📈 economics

Hidden Markov Modelling of Earthquake-Related Market Regimes and State-Dependent Forecasts: Evidence from the Turkish Stock Market

This study utilizes a three-state Gaussian Hidden Markov Model on Borsa Istanbul 100 data to demonstrate that while major earthquakes do not systematically trigger shifts between market regimes, the resulting state-dependent forecasts reveal significant differences in short-horizon downside risk.

Original authors: Gamze Özel

Published 2026-11-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Gamze Özel

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Financial markets are rarely the steady, predictable machines they appear to be on a daily chart. Instead, they often shift between distinct moods, or "regimes," that change how prices behave. Sometimes the market is calm and moves in a tight, predictable range; other times it is gripped by uncertainty, swinging with moderate volatility; and occasionally, it enters a crisis state where prices tumble and chaos reigns. These shifts are not always obvious to the naked eye, but they fundamentally alter the risk an investor faces. Understanding when a market is in one of these states, and how long it might stay there, is crucial for anyone trying to navigate the economy. A new study from Turkey explores whether a specific, terrifying kind of external shock—major earthquakes—forces the stock market into these crisis states, or if the market's internal rhythms are more complex than a single disaster can dictate.

The researchers focused on the Borsa Istanbul 100, the main index of the Turkish stock market, which serves as a barometer for the nation's economic health. Turkey sits on some of the world's most active fault lines, making it a perfect laboratory for this question. The team gathered over four thousand days of trading data, spanning from early 2009 to early 2026, and looked for patterns in how the market moved. They did not simply look at whether the market went up or down on the day of an earthquake. Instead, they used a statistical tool called a hidden Markov model. Think of this tool as a way to listen to the market's underlying heartbeat, trying to detect whether the market is currently in a "normal," "uncertain," or "crisis" mode, even when those modes are not explicitly labeled on a news ticker. By analyzing the daily returns, the model identified three distinct states: a normal state with low volatility, an uncertainty state with moderate swings, and a crisis state characterized by high volatility and negative returns.

Once the model mapped out these three states over the sixteen-year period, the researchers overlaid the history of major earthquakes. They filtered the data to include only significant seismic events, specifically those with a magnitude of 5.5 or higher, and aligned them with the trading days when the market was open. The goal was to see if these earthquake days were disproportionately likely to fall into the "crisis" category. The results were surprising. Despite the intuitive expectation that a major earthquake would immediately plunge the market into a state of panic, the data showed no such systematic link. The earthquake days were distributed across the three states in almost the exact same proportion as any random day in the market's history. Roughly 9 percent of earthquake days fell into the crisis regime, while the rest were split between uncertainty and normal conditions. This distribution was statistically indistinguishable from the market's overall behavior.

The study also examined whether the size of the earthquake mattered. They separated the events into those with a magnitude between 5.5 and 6.49 and those that were even larger, at 6.5 or above. The hypothesis was that the most massive quakes might be the ones capable of triggering an immediate shift into a crisis state. However, the analysis found no evidence for this either. Whether the earthquake was a strong 5.5 or a devastating 6.5, it did not predict an immediate change in the market's underlying regime. The market did not consistently jump from a calm state to a crisis state simply because a large tremor occurred. This suggests that while a specific earthquake might cause significant damage to certain companies or sectors, the aggregate market index does not automatically flip into a "crisis" mode based on the earthquake's magnitude alone. The market's reaction appears to depend on other factors, such as the location of the quake, the timing relative to trading hours, or the broader economic conditions at the time, none of which were captured by the earthquake's size alone.

While the earthquakes did not seem to force the market into a crisis state, the study found that knowing the current state of the market is incredibly powerful for predicting what happens next. The researchers used their model to simulate how the market might behave over the next month, depending on where it started. If the market begins in a crisis state, the simulations show a high probability that it will remain in that state for the next day or two, but it is unlikely to snap back to normal immediately. Instead, it tends to drift into the "uncertainty" state first, acting as a middle ground before eventually returning to calm. Conversely, if the market starts in a normal state, it is very likely to stay there for a while, though the risk of eventually visiting a crisis state does creep up over a thirty-day period. These simulations revealed that the starting point matters immensely. A market beginning in crisis has a much wider range of possible outcomes and a significantly higher chance of ending the month with a loss compared to a market starting in a normal state.

The findings offer a nuanced view of how natural disasters interact with financial systems. They suggest that the market is not a fragile glass that shatters instantly at the first sign of a tremor. Instead, it has a complex internal structure that resists simple cause-and-effect explanations. A major earthquake does not guarantee a market crash, nor does it guarantee a shift into a crisis regime. The market's reaction is filtered through its current condition and other unseen variables. However, once the market is in a crisis state, it tends to stay there for a while, moving through a period of uncertainty before recovering. For investors and analysts, the lesson is that watching for external shocks like earthquakes is only part of the picture. Understanding the market's current hidden state—whether it is calm, uncertain, or in crisis—provides a much clearer map of the risks that lie ahead, regardless of what the ground beneath the trading floor is doing.

Technical Summary: Hidden Markov Modelling of Earthquake-Related Market Regimes and State-Dependent Forecasts

Problem Statement
The paper addresses the challenge of understanding how major earthquakes influence financial market dynamics, specifically the Borsa Istanbul 100 (BIST 100) index. While existing literature extensively documents short-run abnormal returns and volatility changes following specific disasters using event-study methods, these approaches often assume a single underlying return-generating process. The author argues that financial returns exhibit volatility clustering, fat tails, and distinct periods of high or low risk that are better represented by models allowing parameters to shift according to unobserved states. The study seeks to determine whether major earthquakes are systematically associated with changes in latent return and volatility regimes and whether earthquake characteristics can predict immediate shifts in market states.

Methodology
The study employs a three-state Gaussian Hidden Markov Model (HMM) to analyze daily logarithmic returns of the BIST 100 index.

Data: The empirical sample comprises 4,334 daily observations from January 5, 2009, to April 3, 2026. The earthquake catalogue covers the overlapping period from June 19, 2009, to October 27, 2025.
Event Filtering: From an initial 96 earthquakes with magnitude $M \ge 5.5$ , 74 representative events were selected using a 48-hour main-event filter to avoid over-weighting aftershock clusters. Events were aligned with the contemporaneous or next trading day.
Model Specification: The HMM assumes the daily return $r_t$ follows a state-dependent Gaussian distribution $N(\mu_k, \sigma^2_k)$ where $k \in \{1, 2, 3\}$ represents the unobserved market regime. The regimes are inferred post-estimation based on volatility and mean return characteristics.
Analysis Strategy:
1. Regime Estimation: The model estimates transition probabilities, expected regime durations, and state-specific means and variances.
2. Association Tests: A Pearson chi-square goodness-of-fit test compares the distribution of earthquake-linked trading days across the three regimes against the unconditional regime composition of the full sample.
3. Magnitude Analysis: Events are stratified by magnitude ( $5.5 \le M < 6.5$ vs. $M \ge 6.5$ ) to test if larger quakes predict immediate regime changes (using Fisher's exact test).
4. Forecasting: State-conditional forecasts are generated using the transition matrix and Monte Carlo simulations (100,000 paths) to project 30-day cumulative returns and regime probabilities, conditional on the starting state.
Proposed Extension: The paper outlines a theoretical framework for a time-varying transition probability (TVTP) model where earthquake covariates (magnitude, depth, distance) could influence the probability of moving between states, though this extension is not estimated in the current analysis.

Key Results

Regime Characterization: The three-state model identifies distinct regimes:
- Crisis: 7.27% of observations, mean return -0.79%, volatility 3.45%, expected duration 5.3 days.
- Uncertainty: 51.66% of observations, mean return 0.20%, volatility 1.16%, expected duration 12.1 days.
- Normal: 41.07% of observations, mean return 0.19%, volatility 0.50%, expected duration 15.7 days.
- The "Uncertainty" and "Normal" regimes have similar positive mean returns; their primary distinction is volatility.
Transition Dynamics: The crisis regime is persistent (81.1% probability of remaining), but exits from crisis occur predominantly through the "Uncertainty" regime rather than directly to "Normal." The "Normal" regime is the most persistent (93.6% self-transition).
Earthquake Association:
- Hypothesis H3 (Rejection): The distribution of earthquake-linked trading days across the three regimes (7 crisis, 35 uncertainty, 32 normal) does not differ significantly from the unconditional state composition ( $\chi^2 = 0.845, p = 0.655$ ).
- Hypothesis H4 (Rejection): There is no significant association between earthquake magnitude groups and immediate next-day regime changes ( $p = 0.952$ ).
State-Dependent Forecasts: While the pooled earthquake association is weak, state-conditional forecasts reveal substantial differences in risk.
- Starting from a Crisis state, the 30-day expected return is near zero (0.03%) with a 46.0% probability of a negative cumulative return and an 89.9% probability of visiting the crisis regime again.
- Starting from a Normal state, the expected return is 4.28% with only a 20.5% probability of a negative outcome.
- Forecasts show that market stress dissipates gradually through the "Uncertainty" regime rather than snapping back immediately to "Normal."

Significance and Claims
The paper claims to contribute to the literature in four specific ways:

Sample Scope: It utilizes a long-term sample (2009–2026) containing multiple seismic events rather than focusing on a single catastrophic event.
Methodological Separation: It separates the estimation of latent market regimes from the subsequent examination of earthquake days, reducing the risk of circular classification where states are defined solely by disaster dates.
Expanded Metrics: It moves beyond average abnormal returns to report transition probabilities, regime persistence, expected durations, and regime-conditional forecasts.
Theoretical Extension: It proposes a time-varying Markov-switching specification to examine whether earthquake characteristics affect returns within a regime or alter the probability of regime transitions.

Modest Interpretation of Findings
The author explicitly cautions that the pooled results do not establish a systematic causal link between earthquake magnitude and immediate regime shifts. They argue that magnitude alone is an insufficient proxy for financial exposure, as factors like depth, location, economic exposure, and policy responses (e.g., trading suspensions) vary significantly. The study concludes that while earthquakes do not systematically predict regime changes in the aggregate index across this sample, the latent-state framework remains valuable for risk monitoring. State-conditional forecasts demonstrate that the current market regime materially affects short-horizon downside risk distributions, providing a probabilistic tool for assessing market stress even in the absence of a direct earthquake-regime correlation.

More like this