Original title: Explainable Deep Learning for Price-Trade Dynamics: From Black-Box Forecasts to Effective Parametric Models
Manuel Naviglio, Fabrizio Lillo · September 2026
Stock prices and trade volumes interact in complex, changing ways that simple straight-line math cannot catch. Deep neural networks predict these moves much better, but they work like black boxes where nobody knows how they decide. Researchers trained a deep feed-forward neural network on high-frequency stock data to uncover how these trades really work. They used an explainability tool called Shapley values to look inside the network and break down its choices. They found that recent trade events matter most. Prior trading volume pushes prices in the same direction, but this impact levels off. Meanwhile, recent price changes decide what happens next. If a trade did not move the price, the momentum continues. If the price did change, that push weakens or turns around. Using these rules, the team built a simple, clear math model that beats standard linear tools and nearly matches the complex neural network.
Imagine trying to predict whether a friend will laugh at a joke. A simple rule like 'if the joke is about cats, they laugh' works most of the time but misses when they are in a bad mood. A complex model that tracks their mood, the time of day, and who told the joke is more accurate but you cannot explain why. In this paper, the researchers use a complex model to find which factors matter most, then build a simple rule that captures the key patterns without the complexity.
This matters because it bridges the gap between black-box forecasting and economic interpretability. Traders and regulators can now understand the nonlinear forces driving price movements without relying on opaque models. The resulting parametric model offers both accuracy and transparency, making it practical for real-time analysis and policy decisions.
The analysis is limited to high-frequency data for large- and small-tick stocks; results may not generalize to other assets or lower frequencies. The SHAP-based parametric model, while interpretable, may not capture all nonlinear dependencies present in the full neural network. The study does not address out-of-sample stability or transaction costs.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Explainable Deep Learning for Price-Trade Dynamics: From Black-Box Forecasts to Effective Parametric Models — Manuel Naviglio, Fabrizio Lillo
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.