Regime-Aware Explainable Deep Learning and Ensemble Learning for Multi-Asset Tactical Allocation: Transformer-LSTM Evidence from Cross-Asset ETFs

Abstract

Financial return forecasting remains difficult because asset prices are noisy, non-stationary and exposed to changing market regimes. This paper develops a regime-aware and explainable machine-learning framework for multi-asset tactical allocation using five exchange-traded funds representing equities, long-term bonds, short-term bonds, gold and crude oil. The framework compares Elastic Net, Histogram Gradient Boosting, Multi-output Long Short-Term Memory and Transformer Encoder models using 60-day historical feature windows to forecast 25-day cumulative returns. The feature set incorporates daily returns, momentum, realised volatility, return z-scores, macro-market indicators and cross-asset dependence measures. The models are evaluated through forecasting metrics, directional accuracy, cross-sectional rank information, transaction-cost-aware portfolio backtesting, ablation experiments and grouped permutation importance. The results show that the multi-output LSTM provides the strongest average forecasting accuracy, while the Histogram Gradient Boosting model delivers the best transaction-cost-adjusted allocation performance. At 5 basis points transaction cost, the best strategy outperforms the equal-weight benchmark in total return, annualised return, Sharpe ratio and maximum drawdown. The results of the explainability analysis provide insight into the performance-relevant features. Most relevant for predicting the performance based on return data are the features in the group of realised volatility, i.e., the risk signals that are regime-sensitive. Although deep learning can extract complex sequential structures from return data, the findings suggest that simpler ensemble models may provide more robust tradable allocation signals when forecasts are converted into transaction-cost-adjusted portfolio decisions.