From Constituents to Index: Interpretable Price Movement Prediction via Cross-Asset Order Flow

Abstract

This study proposes a novel deep learning architecture for intraday index price movement prediction, designed to capture the spatio-temporal dependencies between a market index and its constituents. Our interpretable, multi-stage framework addresses the limitations of single-asset black box models by: (i) dynamically selecting the 30 most influential constituents, (ii) forecasting each selected asset's future order flow imbalance(OFI) sequence, and (iii) fusing the predicted OFI tensor through a hierarchical Stock-Time Attention block to classify the short-term index direction. Extensive experiments on KOSPI 200 data demonstrate that our model significantly outperforms existing baselines. The results confirm that incorporating cross-asset information, particularly from constituent futures, is critical for enhancing forecast accuracy. We further validate our framework's interpretability by showing that the learned attention weights effectively identify primary market drivers. Assets with high attention scores exhibit price patterns that are quantitatively more similar to the aggregate index.