Time-Series Momentum Portfolios with Deep Multi-Task Learning
Log in to collectAcademic paper
Constructing Time-Series Momentum Portfolios with Deep Multi-Task Learning
J. M. Joel Ong; Dorien Herremans
- SGSingapore University of Technology and Design
Strategy in a nutshell
Construct long-short time-series momentum portfolios across multiple futures using deep multi-task learning with LSTM and feedforward networks. Portfolios are volatility-scaled, trained annually, and optimized for out-of-sample risk-adjusted returns.
Economic rationale
Multi-task learning improves portfolio performance by jointly modeling momentum and auxiliary tasks like volatility forecasting. This method captures complex patterns, yielding superior out-of-sample risk-adjusted returns, even during market crises.
Backtest performance
Annualised return7.9%
Volatility9.75%
Sharpe ratio0.81