A Volatility-Aware Temporal Transformer for Intraday Risk Forecasting with Market Microstructure Signals

Abstract

The accurate forecasting of intraday financial risk is a cornerstone of modern algorithmic trading and systemic stability analysis. Traditional econometric models often fail to capture the nonlinear dependencies and rapid regime shifts characteristic of high-frequency limit order book (LOB) data, while standard deep learning architectures frequently struggle with the extremely low signal-to-noise ratio inherent in microstructure signals. This paper introduces the Volatility-Aware Temporal Transformer (VATT), a novel deep learning architecture designed specifically for intraday realized volatility forecasting. Unlike canonical Transformers, VATT incorporates a specialized Volatility Gating Module (VGM) that dynamically modulates the attention mechanism based on the prevailing market regime, allowing the model to distinguish between transient noise and structural volatility shifts. We leverage granular microstructure signals, including Order Flow Imbalance (OFI) and depth-weighted spread, to enhance the feature space beyond simple price history. Extensive experiments conducted on tick-level data for major equity indices demonstrate that VATT significantly outperforms GARCH-family baselines and standard Long Short-Term Memory (LSTM) networks in terms of Mean Absolute Error and Quasi-Likelihood loss. The results suggest that integrating volatility-specific inductive biases into the Transformer architecture is crucial for robust risk forecasting in high-frequency domains.