BIST 100 earthquake days show no excess crisis exposure

Özel's earthquake results give traders no reason to change BIST 100 exposure on seismic news. The paper finds no evidence that earthquakes push the index into its stress state, and states that conclusion plainly. Its surviving contribution is a volatility-state model with in-sample risk scenarios. Those scenarios may deserve validation. They are far from a sizing rule.

Separate the regimes from the earthquakes

A large quake could move the market into a recurring high-volatility condition, beyond its effect on the day's return. Regime membership might capture that shift even when a trading halt or policy response mutes the event-day move. The premise is reasonable. To avoid circularity, Özel fits regimes using returns alone, then checks which regimes contain the earthquake days.

The sample contains 4,334 daily BIST 100 closes from 5 January 2009 to 3 April 2026, yielding 4,332 decoded log returns. Özel estimates a three-state Gaussian hidden Markov model by EM with multiple starts. Each day receives the state with the highest smoothed posterior, computed using the full sample. After estimation, volatility determines the labels: crisis, uncertainty, normal.

The earthquake sample spans 19 June 2009 to 27 October 2025. A 48-hour main-event filter reduces Ninety-six M≥5.5 events to 74, retaining one event per aftershock cluster. Events are assigned to the same trading day unless they occur after the close or on a non-trading day. In those cases, they move to the next trading day.

Crisis accounts for 315 days (7.27%). Its mean daily log return is -0.79%, its volatility is 3.45%, and its expected duration is 5.29 days. The paper's summary makes two claims. Pooled tests find no systematic earthquake-regime link, while the fitted model generates predictive risk distributions that differ by state. The first holds up. The second follows from the model's construction; whether it survives outside the sample remains open.

Can the event tests detect much?

Among the 74 event days, 7 are classified as crisis, 35 as uncertainty and 32 as normal. Their respective shares are 9.46%, 47.30% and 43.24%, compared with unconditional shares of 7.27%, 51.66% and 41.07%. The chi-square statistic is 0.845, p = 0.655. A decoded state change follows only 5 of the 74 event days. Fisher's test relating magnitude group to that change returns p = 0.952.

This is a weak null. At the base rate, the expected crisis count is about 5.4 events (74 × 7.27%), barely above the paper's own threshold of five. The M≥6.5 group contains 8 events, with one in crisis. Five state changes across the sample leave Fisher's test very little to work with. We also did not find the catalogue's source or geographic bounds in the paper. Including offshore or distant M5.5 events would dilute any signal from destructive quakes. Özel acknowledges that individual disasters, including February 2023, can still cause large losses. The pooled test cannot detect those episodes.

Two calm states, a shared 0.2% daily drift

Uncertainty averages 0.20% a day with 1.16% volatility; normal averages 0.19% with 0.50% volatility. The paper acknowledges that the nearly equal means make this a risk split. Traders should interpret it accordingly. Outside crisis, the daily means differ by 0.01 points, so the model is classifying volatility. Both means are also in nominal lira. We found no inflation or currency adjustment, making a Turkish price-level effect a likely contributor to the positive drift in both calm states.

The most interesting number appears in the transition matrix: crisis-to-normal probability is 0.0000. Crisis exits pass through uncertainty, with a daily probability of 0.1890. Normal has the greatest persistence, at 0.9364.

What a month can lose

The Monte Carlo simulation generates 100,000 thirty-day paths for each starting state.

Starting in crisis gives a mean cumulative return of 0.03%, a median of 1.06% and a 2.5% tail of -24.44%. Loss probability is 46.0%. Starting in normal gives corresponding figures of 4.28%, 5.09%, -14.01% and 20.5%. Uncertainty falls between them, with a 28.1% loss probability. Even a normal start carries a 35.0% chance of visiting crisis within 30 days. The near-zero crisis-start mean may look surprising beside that state's -0.79% daily drift. Part of the explanation is that Pr(crisis) drops from 0.811 at one day to 0.387 at five days.

These scenarios have no earthquake conditioning. They use fitted parameters and assume certainty about the starting state.

Before this can govern exposure

Özel acknowledges most of the distance to a trading rule. The forecasts omit parameter uncertainty, transaction costs and macro covariates. The paper explicitly says they do not establish out-of-sample forecasting superiority. It also specifies a time-varying transition probability (TVTP) extension, through which quake covariates would directly affect transition probabilities, without estimating it. Three further problems appear in the paper and remain unresolved.

State count comes first. Preliminary information criteria favoured four states, the paper reports, but it retained three for interpretability. We did not find four-state estimates in the paper. That choice affects every figure above.

Then there is the labelling problem. The regimes use smoothed posteriors from the full sample, a method the paper itself describes as appropriate for retrospective identification. A live book would have filtered probabilities only. The 46.0% loss probability assumes it knows it is in crisis at day zero. Expected crisis duration is 5.29-day; a one- or two-day filtering lag would therefore leave much of the episode spent before a live book enters. We encountered the same lagging-flag problem in Behal's NSE regime study.

The remaining problem is tails. Emissions are Gaussian. Although the paper discusses a Student-t alternative, it does not estimate one. The -24.44% tail therefore remains unchecked against heavier-tailed emissions.

We could not test these issues ourselves. We have neither BIST 100 history nor an earthquake catalogue, which prevented us from rerunning the regime fit or the event tests.

A rolling backtest using filtered probabilities would change our view if the crisis-start loss probability remained well above the normal-start probability after allowing for label lag. Until then, Özel offers a careful account of Turkish volatility clustering and an inconclusive earthquake null (chi-square p = 0.655, with an expected crisis count of about 5.4). Neither gives us a trade.