Constructing Time-Series Momentum Portfolios with Deep Multi-Task Learning

Abstract: A diversified risk-adjusted time-series momentum (TSMOM) portfolio can deliver substantial abnormal returns and offer some degree of tail risk protection during extreme market events. The performance of existing TSMOM strategies, however, relies not only on the quality of the momentum signal but also on the efficacy of the volatility estimator. Yet many of the existing studies have always considered these two factors to be independent. Inspired by recent progress in Multi-Task Learning (MTL), we present a new approach using MTL in a deep neural network architecture that jointly learns portfolio construction and various auxiliary tasks related to volatility, such as forecasting realized volatility as measured by different volatility estimators. Through backtesting from January 2000 to December 2020 on a diversified portfolio of continuous futures contracts, we demonstrate that even after accounting for transaction costs of up to 3 basis points, our approach outperforms existing TSMOM strategies. Moreover, experiments confirm that adding auxiliary tasks indeed boosts the portfolio's performance. These findings demonstrate that MTL can be a powerful tool in finance.

• The paper introduces a novel deep multi-task learning framework that jointly optimizes time-series momentum signals and risk management.
• It employs LSTM networks for feature extraction and feedforward layers to integrate multiple volatility estimators, enhancing portfolio performance.
• Empirical results demonstrate improved annualized returns, Sharpe ratios, and reduced drawdowns compared to traditional momentum strategies.

Constructing Time-Series Momentum Portfolios with Deep Multi-Task Learning

The paper "Constructing Time-Series Momentum Portfolios with Deep Multi-Task Learning" by Joel Ong and Dorien Herremans presents an innovative approach to the construction of time-series momentum (TSMOM) portfolios utilizing deep multi-task learning (MTL) frameworks. This work bridges the gap in existing TSMOM strategies by integrating the process of return generation and risk management through deep neural network architectures.

Summary of the Research

Momentum strategies, and in particular time-series momentum, have been central to quantitative finance research over the past few decades. A key challenge in developing TSMOM portfolios is the effective estimation of both the momentum signal and the volatility estimator, which have traditionally been treated independently. This paper proposes a novel approach using MTL to jointly learn and optimize portfolio construction alongside various auxiliary tasks related to volatility estimation.

The authors implement a deep learning architecture, specifically utilizing Long Short-Term Memory (LSTM) networks for feature extraction with subsequent feedforward networks for task-specific outputs. This architecture concurrently tackles the main task of portfolio construction and auxiliary tasks such as forecasting volatility using different estimators—close-to-close, Parkinson, Garman-Klass, Rogers-Satchell, and Yang-Zhang.

Key Findings

The empirical analysis, using data from January 2000 to December 2020, demonstrates that the proposed MTL framework results in portfolio performance that exceeds traditional TSMOM strategies. After accounting for realistic transaction costs of up to 3 basis points, the MTL-based strategy achieves a higher annualized return and demonstrates a favorable risk-return profile as judged by metrics such as Sharpe and Sortino ratios. Notably, the proposed model also shows a reduced maximum drawdown period and a shorter recovery time compared to benchmarks, which highlights its improved robustness against market stress.

An ablation study within the research elucidates the contribution of each auxiliary task to the overall model performance. It suggests that no single auxiliary task significantly diminishes performance, but rather the combination of various volatility-related tasks enhances the main task's effectiveness.

Implications and Future Directions

The implications of integrating deep MTL into financial asset management are profound. The approach offers a more unified treatment of risk and return, capturing the synergies between volatile market environments and traditional momentum indicators. Additionally, the ability of the MTL model to provide a lower correlation to broad equity indices, such as the US MSCI Total Return Index, positions this strategy as an attractive diversification tool for portfolio managers.

Future developments could include the introduction of more nuanced task weighting mechanisms or the examination of alternative auxiliary tasks which capture different facets of market risk, thereby possibly enhancing model generalizability and effectiveness. Continuous exploration into the relationship between MTL and different asset classes or market conditions could further augment the utility of this strategy in diverse financial contexts.

In summary, this paper's contribution lies in its innovative use of deep learning architectures to improve traditional portfolio construction methodologies. The evidence suggests that multi-task learning, through the lens of deep models, could play a critical role in developing more adaptive and resilient financial strategies.