Deep Learning for Risk Forecasts in a Long–Short Equity Strategy

Summary

The article describes using a deep neural network to estimate risk for an automated, beta-neutral long–short equity strategy. Financial statement data and historical prices serve as inputs; the model predicts return quantiles over horizons from five to 90 days. Wider predicted quantile ranges represent higher estimated risk. The model is compared with forecasts derived from historical volatility, using quantile coverage error on test data.

The authors report lower average coverage error for the neural network and then apply it to risk assessment in the strategy’s long and short selections. On constituents of the S&P 900, they report improved returns and Sharpe ratio, with lower volatility, drawdown, and value at risk. The document supplies no numerical results for this strategy comparison or detailed information about the data split, costs, or robustness. Its findings therefore describe one reported case study, not evidence that the model will generalize across markets or strategies.

Key ideas

• The model uses financial statements and historical prices to forecast return quantiles at multiple horizons.
• Quantile coverage error is used to compare its forecasts with historical-volatility estimates on test data.
• The neural network is reported to improve average coverage error over the simpler benchmark.
• In one S&P 900 long–short application, the authors report better returns and risk measures, but give limited robustness details.

Tags

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