Market Crash Models Are Moving From Prediction to Explanation

Since the beginning of modern financial markets in the Netherlands in the 1600s, financial crises and bubbles have been a regular occurrence, starting with the famous Tulip Mania. A direct consequence has been the recognition that understanding the conditions that can cause such crises is important, either for the state and regulators to reduce the occurrence and/or severity of crises, or for actors in the financial system to avoid suffering massive losses.

However, until now, the main method has been correlation-based prediction, like looking at metrics such as debt-to-GDP ratio, overvaluation metrics, or investor sentiment. All of these data can indeed correlate to conditions that can cause a crisis, the way dry kindling in a forest can cause fire.

This still does not provide information on what causes a given crisis, the same way that a forest fire starts because of an initial spark, not because of dry wood.

A new study by a researcher at the University of Szczecin in Poland argues that crisis analytics should shift toward models that explain which structural channels drive market collapses. In this paper, the study examines the role of volatility shocks and Treasury-yield shocks in causing financial crises.

It was published in Expert Systems with Applications^1, under the title “Predicting the unpredictable: a counterfactual causal inference framework for financial market collapse during black swan events“.

This can be important data for investors and risk managers as the usual stress tests built on average-market assumptions may underestimate losses when volatility regimes shift.

Predicting Financial Crises

Moving From Correlation To Prediction

A lot of the modern financial system is based on mathematical models that try to make sense of and predict risks. However, they are also based on mathematical assumptions, and abstract statistics rarely match real-life situations, leading to so-called black swan events, a term coined by Nassim Taleb, describing an unpredictable, rare occurrence that has a massive impact on society, economies, or financial markets.

“Conventional predictive models are effective at finding trends in big data, but they frequently fail in explaining why certain rare events occur or how the outcome would change in alternate conditions.

This is also why crises or brutal market moves are often described as “statistically impossible”. Except, of course, this only means that the correlation-based approach is inadequate to properly match real-life conditions.

This is an issue, as risk managers need to know not just that the market fell, but which structural channel drove the collapse.

Similarly, the central banks need to evaluate whether their instruments address the dominant transmission mechanism to reduce such risks.

So overall, stress-test designers need to parameterize sensitivities that are appropriate for extreme scenarios, not long-run averages.

This is why this study advocates for a different approach, called “counterfactual causal inference”, or the process of estimating what would have happened in an alternate, hypothetical reality.

To do so, the research paper used three design principles:

First, the model must be able to respond to an interventional question instead of only a predictive one:

“How would the cumulative crash trajectory have evolved in the absence of a specific shock channel?”

Second, each structural claim must be supported by at least one formal empirical test.

Third, the outcome must be verifiable, for example, with placebo testing on non-crisis periods.

Gathering The Data

The study used two main financial crisis events to demonstrate: the 2007-2009 Global Financial Crisis (GFC) and the COVID-19 pandemic.

A wide array of data was collected to analyze these two crises:

• The daily series of the S&P 500 Index.
• The Monthly Consumer Price Index (CPIAUCSL) for measure of inflation
• The Unemployment rate (UNRATE).
• The CBOE implied-volatility index (VIX).
• The U.S 10-year Treasury constant-maturity yield.
• The Moody’s Baa-corporate-yield-minus-10-year-Treasury credit spread (BAA10Y).
• The TED spread (TEDRATE).

What Causes Financial Crises?

Yield Shocks As Causes Of Crashes

The first part of the analysis looks at yield shocks, or a sudden, unexpected shift in interest rates or bond yields across financial markets.

In a graph observing what market returns would have looked like if the yield shocks were absent (blue line) and what actually happened with the yield shocks included (red line), the two datasets looked very close.

The divergence of the red and blue trajectories indicates a crucial aspect of direction–yields: they are actively influencing the stock returns rather than merely reacting to them.

“Rather, it is observed that the blue trajectory is above the red, especially during COVID crisis and more subtly during the GFC crisis. This demonstrates that changes in the yield are not outcome of the crash, rather it is the predictor.”

However, what determines the severity of the losses is driven by another cause.

Volatility As Loss Amplifier

The other cause of financial crisis, according to this study, is spikes in volatility.

This is maybe not a surprise, as this falls in line with a mainstream explanation for triggers of financial crisis: 1992’s Minsky’s Financial Instability Hypothesis.

Its core idea is that a false sense of security is causing financial actors to take on excessive, dangerous levels of debt. Ultimately, “stability is destabilizing”.

This was determined to have been the main cause of equity prices’ crash during both the GFC and the COVID pandemic.

“In COVID and GFC, the volatility channel accounts for 58.7 % and 28.3 % of the total cumulative drawdown, respectively, while the yield channel accounts for 8.4 % and 12.6 %.”

The study results also suggest that yield sensitivity is increased in turbulent markets, so the more volatility is already happening, the more yield shocks become impactful as well.

Predicting Crises More Accurately

Refining Predictions Further

This is not to say that the framework of the study is a perfect predictor. For once, it uses a linear model, which might not be ideal for extreme circumstances like financial crises.

“Future research should test a structurally non-linear model (such as a regime-switching SVAR or a neural structural equation model).”

Using extra data could also help improve the prediction capacities, like repo-rate dynamics, dealer-balance-sheet constraints, and options-microstructure variables.

Implications For Investors & Policymakers

For risk managers and designers of stress tests, this demonstration of causation and not just correlation should be a vital input to hedging decisions.

So models should use a two-regime sensitivity schedule: the low-VIX-regime coefficient for moderate situations and the high-VIX-regime coefficient (3–5 × greater) for extreme scenarios.

It also means that during different conditions, the focus should be on different metrics:

• When yield-driven losses are increasing in relative importance, duration management (monitoring and adjusting the time sensitivity of an asset or liability) becomes progressively more significant.
• When VIX-driven losses predominate, volatility-overlay strategies are the first-order priority.

For central banks, this means they can measure the benefit of additional equity-stabilizing rate-management instruments, depending on the degree of market panic that already exists.

For policymakers, understanding what kind of crisis they are dealing with is the most important.

“COVID was a 7:1 volatility-to-yield crisis, whereas GFC was a 2.3:1 crisis. As a diagnostic tool for determining which channel is most likely to prevail in real time, these ratios might be used to future Black Swan occurrences.”

Study Referenced

^{1. Guru Ashish Singh. Predicting the unpredictable: a counterfactual causal inference framework for financial market collapse during black swan events. Expert Systems with Applications. 15 December 2026. Article: 133342. Volume: Volume 331, Part C. DOI: https://doi.org/10.1016/j.eswa.2026.133342}