu}_0(H_0)\). The STR score is: \[\begin{align} \hat{\varphi}_t^{(5)} &= \hat{\mu}_0(H_0) + (\hat{\mu}_1(H_1) - \hat{\mu}_0(H_0)) + \frac{R_t}{\hat{r}_0(H_1)} (\hat{\mu}_2(H_2) - \hat{\mu}_1(H_1)) + \frac{O_t}{\hat{r}_0 \hat{p}_0} (\tilde{Y}_t^{\mathrm{corr}} - \hat{\mu}_2(H_2)) \\ &= \hat{\mu}_1(H_1) + \frac{R_t}{\hat{r}_0(H_1)} (\hat{\mu}_2(H_2) - \hat{\mu}_1(H_1)) + \frac{O_t}{\hat{r}_0 \hat{p}_0} (\tilde{Y}_t^{\mathrm{corr}} - \hat{\mu}_2(H_2)), \end{align}\] which is 107 .
Step 2: What is lost.
The authorization gate in the three-gate STR serves two purposes:
Selection correction: Adjusting for the fact that declined transactions are unobserved.
Information exploitation: Using post-authorization signals \(W_1\) through \(\mu_1(H_1) \neq \mu_0(H_0)\).
Purpose (a) is unnecessary when \(e_0 = 1\) (nothing is declined). Purpose (b) still applies if there are informative post-authorization signals \(W_1\). However, for Class 5, the post-authorization signals are typically uninformative (the transaction looks normal because the victim performed it), so \(\mu_1(H_1) \approx \mu_0(H_0)\) and even purpose (b) contributes minimally.
Step 3: The efficiency bound.
The bound follows from [2] with \(e_0 = 1\), which eliminates the authorization component from the efficiency bound’s denominator. ◻
Remark 53 (The remaining two gates are severely impaired). Even after the authorization gate vanishes, the remaining two gates are impaired for Class 5:
Reporting: \(r_{\mathrm{victim}}^{(5)} \approx 0.15\text{--}0.30\). Victims are slow to realize they were scammed (days to months), and psychological barriers (shame, embarrassment, perceived futility) suppress reporting. Romance scam victims may never report.
Corruption: \(\varepsilon_{10}^{(5)}\) is elevated because even when victims report, the transaction was authorized, so the issuer may classify it as “not fraud” (the cardholder consented to the payment).
The combined effect: \(r_{\mathrm{victim}}^{(5)} \cdot \gamma^{(5)}\) is small, making the Class 5 bound high relative to Class 1.
Theorem 54 (Feature non-informativeness). When the transaction feature distribution is identical for fraud and legitimate transactions: \[\label{eq:feature95identical} P(X_t \mid Y^*_t = 1,\, Z_t = 5) = P(X_t \mid Y^*_t = 0),\tag{109}\] the Bayes-optimal transaction-level classifier achieves \(\mathop{\mathrm{AUC}}= 0.5\): \[\label{eq:auc95half} \mathop{\mathrm{AUC}}\!\bigl(\hat{f}^{\mathrm{Bayes}}_0\bigr) = 0.5.\tag{110}\] Moreover, the outcome model in the STR degenerates to a constant: \[\label{eq:mu95constant} \mu_0^{(5)}(H_0) = \mathbb{E}[Y^* \mid H_0,\, Z=5] = \frac{\pi_5(X) \cdot P(X)}{\pi_5(X) \cdot P(X) + (1-\pi_5(X)) \cdot P(X)} = \pi_5(X),\tag{111}\] and when \(\pi_5\) does not vary with \(X\) (scam targeting is independent of transaction features), \(\mu_0^{(5)} = \bar\pi_5\) is a constant.
Proof. Step 1: The Bayes-optimal classifier.
The Bayes-optimal classifier predicts: \[\label{eq:bayes95classifier} f_0^{(5)}(x) = \mathbb{P}(Y^*=1 \mid X=x,\, Z=5) = \frac{\pi_5(x) \cdot p(x \mid Y^*=1, Z=5)}{p(x \mid Z=5)}.\tag{112}\]
Under 109 : \[\label{eq:bayes95simplify} p(x \mid Y^*=1, Z=5) = p(x \mid Y^*=0) = p(x),\tag{113}\] so: \[f_0^{(5)}(x) = \frac{\pi_5(x) \cdot p(x)}{\pi_5(x) \cdot p(x) + (1-\pi_5(x)) \cdot p(x)} = \pi_5(x).\]
When \(\pi_5(x) = \bar\pi_5\) (constant), the classifier assigns the same score \(\bar\pi_5\) to every transaction.
Step 2: AUC = 0.5.
A constant classifier produces no ranking — all transactions receive the same score. The ROC curve is the diagonal from \((0,0)\) to \((1,1)\), and: \[\mathop{\mathrm{AUC}}= \mathbb{P}(\hat{f}_0(X^+) > \hat{f}_0(X^-)) + \tfrac{1}{2}\mathbb{P}(\hat{f}_0(X^+) = \hat{f}_0(X^-)) = 0 + \tfrac{1}{2} \cdot 1 = 0.5,\] where \(X^+\) is a fraud transaction and \(X^-\) is a legitimate transaction.
Step 3: Outcome model degeneracy.
The outcome model \(\mu_0^{(5)}(H_0) = \mathbb{E}[Y^* \mid H_0, Z=5]\) is the posterior fraud probability given the pre-authorization history. Under 109 , the posterior equals the prior: \(\mu_0^{(5)}(H_0) = \pi_5(X)\). When the prior does not depend on \(X\), \(\mu_0^{(5)} = \bar\pi_5\).
This eliminates the outcome model’s ability to reduce variance in the STR. In the standard STR, the outcome model provides a “baseline prediction” that reduces the residual variance of the inverse-propensity-weighted correction. When the outcome model is constant, this variance reduction vanishes entirely, and the STR reduces to a pure IPW estimator — the highest-variance form of the estimator. ◻
Remark 55 (The feature non-informativeness condition is approximately satisfied for scams). Condition 109 is an idealization. In practice, scam transactions may have subtle differences from legitimate transactions:
Slightly unusual amounts: Scam payments may cluster at round numbers ($5,000, $10,000) more than typical spending.
Unusual recipients: The payee may be a recently created account, a cryptocurrency exchange, or an overseas entity.
Temporal patterns: Scam payments may occur after an extended phone call or a series of incoming messages (detectable from session metadata).
These signals provide some discriminative power (\(\mathop{\mathrm{AUC}}> 0.5\)), but they are weak compared to the strong signals available for Class 1 fraud (unusual device, unusual location, unusual merchant, velocity anomalies). The feature non-informativeness theorem characterizes the limiting case and explains why Class 5 detection is fundamentally harder than Class 1 detection even with unlimited data.
Definition 26 (Augmented feature set). The augmented feature set for Class 5 is \(\tilde{X}_t = (X_t, X_{\mathrm{aux},t})\), where \(X_{\mathrm{aux},t}\) includes signals outside the standard transaction feature set:
Communication signals: Whether the victim was on a phone call during the transaction, duration of the call, whether the caller’s number is known or unknown.
Session behavior: Hesitation patterns (pauses, backtracking, error corrections), typing speed anomalies, navigation patterns suggesting external instruction.
Recipient analysis: Age of the recipient account, recipient’s transaction history, recipient’s risk score, whether the recipient has received payments from multiple first-time senders (“money mule” indicator).
Device interaction: Whether the victim’s screen was being shared or mirrored (common in tech support scams), whether a remote access tool is active.
Temporal context: Time since the victim’s last incoming call or message, correlation between communication events and payment initiation.
Proposition 56 (Augmented features restore detectability). If the augmented feature distribution differs between fraud and legitimate transactions: \[\label{eq:augmented95informative} P(\tilde{X}_t \mid Y^*_t = 1,\, Z_t = 5) \neq P(\tilde{X}_t \mid Y^*_t = 0),\qquad{(10)}\] then the Bayes-optimal classifier on \(\tilde{X}_t\) achieves \(\mathop{\mathrm{AUC}}> 0.5\), and the augmented outcome model \(\tilde{\mu}_0^{(5)}(\tilde{H}_0)\) provides non-trivial variance reduction in the STR.
The AUC improvement depends on the strength of the auxiliary signal: \[\label{eq:augmented95auc} \mathop{\mathrm{AUC}}(\tilde{f}_0^{(5)}) = \Phi\!\left( \frac{d(\tilde{F}_1, \tilde{F}_0)}{\sqrt{2}} \right),\qquad{(11)}\] where \(d(\tilde{F}_1, \tilde{F}_0)\) is the Mahalanobis distance between the augmented feature distributions under fraud (\(\tilde{F}_1\)) and legitimate (\(\tilde{F}_0\)), and \(\Phi\) is the standard normal CDF.
Proof. When ?? holds, the augmented Bayes-optimal classifier is: \[\tilde{f}_0^{(5)}(\tilde{x}) = \frac{\pi_5 \cdot \tilde{p}(\tilde{x} \mid Y^*=1, Z=5)}{\tilde{p}(\tilde{x})} \neq \pi_5 \quad\text{(not constant)}.\] A non-constant classifier has \(\mathop{\mathrm{AUC}}> 0.5\) by definition. The AUC formula ?? follows from the standard result for Gaussian location models (the likelihood ratio is monotone in the Mahalanobis distance, and the AUC of a monotone score is determined by the separation of the score distributions). ◻
Example 15 (What auxiliary signals can achieve). Consider a scam detection system with access to:
| Feature set | Estimated AUC | Source of signal |
|---|---|---|
| Transaction features only (\(X_t\)) | 0.52–0.55 | Weak: amount, recipient |
| \(+\) Recipient risk score | 0.62–0.68 | Money mule indicators |
| \(+\) Communication metadata | 0.72–0.78 | Phone call during payment |
| \(+\) Session behavior | 0.78–0.85 | Hesitation, external instruction |
| \(+\) Device sharing detection | 0.85–0.92 | Screen sharing, remote access |
| Full augmented set (\(\tilde{X}_t\)) | 0.85–0.92 | All of the above |
Transaction features alone give near-random performance (\(\mathop{\mathrm{AUC}}\approx 0.53\)), confirming Theorem 54. Each auxiliary signal category adds 5–10 AUC points. The full augmented feature set achieves \(\mathop{\mathrm{AUC}}\approx 0.88\) — usable but still substantially below the \(\mathop{\mathrm{AUC}}> 0.95\) achievable for Class 1 fraud with transaction features alone.
The theoretical message: for Class 5, the bottleneck is the feature set, not the model or the labels. Investing in auxiliary signal collection (communication metadata, session analytics, recipient scoring) has higher marginal value than investing in better models or better label correction.
Proposition 57 (Augmented Class 5 efficiency bound). With the augmented feature set \(\tilde{X}_t\), the Class 5 efficiency bound becomes: \[\label{eq:class595augmented95bound} \sigma^2_{\mathrm{eff}}(5, \tilde{X}) = \mathbb{E}\!\left[ \frac{\tilde{\pi}_5(\tilde{X}) (1-\tilde{\pi}_5(\tilde{X}))}{r_{\mathrm{victim}}^{(5)}(\tilde{H}_1) \cdot p_0^{(5)}(\tilde{H}_2) \cdot \gamma^{(5)}} \right],\qquad{(12)}\] where \(\tilde{\pi}_5(\tilde{x}) = \mathbb{P}(Z=5 \mid \tilde{X} = \tilde{x})\) is the type probability conditional on the augmented features.
This bound is strictly lower than the non-augmented bound 108 whenever the auxiliary features are informative: \[\label{eq:augmented95improvement} \sigma^2_{\mathrm{eff}}(5, \tilde{X}) < \sigma^2_{\mathrm{eff}}(5, X) \qquad\text{when } P(\tilde{X} \mid Z=5) \neq P(\tilde{X} \mid Z \neq 5).\qquad{(13)}\]
Proof. The augmented features allow the outcome model to be non-constant: \(\tilde{\mu}_0^{(5)}(\tilde{H}_0) \neq \bar\pi_5\). The non-constant outcome model reduces the residual variance in the STR’s inverse-propensity-weighted correction. Formally, the efficiency bound is the variance of the efficient influence function, which decreases when the outcome model has better predictive power (lower residual variance).
The strict inequality follows from the fact that \(\mathrm{Var}(\tilde{\pi}_5(\tilde{X})) > \mathrm{Var}(\pi_5(X))\) when the auxiliary features carry information, which increases the “explained variance” component and decreases the “residual” component of the efficiency bound. ◻
Remark 58 (The investment priority for Class 5). The results of this section establish a clear investment priority for Class 5 fraud:
First: collect auxiliary signals. The feature non-informativeness theorem (Theorem 54) shows that transaction features alone are useless. The largest marginal gain comes from adding communication metadata, recipient scoring, and session behavior to the feature set.
Second: improve reporting. The victim reporting rate \(r_{\mathrm{victim}}^{(5)}\) is the binding constraint on label quality. Reducing shame barriers (education campaigns), simplifying reporting processes, and extending chargeback rights to authorized transactions all increase \(r_{\mathrm{victim}}^{(5)}\) and lower the efficiency bound.
Third: apply the (augmented) STR. Once auxiliary features and improved reporting are in place, the two-gate STR 107 with augmented features achieves the efficiency bound ?? .
This ordering reverses the typical fraud-detection investment priority (which focuses on models first, features second, and labels third). For Class 5, the binding constraints are features and labels, not models.
The decomposed STR requires estimating the type membership probabilities \(\hat{\pi}_k(x)\) (Definition 17). In practice, type identification relies on mature, fully resolved labels (Assumption 3).
Remark 59 (Type identification data sources). The following data sources support type identification for each class:
Class 1 (third-party fraud): Chargeback reason codes (e.g., Mastercard 4837 “no cardholder authorization,” Visa 10.4 “other fraud – card absent environment”) directly identify third-party fraud. Confirmation comes from cardholder attestation that the transaction was unauthorized.
Class 2 (first-party fraud): Identified through representment outcomes (the merchant provides proof of delivery and the chargeback is reversed), repeated dispute patterns on the same account, and specialized first-party fraud models that flag inconsistencies between the dispute claim and the transaction evidence.
Class 3 (bust-out): Identified by the charge-off event combined with the account’s credit history pattern: rapid credit utilization increase, payment cessation, and (for synthetic identities) the absence of a verifiable identity trail.
Class 4 (merchant fraud): Identified through network monitoring programs (MATCH, VMAS), acquirer investigation outcomes, and law enforcement referrals.
Class 5 (scams): Identified through victim self-reporting (“I was tricked into sending money”), police reports, and specialized scam-detection flags in the dispute process.
Remark 60 (Type misclassification and its effects). In practice, type identification is imperfect. The most common misclassifications are:
Class 1 \(\leftrightarrow\) Class 2: Distinguishing third-party fraud from first-party fraud is inherently difficult — both produce chargebacks, and the issuer cannot directly observe the cardholder’s intent. Misclassification between these two classes is the most frequent error.
Class 3 \(\to\) Class 1: During the exploitation phase, bust-out transactions generate chargebacks that resemble Class 1 fraud. Without examining the account’s full credit history, the exploitation-phase chargebacks may be misclassified as standard third-party fraud.
Class 5 \(\to\) “not fraud”: Because the victim authorized the transaction, issuers may classify scam losses as “not fraud,” leading to underestimation of Class 5 prevalence.
The effect of type misclassification on the decomposed STR is analogous to the effect of propensity model misspecification in the pooled STR: it introduces bias that is controlled by the robustness properties. Specifically, if the type classifier achieves accuracy \(1-\epsilon_Z\) (fraction of transactions correctly classified), the additional bias introduced by misclassification is \(O(\epsilon_Z)\) and the additional variance is \(O(\epsilon_Z^2)\). The decomposed STR dominates the pooled STR as long as: \[\label{eq:misclass95threshold} \epsilon_Z < 1 - \frac{\sigma^2_{\mathrm{decomp}}}{\sigma^2_{\mathrm{pooled}}},\tag{114}\] i.e., the type classification error is smaller than the relative efficiency gain from decomposition. For the parameters of Example 9 (36% efficiency gain), the decomposed STR dominates as long as the type classifier is better than 64% accurate — a very weak requirement.
Remark 61 (Is five the right number?). The five-class partition is derived from the structural signature (Definition 9) and proved to be minimal and complete (Theorem 12). However, the partition depends on the granularity of the structural questions Q1–Q5. Two considerations:
Coarser partitions. One could merge Classes 4 and 5 (both have \(e_0 = 1\)) or merge Classes 1 and 5 (both have victim-generated labels). Theorem 11 shows that every such merger loses a structural distinction that changes the efficient estimator. Coarser partitions are provably suboptimal.
Finer partitions. One could split Class 1 into card-present and card-not-present subtypes, or split Class 5 into APP fraud and romance scams. Proposition 26 shows that within-class decomposition yields diminishing returns when the subtypes share the same pipeline topology. Finer partitions increase modeling complexity without proportionate efficiency gains.
The five-class partition represents the bias-complexity tradeoff: it captures all structurally distinct observation mechanisms (eliminating the Jensen penalty from pooling heterogeneous pipelines) without introducing unnecessary modeling complexity from splitting homogeneous pipelines.
Remark 62 (Cross-class transactions). Some fraud schemes involve elements of multiple classes. For example:
Bust-out via collusive merchant (Class 3 \(\cap\) Class 4): A synthetic identity builds credit (Class 3) and then exploits it through a collusive merchant that launders the proceeds (Class 4). The observation pipeline has both the temporal two-phase structure of Class 3 and the acquirer-side monitoring of Class 4.
Scam followed by friendly fraud (Class 5 \(\to\) Class 2): A victim is scammed (Class 5) and then, realizing the loss, files chargebacks on other legitimate transactions to recover the loss (Class 2). The same account generates both Class 5 and Class 2 fraud.
Account takeover with authorized transactions (Class 1 \(\to\) Class 5): An account takeover (Class 1) in which the criminal socially engineers the victim into authorizing transactions themselves (e.g., “your account is compromised, please transfer funds to this safe account”). The early transactions are unauthorized (Class 1); the later ones are victim-authorized (Class 5).
These cross-class cases are handled by the soft type assignment: the type probabilities \(\hat{\pi}_k(x)\) can assign positive probability to multiple classes simultaneously. The decomposed STR then applies a weighted combination of class-specific corrections. The theoretical guarantees hold as long as the true type distribution is in the support of the estimated distribution.
Remark 63 (The fraud mix changes over time). The prevalence \(\bar\pi_k\) of each fraud class is not static. Industry trends show:
Class 1 (third-party card fraud) is declining in relative terms, due to chip-and-PIN adoption, 3D Secure, and tokenization.
Class 2 (first-party fraud) is increasing, driven by frictionless dispute processes and the growth of e-commerce returns.
Class 3 (bust-out/synthetic identity) is increasing, driven by the availability of compromised personal data and the growth of digital-only banking.
Class 5 (scams) is rapidly increasing, particularly authorized push payment fraud, driven by social media, cryptocurrency, and real-time payment systems.
The Jensen penalty (Theorem 18) changes with the type distribution: as poorly-observed classes (3, 4, 5) grow in prevalence, the penalty from pooling increases. This means that the value of decomposition is increasing over time as the fraud landscape shifts toward harder-to-observe fraud types.
Remark 64 (Practical estimation for Class 2). Theorem 31 requires estimating the dispute game’s primitives \((\alpha, c, \beta^*, p, T)\). In practice:
\(\alpha(x)\) (chargeback success rate) is directly observable from historical dispute outcomes, stratified by transaction features.
\(\beta^*(x)\) (first-party fraud detection rate) is observable from the fraction of disputes that are successfully challenged via representment, stratified by transaction features.
\(p(x,a)\) (penalty) is set by issuer policy and is therefore known.
\(c(x)\) (dispute cost) and \(T\) (rationality) are not directly observable but can be estimated from the observed dispute rate as a function of transaction amount. The logistic model 72 implies: \[\label{eq:logit95estimation} \log\!\left(\frac{d^*}{1-d^*}\right) = \frac{\alpha a - c - \beta^* p}{T},\tag{115}\] which is a linear model in \(a\) (conditional on \(x\)) with slope \(\alpha/T\) and intercept \(-(c+\beta^* p)/T\). Standard logistic regression on the observed dispute indicator \(D_t\) against the transaction amount \(a_t\) identifies \(c/T\) and \(1/T\) (and hence \(c\) and \(T\) separately if the scale of \(T\) is normalized).
Remark 65 (Key limitations).
Type observability assumption. Assumption 3 requires that fraud types are identifiable for mature, fully resolved transactions. This is approximately true for Classes 1–3 (chargeback reason codes, charge-off indicators) but weaker for Classes 4–5 (acquirer investigations and scam reports are less standardized). Misclassification between Classes 4 and 5 is the most consequential limitation.
Independence of types. The decomposition dominance theorem (Theorem 16) assumes that the type-specific STRs are asymptotically independent. In practice, a single transaction can have elements of multiple classes (Remark 62), introducing dependence. The soft type assignment handles this approximately, but a formal treatment of overlapping classes is deferred to future work.
Static game for Class 2. The dispute game (Section 7) is a one-shot game. In practice, the cardholder and issuer interact repeatedly, and the cardholder learns from past dispute outcomes. A dynamic (repeated game) formulation would capture reputation effects and learning, but is substantially more complex.
Constructive DGP for Class 3. Theorem 42 proves insufficiency by constructing a specific adversarial DGP. Real bust-out accounts may not perfectly mimic legitimate behavior — subtle behavioral differences (slightly higher velocity, slightly different merchant mix) may provide weak individual-level signal. The theorem establishes a lower bound on the difficulty, not an exact characterization.
No empirical validation. This paper is purely theoretical. Empirical validation of the Jensen penalty magnitude, the dispute game equilibrium, and the ring-level detection power on real payment network data is an important direction for future work.
Remark 66 (Open directions). Several extensions of this work are natural:
Economics of the observation pipeline. The dispute game (Theorem 31) shows that the corruption rate in Class 2 is an economic equilibrium, not a statistical parameter. The same logic applies to all five classes: each actor in the payment network (issuer, acquirer, merchant, cardholder) has economic incentives that shape the observation pipeline’s propensities. Deriving optimal mechanism designs — incentive structures that maximize information quality across all five classes subject to each actor’s participation constraints — is an important open problem.
Adversarial dynamics. The Class 2 analysis (Section 7) models one form of strategic adaptation: dispute behavior responds to detection capability. More broadly, fraudsters across all classes adapt to detection models by shifting tactics, evading features, and exploiting blind spots. Modeling the co-evolution of detection and evasion as a dynamic game would extend the static equilibrium of Theorem 31 to a richer setting.
Graph-level detection under censorship. Theorem 42 shows that Class 3 detection requires cross-account graph signals. However, the transaction graph is itself subject to the observation pipeline: declined transactions create missing edges, and unreported fraud creates missing node labels. Extending the semiparametric framework of [2] to graph-level estimation under edge and node censorship is a natural next step.
Dynamic type assignment. The type probabilities \(\pi_k(x)\) are currently estimated from a static snapshot. A dynamic model that updates type probabilities as the fraud landscape evolves (e.g., using online learning or Bayesian updating) would improve the decomposed STR’s adaptiveness.
Repeated dispute game. Extending Theorem 31 to a repeated game would capture the cardholder’s learning (“my last dispute succeeded, so I’ll try again”) and the issuer’s reputation building (“this cardholder has disputed three times, so flag future disputes”).
Optimal monitoring design for Class 4. The acquirer-side bound (Theorem 47) shows that \(r_{\mathrm{acq}}\) is the bottleneck for merchant fraud detection. Designing monitoring programs that maximize \(r_{\mathrm{acq}}\) subject to acquirer cost constraints is a mechanism design problem with direct regulatory implications.
Cross-class information sharing. The current framework treats the five classes independently. In practice, information from one class can improve detection in another (e.g., a merchant flagged for Class 4 fraud may also be associated with Class 3 synthetic identities). Formalizing these cross-class information flows is an open problem.
Empirical validation. Computing the Jensen penalty from real network data, estimating the dispute game equilibrium from observed dispute rates, and measuring ring-level detection power for known bust-out rings would validate the theoretical results.
This paper establishes three results.
First, the industry’s extensive fraud taxonomy collapses into exactly five observation-mechanism classes, each defined by a structurally distinct censorship pipeline. The classification is minimal (no two classes can be merged without losing estimation-relevant structural distinctions) and complete (every fraud type maps to exactly one class). The classification criterion is the structural signature — a 5-tuple characterizing the perpetrator–cardholder relationship, the label-generating agent, the authorization gate informativeness, the reporting gate structure, and the temporal observability — and two fraud types belong to the same class if and only if the Sequential Triply Robust estimator of [2] has the same functional form for both.
Second, treating the heterogeneous fraud mixture as homogeneous is provably suboptimal. The type-decomposed STR — which applies class-specific propensities and aggregates — strictly dominates the pooled STR in mean squared error. The efficiency gap is a Jensen penalty from averaging heterogeneous inverse-propensity weights, and is substantial for realistic network parameters (approximately 36% efficiency waste in our numerical illustration, dominated by the poorly-observed Class 3).
Third, each class has a unique structural feature that requires a different theoretical treatment:
Class 1 (victim-reported institutional): the STR of [2] applies directly with type-specific propensities.
Class 2 (adversarial label generation): the false-positive corruption rate is the Nash equilibrium of a dispute game, with comparative statics showing that corruption increases with transaction amount and decreases with dispute friction and detection capability. Multiple equilibria are possible, with discontinuous jumps between low-fraud and high-fraud regimes.
Class 3 (deferred observability): the STR’s positivity assumption fails structurally during the bust-out build-up phase, and no account-level classifier exceeds random guessing — detection requires cross-account graph signals whose power grows with ring size.
Class 4 (acquirer-observed): the detection floor depends on the acquirer’s monitoring propensity rather than the issuer’s reporting propensity, and is an order of magnitude higher.
Class 5 (victim-authorized with non-informative features): the authorization correction vanishes, the Bayes-optimal classifier achieves \(\mathop{\mathrm{AUC}}= 0.5\) on transaction features alone, and detection requires auxiliary signals outside the payment data stream.
Together with the companion papers on information-theoretic limits [1] and optimal label recovery [2], this paper completes the theoretical foundation for fraud detection in payment networks: [1] proves the floor exists, [2] achieves it under homogeneity, and the present paper decomposes it by fraud type, showing that the homogeneous treatment is provably suboptimal and that each observation-mechanism class demands its own detection strategy.
The decomposition dominance result follows from a general principle in semiparametric efficiency theory: conditioning on additional information cannot increase the efficiency bound.
Let \(\mathcal{M}_{\mathrm{pooled}}\) denote the semiparametric model in which the observation propensity \(q_0\) and corruption \(\gamma\) are arbitrary functions of \((H_0, H_2)\), with the fraud type \(Z\) treated as latent. Let \(\mathcal{M}_{\mathrm{decomp}}\) denote the model in which the propensity and corruption are parameterized type-conditionally: \(q_0^{(k)}(H_0)\) and \(\gamma^{(k)}(H_2)\) for each \(k\), with \(Z\) observed (or estimated).
The decomposed model conditions on strictly more information (\(Z\) in addition to \(H_0, H_2\)). By the general result of [11], the efficient influence function in a model that conditions on additional information has variance no larger than the efficient influence function in the marginal model: \[\label{eq:submodel95detail} \mathrm{Var}(\varphi_{\mathrm{eff}}^{\mathrm{decomp}}) = \mathbb{E}[\mathrm{Var}(\varphi_{\mathrm{eff}}^{\mathrm{pooled}} \mid Z)] \leq \mathrm{Var}(\varphi_{\mathrm{eff}}^{\mathrm{pooled}}),\tag{116}\] where the inequality is the law of total variance (the total variance equals the expected conditional variance plus the variance of the conditional mean; the left-hand side is only the first term).
The gap is: \[\label{eq:gap95detail} \mathrm{Var}(\varphi_{\mathrm{eff}}^{\mathrm{pooled}}) - \mathrm{Var}(\varphi_{\mathrm{eff}}^{\mathrm{decomp}}) = \mathrm{Var}_Z(\mathbb{E}[\varphi_{\mathrm{eff}}^{\mathrm{pooled}} \mid Z]) \geq 0,\tag{117}\] with equality iff \(\mathbb{E}[\varphi_{\mathrm{eff}}^{\mathrm{pooled}} \mid Z]\) is constant in \(Z\) — i.e., the efficient influence function does not depend on the fraud type, which requires identical propensity structures across classes.
The ring cohesion statistic \(T_{\mathcal{R}} = \sum_{i<j} S(a_i, a_j)\) is a U-statistic of order 2. Under the null hypothesis (\(H_0\): independent legitimate accounts), the variance of \(T_{\mathcal{R}}\) is: \[\label{eq:ustat95var} \mathrm{Var}(T_{\mathcal{R}} \mid H_0) = \binom{m}{2}\mathrm{Var}(S(a_1,a_2)) + 4\binom{m}{3}\mathrm{Cov}(S(a_1,a_2), S(a_1,a_3)),\tag{118}\] where the covariance term arises from pairs sharing a common account.
Under the independence assumption of \(H_0\), the covariance term is small relative to the first term (it arises only from the shared account’s marginal effect on the kernel \(S\)). Therefore: \[\label{eq:ustat95order} \mathrm{Var}(T_{\mathcal{R}} \mid H_0) = O(m^2),\tag{119}\] and the signal-to-noise ratio is: \[\mathrm{SNR} = \frac{O(m^2)}{O(m)} = O(m) \to \infty.\]
Under \(H_1\) (the ring hypothesis), the pairwise similarities are correlated through the shared ring structure, and \(\mathrm{Var}(T_{\mathcal{R}} \mid H_1) = O(m^2)\) as well. The SNR calculation uses the \(H_0\) variance as the denominator (standard in hypothesis testing), giving \(\mathrm{SNR} = O(m)\) as stated in the proof of Theorem 42.
| Symbol | Meaning |
|---|---|
| Symbol | Meaning |
| \(X_t \in \mathcal{X}\) | Transaction features |
| \(I_t \in \mathcal{I}\) | Issuer identity |
| \(Y^*_t \in \{0,1\}\) | Latent true fraud indicator |
| \(Z_t \in \{0,1,\ldots,K\}\) | Latent fraud type (\(0\) = legitimate, \(k \geq 1\) = fraud type \(k\)) |
| \(A_t\) | Authorization indicator |
| \(R_t\) | Reporting indicator |
| \(M_t\) | Maturity indicator |
| \(O_t = A_t R_t M_t\) | Observation indicator |
| \(\tilde{Y}_t\) | Observed label (when \(O_t=1\)) |
| \(\tilde{Y}_t^{\mathrm{corr},(k)}\) | Corruption-corrected label for class \(k\) (eq. ?? ) |
| Stage histories | |
| \(H_0 = (X, I, \Delta)\) | Pre-authorization history |
| \(H_1 = (H_0, W_1)\) | Post-authorization history |
| \(H_2 = (H_1, W_2)\) | Post-reporting history |
| \(H_{\mathrm{merch}}\) | Merchant-level features (Class 4) |
| \(\mathcal{H}_a^{(t)}\) | Account \(a\)’s history up to time \(t\) (Class 3) |
| Type distribution | |
| \(\pi_k(x) = \mathbb{P}(Z=k \mid X=x)\) | Type probability (Def. 2) |
| \(\bar\pi_k = \mathbb{E}[\pi_k(X)]\) | Marginal type prevalence |
| \(K\) | Number of fraud types |
| Type-specific propensities (Def. 4) | |
| \(e_0^{(k)}(H_0)\) | Authorization propensity for type \(k\) |
| \(r_0^{(k)}(H_1)\) | Reporting propensity for type \(k\) |
| \(p_0^{(k)}(H_2)\) | Delay propensity for type \(k\) |
| \(q_0^{(k)} = e_0^{(k)} r_0^{(k)} p_0^{(k)}\) | Total observation propensity for type \(k\) |
| \(\varepsilon_{10}^{(k)}\) | False-negative corruption for type \(k\) |
| \(\varepsilon_{01}^{(k)}\) | False-positive corruption for type \(k\) |
| \(\gamma^{(k)} = (1-\varepsilon_{10}^{(k)}- \varepsilon_{01}^{(k)})^2\) | Corruption penalty for type \(k\) |
| Class-specific parameters | |
| \(r_{\mathrm{acq}}^{(4)}\) | Acquirer monitoring propensity (Class 4) |
| \(p_{\mathrm{acq}}^{(4)}\) | Acquirer investigation delay (Class 4) |
| \(r_{\mathrm{victim}}^{(5)}\) | Victim reporting propensity (Class 5) |
| \(D_t \in \{0,1\}\) | Strategic dispute decision (Class 2, eq. 31 ) |
| \(\theta_a \in \{0,1\}\) | Account-level bust-out indicator (Class 3, eq. 82 ) |
| \(\tau_a\) | Bust-out time (Class 3, Def. 23) |
| Structural signature (Def. 9) | |
| \(\mathcal{S}(k)\) | Structural signature of fraud type \(k\) |
| \(\mathrm{Q1}\)–\(\mathrm{Q5}\) | Structural classification questions |
| \(\mathcal{S}^{(c)}\) | Signature of class \(c \in \{1,\ldots,5\}\) |
| Efficiency bounds | |
| \(\sigma^2_{\mathrm{eff}}(k)\) | Semiparametric efficiency bound for type \(k\) (eq. 14 ) |
| \(\sigma^2_{\mathrm{pooled}}\) | Efficiency bound for pooled STR (eq. 46 ) |
| \(\sigma^2_{\mathrm{decomp}}\) | Efficiency bound for decomposed STR (eq. 50 ) |
| \(\Delta\) | Jensen penalty (Thm. 18, eq. 57 ) |
| Estimators | |
| \(\hat{\Psi}_{\mathrm{pooled}}\) | Pooled STR estimator (Def. 16) |
| \(\hat{\Psi}_{\mathrm{decomp}}\) | Type-decomposed STR estimator (Def. 17) |
| \(\hat{\varphi}_t^{(k)}\) | STR score for class \(k\) |
| \(\hat{\pi}_k(X_t)\) | Estimated type probability |
| Dispute game (Section 7) | |
| \(\alpha(x)\) | Chargeback success probability (eq. 67 ) |
| \(c(x)\) | Dispute filing cost (eq. 68 ) |
| \(\beta(x)\) | First-party fraud detection probability (eq. 69 ) |
| \(p(x,a)\) | Penalty for identification (eq. 70 ) |
| \(T\) | Cardholder rationality parameter (Def. 21) |
| \(d^*(x,a;\beta)\) | Cardholder’s equilibrium dispute probability (eq. 72 ) |
| \(\Phi(\varepsilon)\) | Fixed-point mapping for equilibrium computation (eq. 80 ) |
| \(U_C, U_I\) | Cardholder and issuer payoffs (Defs. 19, 20) |
| \(\ell_{\mathrm{FP}}\) | Cost of falsely accusing honest cardholder (eq. 71 ) |
| Ring detection (Section 8) | |
| \(T_{\mathcal{R}}\) | Ring cohesion statistic (eq. 92 ) |
| \(S(a_i, a_j)\) | Pairwise account similarity (eq. 93 ) |
| \(\mathcal{R}\) | Candidate ring of accounts |
| \(m = |\mathcal{R}|\) | Ring size |
| \(\mu_1, \mu_0\) | Expected pairwise similarity under ring (\(H_1\)) and null (\(H_0\)) hypotheses |
| Augmented features (Section 10) | |
| \(\tilde{X}_t = (X_t, X_{\mathrm{aux},t})\) | Augmented feature set (Def. 26) |
| \(X_{\mathrm{aux},t}\) | Auxiliary (non-transaction) features |
| \(\tilde{\pi}_5(\tilde{x})\) | Type probability conditional on augmented features |
| Target estimands | |
| \(\Psi = \mathbb{E}[Y^*]\) | Population fraud rate |
| \(\Psi_k = \bar\pi_k\) | Type-\(k\) fraud rate |