Explainable Deep Learning for Price–Trade Dynamics: From Black-Box Forecasts to Effective Parametric Models

Battle Count: 77

Highly relevant to quantitative trading. The paper directly addresses the joint dynamics of prices and trades, which is fundamental to algorithmic execution, market making, and short-term trading strategies. The SHAP-inspired parametric model provides an interpretable framework for understanding nonlinear price impact and order-flow persistence. The distinction between large-tick and small-tick regimes is practically important for execution strategy design. The model's ability to capture state-dependent effects (continuation vs. reversal based on whether previous trade moved the price) has direct implications for trading signal generation. The parsimonious parametric form is suitable for real-time implementation and calibration.

Moderate complexity. The neural network architecture is straightforward (3 hidden layers, standard activations). SHAP computation using DeepExplainer is well-supported by existing libraries. The parametric model involves nonlinear least squares with reparametrization constraints (softplus for positivity), which is standard but requires careful implementation. The multi-lag extension with exponential decay adds some complexity. Main challenges include: proper data preprocessing from LOBSTER, handling the binning/clustering procedures for response surfaces, ensuring numerical stability in the parametric calibration, and the computational cost of SHAP value computation for 40 input features (20 lags × 2 variables).

The paper provides detailed architecture specifications (3 hidden layers with 64-64-32 neurons, tanh activations, Adam optimizer, MSE loss), clear data preprocessing steps (LOBSTER data, execution events only, quantile filtering, 50/50 train-test split), and explicit parametric model equations. Data source (LOBSTER) is publicly accessible. However, no code repository is provided, and specific hyperparameter tuning details (learning rate, batch size, epochs) are not fully specified. The SHAP computation uses the DeepExplainer method from the SHAP library.