Macro–Market Fusion with Cross-Attention for Equity Return Prediction

Macroeconomic conditions are widely believed to influence the direction of equity markets, yet most forecasting models either ignore macroeconomic information or incorporate it through a small set of ad hoc predictors. We propose XAttnFusion, a macro–market fusion architecture that jointly learns from high-frequency market data and lower-frequency macroeconomic time series for equity return prediction. The model comprises three branches: a 1D convolutional network that encodes 40-day market windows (price, volume, and technical indicators), a temporal convolutional network that encodes 24-month macro sequences, and a feedforward branch for volume-at-price structure features. These representations are integrated through multi-head cross-attention, in which the current market state queries the macro sequence to produce a fused representation for directional forecasting. We evaluate XAttnFusion on daily SPY returns from 2012 to 2024 using purged cross-validation with a 5-day embargo to prevent information leakage. To address potential look-ahead bias from macroeconomic publication lags, all macro inputs are lagged by two months. The model achieves a mean out-of-sample AUROC of 0.63±0.05, representing a 27% improvement over random and an 8.1% improvement over the best concatenation baseline. In a fair comparison where each model is independently hyperparameter-tuned, cross-attention fusion improves AUROC by 0.047 over concatenation (p=0.031, Wilcoxon signed-rank test). The model also generalizes to QQQ and IWM, where cross-attention consistently outperforms concatenation fusion. Crucially, the model’s discriminative ability is state-dependent, indicating that the value of macro–market fusion is itself conditioned on market structure. Permutation-based feature importance shows that macro and market branches contribute on a comparable scale (approximately 48% and 36%, respectively), so the gains come from jointly fusing two comparably weighted sources rather than from a single dominant input. Our results show that explicitly modeling macro–market interactions with interpretable attention improves predictive accuracy over naive fusion strategies and provides insight into the time-varying relevance of macroeconomic information in financial forecasting and equity market prediction.