AI Reveals Hidden Market Shocks in Stock Data

Financial markets are not just about smooth trends and gradual changes; they are punctuated by sudden, unpredictable shocks that can upend portfolios in moments. A new study analyzing high-frequency data from the New York Stock Exchange (NYSE) reveals that these abrupt shifts, often overlooked by conventional models, are the dominant force behind market volatility. Researchers have developed a framework that separates the continuous, diffusive movements from the discontinuous jumps, providing a clearer picture of how markets really behave and why risk management often falls short.

The key finding is that for most statistical models of volume-price data—a composite measure combining price and trading volume—the scale parameter, which reflects the magnitude of fluctuations, is driven primarily by jump-diffusion dynamics. This means rare, large discontinuities account for a significant share of total variance, between 40% and 63% for models like Gamma, Inverse Gamma, and Weibull distributions. In contrast, the shape parameter, which describes the distribution's form, behaves as a pure diffusion with linear mean reversion, indicating smoother, more predictable adjustments. The log-normal model shows an inverted pattern, where the scale is diffusive and the shape exhibits weak jump signatures, highlighting how logarithmic transformations alter stochastic behavior.

Ology involved analyzing 1,750 NYSE-listed companies over 976 trading days, with data sampled every 10 minutes. Volume-price distributions were fitted to four models: Gamma, Inverse Gamma, Weibull, and Log-Normal, each parameterized by shape (φ) and scale (θ) parameters. Daily trends were removed using a 21-day moving average to focus on stochastic fluctuations. The researchers then applied Kramers-Moyal (KM) coefficient analysis, estimating coefficients up to the sixth order to classify dynamics as either diffusive or jump-diffusion. This involved verifying the Markov property, using adaptive binning for state-space discretization, and computing conditional moments to extract infinitesimal moments, with corrections for finite-lag biases.

, Detailed in figures such as Figure 3 and Figure 4, show stark contrasts. For the Gamma model's scale parameter (θ), fourth- and sixth-order KM coefficients reach magnitudes as high as 10^22 and 10^33, with a diagnostic ratio D^(4)/D^(2) far exceeding 0.1, confirming jump-diffusion. In contrast, the shape parameter (φ) has negligible higher-order terms, with D^(4)/D^(2) below 0.10, indicating pure diffusion. Similar patterns hold for Inverse Gamma and Weibull models, while the Log-Normal model shows the opposite, with θ being diffusive and φ showing weak jumps. Global jump parameter inversions, summarized in Table D4, reveal jump rates of 4–15 events per unit time for θ, contributing 40–63% of variance, whereas φ has lower rates of 1–2 and 14–22% variance shares.

This research matters because it s traditional financial models that assume continuous price movements, such as Geometric Brownian motion, which struggle to account for extreme events like market crashes. By quantifying the role of jumps, the framework offers practical for risk management, enabling better volatility controls, stop-loss placement, and stress scenario analysis. For everyday investors and regulators, it underscores that market stability is often an illusion, with hidden discontinuities driving much of the observed risk. The split between shape and scale dynamics suggests that while structural signals for trading execution may be diffusive, volatility is jump-prone, necessitating more robust safeguards.

However, the study has limitations. The analysis relies on high-order moment estimates, which are sensitive to tail events and finite sampling, potentially biasing jump parameter inversions, especially for the Log-Normal model. The dataset, while extensive, covers only NYSE equities over a specific period, and may not generalize to other markets or timeframes. Additionally, assumes weak stationarity after detrending, but residual intraday heterogeneity could affect estimates. Future work with larger datasets, alternative detrending s, and cross-market validation would strengthen these and enhance microstructural interpretations.