1 Introduction↩︎

Financial time series exhibit reproducible regime shifts: discrete intervals during which the data-generating process is approximately stable, separated by transitions arising from external structural change. US monetary policy makes the examples concrete. Around the December-2015 lift-off, the Federal Reserve raised rates for the first time since the global financial crisis and exited the zero lower bound: the 2010–2015 quantitative-easing regime, in which the Fed-controlled term premium drove the yield curve, gave way to dynamics increasingly set by market-driven inflation expectations. Conversely, between March 2020 and early 2022 the policy rate returned to the zero lower bound under the COVID-19 emergency response, with large-scale asset purchases (quantitative easing) again the binding instrument: term premium, dealer balance-sheet capacity, and forward-guidance text drove yield-curve dynamics while the conventional short-rate channel was effectively shut.

The phenomenon is not confined to these episodes: the stock–bond correlation flipped sign as inflation re-emerged in 2022 [1], and the canonical flight-to-quality channel from equities to long-dated Treasuries broke down within days in March 2020 [2]. Each is, in time-series terms, a structural break in the data-generating process. Locating these breaks accurately is foundational to understanding financial markets: the interactions among yields, term premium, and macro factors are stable, and therefore estimable, only within a single regime, so a misplaced boundary blends two distinct dependence structures into one. For example, Figure 1 shows that around the December-2015 lift-off the bond market’s lagged dependence structure reorganises outright: the directed-acyclic-graph (DAG) root recovered by rolling PCMCI [3] swaps from the Fed-controlled term premium of the quantitative-easing regime to market-driven inflation expectations. The problem we address is therefore detecting the timestamps at which such regime shifts occur accurately using all available signal.

Regime shift detection in multivariate time series is a mature subfield, spanning Bai–Perron structural-break testing [4], PELT [5], Markov-switching VARs [6], and, in finance, regime-aware causal-discovery methods [7], [8] (see [9] for a survey). All of these methods rely entirely on the data signal, which is a structural limitation for treasury market data: the daily series are noisy and heavily multicollinear, since adjacent-maturity yields, the additive term-premium and risk-neutral components, and curve factors that are linear functions of the yields all co-move. The data-side fingerprint of a true regime shift is therefore easily masked by everyday co-movement that any \(L_2\)-cost or likelihood-based detector must absorb as noise.

For regime shifts driven by external policy decisions, such as central-bank monetary-policy changes, an informative source of evidence is available before, during, and after the shift: contemporaneous text. FOMC minutes,¹ central-bank speeches, and policy communications explicitly state new policy stances, often weeks before the new dynamics materialise in observed yields, yet none of the standard detectors uses this signal. Following a collective-intelligence approach [10], combining complementary information from multiple sources can provide a richer view of regime change than any single modality in isolation.

We propose a text-enhanced regime shift detection pipeline that combines a text channel (LLM-based proposals from FOMC minutes) with a data channel (Treasury/macro panel data), with bidirectional cross-modal validation: text-proposed candidates pass a data-side likelihood-ratio test, and data-side candidates from any regime shift detector pass an LLM lenient-prompt text check. The data channel is detector-agnostic: the ratification stage consumes a candidate set, not a particular algorithm’s internals, so any data-driven regime shift detection method can fill it; we evaluate four (PELT, binary segmentation, Bai-Perron, and rolling PCMCI). The LLM is restricted to a regime-level decision where text carries clear signal advantage.

1.0.0.1 Contributions.

(i) A detector-agnostic text-enhanced regime-detection pipeline: LLM-based text proposals on FOMC minutes are validated by a likelihood-ratio test on the Treasury/macro panel, while the candidate set of any data-driven regime shift detection method is ratified through a lenient cross-modal text check; we demonstrate the framework with four interchangeable detectors. (ii) An empirical evaluation on 2010–2024 Federal Reserve communications paired with U.S. Treasury data, with every method scored on one consistent 14-variable panel (no panel-mismatch confound). The text channel alone attains F1 = 0.82 against a 26-event verified anchor list with same-day modal detection latency; the cross-validation pipeline is statistically comparable for every data-driven detector (F1 up to 0.82) and exceeds the strongest pure data-only baseline (0.68). Text-side and data-side errors are complementary, which is what motivates the cross-validation design.

2 Related Work↩︎

2.0.0.1 Regime shift detection:

Detecting structural breaks in multivariate time series has a long history. The Bai–Perron framework [4] performs multiple structural-break testing under full-likelihood, with sequential and dynamic-programming estimators of break dates. PELT [5] formulates regime shift detection as cost-function optimisation with \(O(n)\) exact-solution complexity for separable costs, and is now a standard baseline. Markov-switching VARs [6] identify regime labels jointly with per-regime dynamics via EM. Truong et al. [9] survey the offline regime shift detection literature. Applied to multi-regime financial data, all these methods derive their signal from data alone and ignore contemporaneous policy text; in §4 we report PELT, BinSeg, Bai-Perron, and PCMCI as data-only baselines.

2.0.0.2 Regime-aware causal discovery:

CD-NOD [11] treats nonstationarity as a surrogate variable and Regime-PCMCI [8] jointly estimates discrete regime labels and per-regime causal structure via EM. Sadeghi et al. [7] adapt CD-NOD to a lagged time-series setting (CD-NOTS) and apply it to financial markets. Although the headline contribution of this line is per-regime structure recovery, the regime-detection sub-problem is identical to ours and we treat the regime-detection output of these methods as direct baselines. We focus on detection in this paper; downstream per-regime structure recovery is left to future work.

2.0.0.3 LLMs in causal and time-series reasoning:

Initial work used LLMs as priors over causal edges [12]–[14]. Subsequent work moves toward LLM-error-resistant interfaces: COAT [15] pairs LLM-driven factor proposal with statistical structure learning; Vashishtha et al. [16] show that LLM ordering queries are more reliable than pairwise edge-direction queries. Zečević et al. [17] document that LLMs frequently pattern-match on memorised content rather than reason causally.

2.0.0.4 LLMs in finance:

Domain-adapted financial LLMs such as BloombergGPT [18] and the open-source FinGPT [19] (surveyed by Li et al. [20]) show that large models capture finance-specific language. A parallel line uses general-purpose LLMs to extract predictive signal directly from financial text: Lopez-Lira and Tang [21] report that ChatGPT sentiment on news headlines forecasts subsequent returns, and recent IMF work applies LLMs at scale to central-bank communication [22]. These approaches treat the LLM as a forecaster or sentiment extractor; we instead restrict it to a regime-boundary judgement and cross-validate that output against a statistical structural-break test, so it is checked rather than trusted.

3 Enhancing Regime Detection From Text Data↩︎

A regime change in monetary policy can manifest in two complementary ways: (i) as a narrative shift in central-bank communication that announces or pre-commits to a new policy stance, and (ii) as a structural break in the joint dynamics of yields, spreads, and macro variables. Statistical detection on time-series panels alone is sensitive to (ii) but blind to (i) when the narrative leads the data, while a purely text-based detector misses regime changes that originate outside the central-bank corpus (e.g., between-meeting speeches, financial-stress events). Our pipeline reconciles both signals through a bidirectional cross-validation design.

Concretely, the pipeline consumes (i) a dated corpus of FOMC minutes \(\mathcal{D} = \{D_1, \ldots, D_M\}\) ordered by release date, and (ii) a multivariate financial time-series panel \(X \in \mathbb{R}^{T \times N}\) over the same period. It outputs a set of validated regime-boundary dates \(\mathcal{B}\). Figure 2 illustrates the four stages.

3.0.0.1 Stage A: LLM proposer.

The strict prompt (Algorithm 3; full text in Figure 4) anchors three labels to explicit definitions: a regime is a multi-meeting stance with stable reaction function, dominant tool, and trajectory; a major_pivot is a transition between regimes (taper start, first hike, zero-lower-bound exit, framework adoption, “transitory”\(\to\)“persistent” reframing). A calibration prior of 0–1 pivots per year (FOMC base rate) discourages over-detection, and the LLM is asked to default to incremental under ambiguity.

3.0.0.2 Stage B: Bootstrap likelihood-ratio validation.

Each LLM candidate \(\tau\) is validated on the full 14-variable panel \(X \in \mathbb{R}^{T \times N}\) (\(N{=}14\)) by a Chow-style structural-break test [23] on a lag-1 Gaussian VAR, \[X_t = c + \Phi X_{t-1} + \varepsilon_t, \qquad \varepsilon_t \sim \mathcal{N}(0, \Sigma). \label{eq:var1}\tag{1}\] The test contrasts \(H_0\): a single VAR(1) \((c,\Phi,\Sigma)\) governs the whole window \([\tau{-}W,\tau{+}W]\) (no regime shift at \(\tau\)), against \(H_1\): the pre- and post-\(\tau\) segments follow different VAR(1) parameters (a structural break at \(\tau\)). Algorithm 5 fits 1 on the pre-, post-, and full \(W\)-day windows around \(\tau\) and forms the split likelihood-ratio statistic \[\mathrm{LR}(\tau) = -2\bigl[\ell_\text{full} - (\ell_\text{pre}+\ell_\text{post})\bigr], \label{eq:lr}\tag{2}\] where \(\ell_\text{pre},\ell_\text{post},\ell_\text{full}\) are the fitted Gaussian log-likelihoods.

The asymptotic \(\chi^2\) calibration of this statistic is invalid here. The unrestricted model releases \(\mathrm{df} = N + N^2 + N(N{+}1)/2 = 315\) intercept, autoregression, and residual-covariance parameters across the boundary, yet each 90-day half-window holds only \(\approx\!63\) daily observations; the per-segment VAR is therefore over-parameterised, overfits segment noise, and inflates \(\mathrm{LR}\). We instead calibrate the null by a residual bootstrap, which additionally relaxes the Gaussian-innovation assumption of 1 : fit the pooled single-regime VAR(1) on \(X_{[\tau-W,\tau+W]}\), resample its residuals i.i.d.to generate \(B{=}500\) datasets under \(H_0\), recompute the split \(\mathrm{LR}\) on each, and set \[p = \frac{1 + \sum_{b=1}^{B} \mathbf{1}\!\left[\,\mathrm{LR}^*_b \geq \mathrm{LR}(\tau)\,\right]}{B+1}, \label{eq:bootp}\tag{3}\] where \(\mathbf{1}[\cdot]\) is the indicator function, so \(p\) is the (add-one smoothed) fraction of bootstrap replicates at least as extreme as the observed statistic. A candidate is retained iff \(p \leq \alpha\); Algorithm 5 summarises the whole procedure.

3.0.0.3 Stage C: Lenient cross-validation of detector candidates.

On the other hand, the data channel runs a data-driven regime shift detector on the panel \(X\), producing a candidate set \(\mathcal{C}_\text{D}\). Stage C consumes only this set, never the detector’s internals, so any data-driven regime shift detection method can fill the slot, e.g., PELT, binary segmentation, Bai-Perron, and rolling-window PCMCI. Each candidate \(\tau \in \mathcal{C}_\text{D}\) is then ratified under a distinct, more lenient prompt (Figure 6) that sees only the two FOMC documents straddling \(\tau\), not the panel data. The LLM is asked whether either document contains substantive monetary-policy content (a rate decision, balance-sheet change, forward-guidance shift, crisis acknowledgement, or framework or diagnosis update) that could plausibly accompany a structural break around \(\tau\). The prompt thus tests text-content sufficiency, not literal explanation of the data signal: a positive verdict means the surrounding minutes carry enough policy substance to make the break credible as a regime boundary, not that the LLM has matched specific data movements to specific sentences.

3.0.0.4 Stage D: Union and deduplication.

Stage D greedily clusters validated candidates within \(\Delta = 14\) days, retains the earliest in each cluster, and tags each output by source set (LLM-only, detector-only, or both). The final boundary set is \(\mathcal{B} = \mathrm{dedup}_{\Delta}(\mathcal{C}_\text{LLM-val} \cup \mathcal{C}_\text{D-cv})\).

4 Experiments on Real Data↩︎

4.0.0.1 Setup.

We evaluate our pipeline on Federal Reserve communications and US Treasury data over 2010–2024.² The FOMC corpus contains 120 minutes documents (8 per year over 15 years) sourced from federalreserve.gov. The financial panel is a lean 14-variable bond-market panel (3,752 daily observations after alignment) covering eight categories: Treasury yields and one curve-shape factor; the 10-year ACM term-structure decomposition into expected-rates and term-premium components; the policy rate; three complementary inflation perspectives (headline CPI, core PCE, and 10-year breakeven); the unemployment rate; banking-funding stress (TED spread); cross-asset and bond-market implied volatility (VIX and the ICE BofAML MOVE Index); and Fed total assets, the direct quantitative-easing / quantitative-tightening framework observable. Redundant intermediate-maturity yields, derived curve-slope and curvature factors, and the FOMC target-rate duplicate are removed in preprocessing to reduce multicollinearity.

Stage B’s bootstrap likelihood-ratio test operates on the full 14-variable panel, rolling z-scored over a 252-day window. Running Stage B on the same panel as the multivariate baselines removes any panel-mismatch confound, and the residual bootstrap (rather than an asymptotic \(\chi^2\)) keeps the test valid despite the resulting over-parameterised per-window VAR. The data channel accepts any data-driven regime shift detection method; we evaluate PELT, binary segmentation, Bai-Perron, and rolling PCMCI. All four detectors run on the same 14-variable z-scored panel.

The LLM is Anthropic’s Claude Sonnet 4.6 at temperature 0.2; the likelihood-ratio-test window is \(W=90\) days; rolling-PCMCI uses a 200-day window with 20-day step.

4.0.0.2 Anchor list.

Ground-truth regime boundaries are taken from a 26-event anchor list. The list was first proposed by prompting OpenAI GPT-5.5 for canonical US monetary-policy regime boundaries over 2010–2024, and then verified by the authors event-by-event against federalreserve.gov press releases and Federal Reserve Board policy timelines; events without a primary-source confirmation were dropped, and the surviving 26 dates are those the authors agreed on. The list comprises canonical Fed actions (quantitative-easing / quantitative-tightening announcements, rate-cycle inflections, crisis interventions) together with well-documented forward-guidance and framework events (Operation Twist, the Evans rule, Flexible Average Inflation Targeting framework adoption, calendar-based guidance). Note that the proposing LLM (GPT-5.5) is distinct from the pipeline’s proposer/ratifier LLM (Claude Sonnet 4.6), so the evaluation does not score a model against its own outputs.

4.0.0.3 Baselines.

We compare three groups, all on the same 14-variable z-scored panel (no panel-mismatch confound). (i) Four data-only baselines, each detector run alone: PELT [5] (\(L_2\)-cost, BIC-scaled penalty \(10\,d\log n\)); BinSeg (greedy binary segmentation, ruptures default, \(L_2\) cost, BIC penalty); Bai-Perron [4] (dynamic-programming placement of \(K{=}25\) structural breaks, grid stride \(\mathrm{jump}{=}5\)); and PCMCI, the prevailing partial-correlation approach used in finance [7] (rolling windows, Jaccard distance \(>0.80\)). (ii) LLM only: Stage A proposals validated by the Stage B residual-bootstrap likelihood-ratio test, isolating text-driven detection. (iii) Cross Validation, our full pipeline, whose data channel accepts any of the four detectors.

The standalone PCMCI row is exactly the rolling-PCMCI detector without the cross-modal stages, so the gap to Cross Validation \(+\) PCMCI (0.46 \(\to\) 0.82) isolates the contribution of the LLM channel.

6pt

Table 1 reports the headline results. We highlight four findings.

4.1 Same-day Detection Latency for the Text Channel↩︎

The text channel caught 20 anchors. Most were dated to the exact day of the FOMC release, giving a mean signed offset of only \(+3.1\) days. The few larger offsets arise when a pivot is stated more clearly at the following meeting than at the one that triggered it: the 2018-12 Powell pivot, for instance, was picked up at the 2019-01 FOMC (\(+42\) days). The data-only detectors time the same breaks far less precisely, with mean offsets from \(-10\) to \(+15\) days (Table 1, Off column), because a structural break surfaces in prices only gradually. This near-zero modal latency is mechanical. The text channel reads the same minutes that constitute the regime change, so it flags the shift as soon as they become public.

4.2 Cross-modal Complementarity↩︎

We illustrate with PCMCI, the same text/data complementarity holds for the other three detectors. On the consistent panel the text channel catches 20 of the 26 anchors and standalone PCMCI catches 9, with 7 found by both. The two anchors PCMCI recovers that the text channel misses (the 2021-Q1 Summary-of-Economic-Projections hike-dot shift and the 2023-03 Silicon-Valley-Bank-era hike) move market microstructure between FOMC meetings, where a text-anchored detector has nothing to read; conversely, standalone PCMCI misses thirteen of the anchors the text channel finds. Because the two channels fail on largely disjoint events, cross-validating them recovers what either alone would miss: the pipeline ratifies 8 of PCMCI’s 13 candidates, unions them with the bootstrap-validated text candidates, and reaches F1 = 0.82 at 23/26 recall (Table 2), statistically tied with the text channel (0.82) and far above standalone PCMCI (0.46). The lenient ratification prompt is load-bearing here.

2.5pt

4.3 Data-only Baselines Fall Short↩︎

On the same z-scored 14-variable panel, the four data-only detectors reach F1 between 0.46 (PCMCI) and 0.68 (BinSeg): they locate some structural breaks but at low precision, because adjacent-maturity yields and the additive term-premium / risk-neutral components co-move so strongly that a genuine policy break is hard to separate from everyday churn. The text channel alone clears the strongest of them (0.82 vs., \(+0.14\)), and every cross-validation instantiation also exceeds it (0.69–0.82), by adding contemporaneous text candidates at a naturally calibrated frequency (\(\approx 1.5\) candidates/year) that \(L_2\)-cost detectors structurally cannot produce.

4.4 Interpretable False Positives↩︎

Cross-Validation \(+\) PCMCI produces 30 detections at the reported setting, of which 23 match an anchor (precision 0.77) and 7 do not. Classifying each non-matching detection post-hoc against the same primary sources used to build the anchor list (Table 3), all seven are interpretable early signals or intra-regime escalations rather than spurious noise: e.g.the 2012-08-14 pre-QE3 run-up, the 2019-06-19 dovish FOMC that pre-signalled the July cut, the 2021-09-22 pre-taper signalling that bridges to the 2021-11 anchor, and the 2024-08-14 carry-trade unwind that preceded the September cut. The reported precision is therefore a conservative lower bound: several “false” alarms are early detections of regime changes that the anchor list dates to a later announcement. This pattern is detector-independent. The three text-side false positives (2019-06-19, 2021-09-22, 2022-01-26) are identical across all four instantiations because the bootstrap-validated text candidates are shared; only the data-side false positives differ. PCMCI contributes just four, all interpretable, whereas the \(L_2\)-cost detectors each add 10–15 that cluster on recurring volatility and turn-of-year dates with no FOMC correlate (e.g.-12-31, 2016-11-07, 2020-02-28, 2022-10-31): the everyday co-movement that depresses their precision in Table 1, not interpretable signal.

3pt

5 Ablation Studies↩︎

5.0.0.1 Detector-agnosticism and the Stage-C threshold \(\theta_C\).

Stage C consumes only a candidate set, so the data-side detector is a free choice among data-driven regime shift detection methods. Holding Stage A and Stage B fixed, we run the pipeline with four interchangeable detectors on the same 14-variable panel and sweep the single global Stage-C confidence threshold \(\theta_C\) (Table 4). Two patterns are stable across the sweep. (i) Every detector instantiation is comparable to the text-only LLM (F1 \(=\) 0.82) rather than clearly above or below it: on a rigorously consistent panel the cross-modal effect is modality robustness, not a headline lift. (ii) The detector ranking is stable, with rolling PCMCI strongest — its Jaccard-on-causal-structure criterion is the sparsest of the four (\(\approx 1\) candidate/year vs.\(\approx 1.6\) for the \(L_2\)-cost detectors), so its union with the text channel adds the fewest false alarms. At the operating point reported in Table 1 (\(\theta_C{=}0.8\)) all four instantiations exceed the strongest pure data-only baseline (BinSeg, 0.68) and Cross-Validation\(+\)PCMCI matches LLM only. We stress that \(\theta_C\) is a single global hyperparameter applied identically to every detector, not tuned per detector; the full grid is reported only for transparency.

6pt

5.0.0.2 Stage C lenient vs strict cross-validation prompt.

The lenient prompt in Stage C asks whether the FOMC pair straddling a detector candidate contains any substantive monetary-policy content. A strict alternative would only ratify candidates that the LLM also classifies as major_pivot under Stage A’s prompt. Substituting the strict prompt for the lenient one in our otherwise-identical pipeline sacrifices the two anchors that the PCMCI instantiation uniquely recovers — the 2021-Q1 Summary-of-Economic-Projections hike-dot shift and the 2023-03 Silicon-Valley-Bank-era hike — because the LLM under the strict prompt declines to call them major pivots: they fall between FOMC meetings or are couched in policy-continuation language. Losing them pulls Cross-Validation \(+\) PCMCI back toward the text-only operating point, erasing the data channel’s contribution. The lenient prompt is therefore a load-bearing design choice: it widens what the LLM will ratify just enough to preserve the modality complementarity that motivates cross-validation in the first place.

5.0.0.3 Other hyperparameters.

\(\theta_C\), swept above, is the only tunable knob in the pipeline; the remaining settings are fixed by design rather than chosen on the data. The Stage-A confidence cutoff \(\theta_A{=}0.6\) is the FOMC-base-rate prior of Algorithm 3. The Stage-B residual bootstrap clears every text candidate by a wide margin on the 14-variable panel, so the likelihood-ratio window is not a load-bearing choice; \(W{=}90\) d and the \(\pm 90\) d matching tolerance are held fixed throughout.

6 Discussion and Limitations↩︎

6.0.0.1 Discussion:

Evaluated on a single consistent panel, the contribution is not a headline accuracy lift but a robustness result: the text channel alone attains F1 = 0.82, and wrapping any of four interchangeable data-driven detectors in the cross-validation pipeline matches that level while every pure data-only baseline tops out at 0.68. Stage B gives LLM proposals an explicit statistical grounding (a residual-bootstrap structural-break test that stays valid under an over-parameterised VAR), and Stage C lets an arbitrary detector contribute without exposing the rest of the pipeline to its internals. The design principle, restricting LLM authority to a regime-level judgement where contemporaneous text carries genuine signal and delegating structural testing to statistics, transfers to any domain with a contemporaneous text channel and a structural-break-testable data channel.

6.0.0.2 Limitations:

On this panel cross-validation reaches parity with, not improvement over, the text-only channel: the LLM already recovers most policy-announced anchors, so the data channel’s marginal contribution is a handful of intra-meeting microstructure events at the cost of added false alarms. Regime shifts that are not announced in text, being purely market-driven, are unreachable by the text channel by construction. The 26-event anchor list is a curated ground truth, and precision is sensitive to its completeness: several nominal false positives are interpretable early signals or genuine policy actions absent from the list. The Stage-B VAR(1) is over-parameterised at \(N{=}14\) (the bootstrap repairs calibration but not power) and fires on joint mean and volatility breaks. Finally, we evaluate only the US Treasury market, chosen because its data is the most complete and liquid and its policy record (FOMC minutes) the best documented; future work will extend the pipeline to the UK, euro-area, and Asian sovereign bond markets.

7 Conclusion↩︎

We presented a text-enhanced regime-shift detection pipeline for bond markets that allocates LLM and statistical components by where each carries genuine signal. Evaluated with every method on one consistent 14-variable panel (no panel-mismatch confound), the text channel alone reaches F1 = 0.82 on the 2010–2024 Federal Reserve corpus with same-day modal detection latency; the cross-validation pipeline matches it (F1 up to 0.82) and lifts the weakest data-only detector (standalone PCMCI, 0.46) to that level, while every pure data-only baseline tops out at 0.68. The underlying design principle is to restrict LLM authority to the one question where contemporaneous text genuinely carries a signal advantage, and to delegate edge-level identification to statistical methods. Because the cross-modal ratification consumes only a candidate set, the data channel accepts any data-driven regime shift detection method: PELT, binary segmentation, Bai-Perron, and rolling PCMCI all clear the strongest pure data-only baseline once wrapped in the pipeline (§5). The principle generalises beyond finance to any domain with a contemporaneous text channel and a structural-break-testable data channel. Epidemiological surveillance, climate regime shifts, and macro-policy transmission are immediate candidates.

Acknowledgments↩︎

This work is funded by UKRI EPSRC grant No. EP/Y028392/1: AI for Collective Intelligence (AI4CI).

References↩︎