DeepUnifiedMom introduces a deep learning framework that unifies time-series momentum strategies across multiple time horizons using multi-task learning and a multi-gate mixture of experts architecture. The model consists of shared LSTM experts that extract common features, task-specific gating networks that selectively combine expert outputs for fast, medium, and slow momentum predictions, and a Capital Allocation Network that learns to weight these three momentum portfolios to form a single unified portfolio. The training process minimizes a combined loss: RMSE for each task-specific network and a soft-capped negative Sharpe ratio for the capital allocation network, which helps prevent overfitting to noisy financial data.

Extensive backtesting on 49 futures contracts across equities, fixed income, foreign exchange, and commodities from 1990 to 2023 shows that DeepUnifiedMom consistently achieves higher risk-adjusted returns than traditional time-series momentum strategies and other deep learning baselines, even after accounting for transaction costs. The unified approach effectively captures momentum opportunities across different trend speeds, leading to improved diversification and reduced drawdowns. This work is the first to apply multi-task learning with multi-gate mixture of experts to portfolio construction, offering a practical tool for practitioners seeking to exploit the full spectrum of momentum in financial markets.