A Regime-Switching Neural Network for Equity Momentum and Reversal

Summary

The paper describes a deep-learning approach to portfolio selection in US equities. Its residual switching network uses one module to identify market conditions and another to produce momentum and reversal forecasts, switching between the two types of signal as regimes change. The architecture combines residual networks, stacked blocks intended to help control overfitting in noisy financial data, and an attention mechanism intended to strengthen prediction.

For the period from 2008 through the first half of 2017, the authors report an average annual Sharpe ratio of 2.22 for the proposed strategy, compared with 0.81 for an artificial-neural-network strategy and 0.69 for a linear model. These are the reported out-of-sample results; the excerpt does not give details about portfolio construction, trading costs, benchmark implementation, or robustness across other periods and markets. The figures therefore describe the study’s tested setting and do not establish that the method will retain its performance in live trading or under different conditions.

Key ideas

• The network uses a switching module to identify market regimes and a separate module to predict momentum or reversal.
• Residual blocks are used to help control overfitting in noisy financial data.
• The authors report stronger out-of-sample Sharpe performance than two benchmark models over the stated sample.
• The excerpt does not describe trading costs or performance outside the tested period.

Tags

# Residual Switching Network for Portfolio Optimization


# Residual Switching Network for Portfolio Optimization









This paper studies deep learning methodologies for portfolio optimization in the US equities market. We present a novel residual switching network that can automatically sense changes in market regimes and switch between momentum and reversal predictors accordingly. The residual switching network architecture combines two separate residual networks (ResNets), namely a switching module that learns stock market conditions, and the main module that learns momentum and reversal predictors. We demonstrate that over-fitting noisy financial data can be controlled with stacked residual blocks and further incorporating the attention mechanism can enhance powerful predictive properties. Over the period 2008 to H12017, the residual switching network (Switching-ResNet) strategy verified superior out-of-sample performance with an average annual Sharpe ratio of 2.22, compared with an average annual Sharpe ratio of 0.81 for the ANN-based strategy and 0.69 for the linear model.

Shown in full with attribution under the source's licence. Licence: abstract CC0

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