Momentum strategy provides distinguishable risk-adjust returns and a hedging mechanism for left-tail events. For long, momentum also has a caveat of significant drawdown. Therefore volatility-based position sizing is often implemented in addition to cross-sectional and time momentums. This article further examines a deep-learning approach to combine the two steps.
The article looks into a portfolio of 88 continuous futures contracts. The outcome shows a promising result before considering transaction costs: the Long-Short Term Memory Sharpe-optimized deep network generates 2.804 Sharpe Ratio from 1995 to 2015.
As we have seen a similar enhancement in Gu, Kelly, and Xiu (2017) for a machine-learning-based stock long-short portfolio with feed-forward neural network implement. We are excited to initiate research to combine the two outstanding works.
Enhancing Time-Series Momentum Strategies Using Deep Neural Networks, B. Lim, S. Zohren, and S. Roberts, The Journal of Financial Data Science, 1 (4)
Gu, S., B. T. Kelly, and D. Xiu. 2017. “Empirical Asset Pricing via Machine Learning.” Chicago Booth research paper No. 18-04, 31st Australasian Finance and Banking Conference.