# Regime-Switching Models

> Detect regime shifts at the narrative level. NOSIBLE WORLD captures the story turn in 95 languages, days or weeks before variance confirms the break.

**URL:** https://nosible.com/regime-switching-models

This is page 11 of 13 in the NOSIBLE instrument suite. Live, global. For quants modelling regime and volatility states.

## Detect regime shifts at the narrative level

Volatility-based regime detection lags. By the time variance spikes, the move has happened.

The narrative shifts days or weeks earlier. NOSIBLE WORLD captures that shift in 95 languages, at event granularity, with tone, severity, momentum, and geography per event.

CTAs:
- **Start Trial** → [trial hub](https://nosible.com/start-trial)
- **See the research** → [#evidence](https://nosible.com/regime-switching-models#evidence)

## 01 — The gap (Variance lag, narrative lead)

**Markets switch regimes because the story shifted.**

Variance confirms a regime change. It does not lead one. Sanctions announcements, central-bank pivot signals, and supply-chain shocks hit the open web before they hit implied or realised variance. A price-only filter is lagging by construction.

## 02 — Evidence: Recent work on narrative regime detection (2024–2026)

Peer-reviewed and arXiv papers from the last two years, covering narrative econometrics, topic-model HMMs, news-changepoint detection, and LLM-augmented regime identification. Each method takes an event ledger as input.

These picks deliberately avoid the standard foundations (Hamilton 1989, Kim 1994, Tong 1990, Heston 1993). Every paper cited extends those foundations with text, narrative, or LLM machinery.

**Reading order:** Papers 01 to 03 treat narrative as a leading variable. 04 to 06 wire it into the regime filter. 07 and 08 are LLM-augmented detection from 2025 and 2026.

| # | Title | Authors | Year | Journal | Link |
|---|-------|---------|------|---------|------|
| 1 | Narrative Shift Detection: A Hybrid Approach of Dynamic Topic Models and Large Language Models | Lange, Schmidt, Reccius, Mueller, Roos & Jentsch | 2025 | arXiv | [arxiv.org/abs/2506.20269](https://arxiv.org/abs/2506.20269) |
| 2 | History Rhymes: Macro-Contextual Retrieval for Robust Financial Forecasting | Khanna, Berger, Chopra, Berghaus & Sifa | 2025 | arXiv | [arxiv.org/abs/2511.09754](https://arxiv.org/abs/2511.09754) |
| 3 | Narrative Momentum | Lee, Lou, Ozik & Sadka | 2024 | SSRN | [papers.ssrn.com](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4912496) |
| 4 | Can LLMs Learn Macroeconomic Narratives from Social Media? | Gueta, Feder, Gekhman, Goldstein & Reichart | 2024 | NAACL | [arxiv.org/abs/2406.12109](https://arxiv.org/abs/2406.12109) |
| 5 | From Headlines to Forecasts: Narrative Econometrics in Equity Markets | Hayrapetyan & Gevorgyan | 2025 | JRFM | [mdpi.com/1911-8074/18/9/524](https://www.mdpi.com/1911-8074/18/9/524) |
| 6 | Transformers Beyond Order: A Chaos-Markov-Gaussian Framework for Short-Term Sentiment Forecasting | Pathan | 2025 | arXiv | [arxiv.org/abs/2506.17244](https://arxiv.org/abs/2506.17244) |
| 7 | Cross-Platform Narrative Prediction: Leveraging Platform-Invariant Discourse Networks | Gerard, Luceri, Blas & Ferrara | 2025 | arXiv | [arxiv.org/abs/2510.09464](https://arxiv.org/abs/2510.09464) |
| 8 | Narratives from GPT-derived Networks of News, and a Link to Financial Markets Dislocations | Miori & Petrov | 2024 | PLOS | [arxiv.org/abs/2311.14419](https://arxiv.org/abs/2311.14419) |

### Why each paper matters

1. **Narrative Shift Detection (Lange et al., 2025, arXiv)** — Dynamic topic model plus LLM finds narrative shifts in news, separating real regime turns from content drift.
2. **History Rhymes (Khanna et al., 2025, arXiv)** — Retrieval-augmented forecasting jointly embeds macro indicators and news sentiment, conditioning on analogous regimes when static baselines fail.
3. **Narrative Momentum (Lee, Lou, Ozik & Sadka, 2024, SSRN)** — Roughly 350 narrative-intensity series from real-time news. Narrative-mimicking long-short portfolios beat declining ones by eight percent annually.
4. **Can LLMs Learn Macroeconomic Narratives from Social Media? (Gueta et al., 2024, NAACL)** — Builds narrative datasets from X, injects LLM-extracted representations into macro prediction. Feed matters more than model in their ablations.
5. **From Headlines to Forecasts (Hayrapetyan & Gevorgyan, 2025, JRFM)** — BERTopic narrative-decay factors on Microsoft news beat the EPU index for monthly returns.
6. **Transformers Beyond Order (Pathan, 2025, arXiv)** — Chaos-Markov-Gaussian block inside a transformer. The Markov chain models regime shifts in sentiment, distinct from price-driven variance.
7. **Cross-Platform Narrative Prediction (Gerard et al., 2025, arXiv)** — Online narratives appear on one platform days before others. Discourse graphs predict cross-platform diffusion above ninety percent AUC.
8. **Narratives from GPT-derived Networks of News (Miori & Petrov, 2024, PLOS)** — GPT extracts entities from Wall Street Journal news. Weekly co-occurrence graph entropy maps to cross-asset dislocation dates.

**Field note:** "A price-only filter fires after the regime switches."

## 03 — The lag, traced: Narrative versus variance

**A real regime, side by side. Narrative versus variance.**

The August 2024 yen carry unwind, traced day by day across eleven sessions. The narrative regime turns on Jul 29. The volatility regime confirms on Aug 5. A price-only filter treats the window between as a single regime.

- **Episode:** USD/JPY · Tokyo · Aug 2024
- **Lead:** 5 calendar days

The trace runs over an 11-day window: Jul 25, Jul 26, Jul 29, Jul 30, Jul 31, Aug 1, Aug 2, Aug 5, Aug 6, Aug 7, Aug 8.

**Narrative lane** (alert state begins Jul 29):
- Jul 29 — BoJ tone hardens
- Jul 31 — Rate hike confirmed
- Aug 1 — Carry-unwind chatter
- Aug 2 — USD/JPY breach

**Vol-regime lane** (alert state begins Aug 5):
- Aug 5 — VIX prints 65

The narrative regime is in the alert state for five sessions before realised variance moves enough for a vol-regime filter to fire.

*Trace · NOSIBLE WORLD · open-web events · Tokyo session*

## 04 — NOSIBLE WORLD: the narrative, encoded at event granularity

**NOSIBLE WORLD · narrative-tagged events**

We index the open web in 95 languages, find every event, date it to the source-publication minute, and stamp each row with tone, severity, momentum, and geography. Thirty years of point-in-time history in a single database.

| Stat | Value |
|------|-------|
| Events | 100M+ |
| Sources | 300K+ |
| Languages | 95 |
| Point-in-time | 30 years |

### Narrative enrichments (covariates that drop straight into the regime filter)

| Enrichment | Detail |
|------------|--------|
| Tone | Sentiment per event, per source |
| Severity | Impact-weighted, normalized to 1 |
| Momentum | Mention-rate first derivative |
| Geography | Country, region, supply lane |

### Named cases (real, dated regime shifts caught in the open web)

| Risk | Place | Date | Event |
|------|-------|------|-------|
| Liquidity | New York | 2008·09 | CDS-spread chatter on Lehman built for months before the bankruptcy filing and the VIX spike. |
| Carry | Tokyo | 2024·08 | BoJ-hike narrative built for days before USD/JPY reversed sharply and the VIX printed 65. |
| Inflation | Global | 2021·Q3 | Supply-chain language shifted from transitory to persistent months ahead of CPI prints. |

## 05 — What you build: Wire the narrative into the regime filter

Builds that wire into a Markov-switching filter, a score-driven update, or a volatility model.

- **§01 — Narrative-conditioned transition matrices:** Drive regime-to-regime transition probability with dated narrative counts by topic. Switches fire earlier and each row identifies the proximate event.
- **§02 — Tone-momentum stress index:** Aggregate severity, momentum, and geographic spread of breaking events into a continuous stress index. An exogenous covariate for a sector GARCH.
- **§03 — Sentiment-dispersion regime filter:** Measure dispersion of tone across sources on the same event. Regime shifts when dispersion compresses or widens sharply, ahead of any move in implied or realised variance.

## 06 — Get started

**Talk to us about wiring NOSIBLE WORLD into your regime filter.**

CTA: **Start Trial** → [trial hub](https://nosible.com/start-trial)

## Related
- [Home](https://nosible.com)
- [NOSIBLE Search API](https://nosible.com/search-api)
- [Research](https://nosible.com/blog)
- [Start Trial](https://nosible.com/start-trial)