Deep Momentum Networks for Risk-Aware Multi-Asset Trend Following

Summary

The article explains the Deep Momentum Network (DMN) approach to multi-asset trend following. Rather than predict returns and determine portfolio weights separately, the model outputs each asset’s next-period position, bounded between long and short exposure. Training directly optimizes a Sharpe-based objective, which incorporates portfolio returns and risk. The article uses an LSTM as an example and discusses why batches must preserve sequences of signals and returns so the objective can be calculated over time. It also describes validation-based early stopping and a rolling training schedule with separate training and validation periods.

Inputs include normalized returns over different horizons and MACD. The article reports that the original study achieved a Sharpe ratio near three over nearly 20 years of historical testing, and says adding Gaussian-process changepoint scores improved reported Sharpe and drawdown relative to the original model. These figures are reported secondhand; the article does not provide full test details, costs, or independent replication. It also notes that the proposed batch construction is the author’s interpretation because the paper did not specify those details.

Key ideas

• DMN predicts position sizes directly, combining directional signals and portfolio weights in one model.
• The model uses a Sharpe-based loss to optimize portfolio risk-adjusted performance rather than return prediction error.
• Batches need ordered sequences of signals and returns to calculate the multi-period portfolio objective.
• The described training setup uses validation data for early stopping and rolling retraining.
• Changepoint scores are proposed as additional features to help the model respond to trend reversals.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.