An analysis of neural-network forecasts from a large-scale, real-world stock selection system

LBS Capital Management employs a system for managing several large investment portfolios that is founded on financial engineering principles. Over three thousand neural networks form the backbone of this system. Network forecasts spanning several weeks in the recent past are analyzed with respect to their horizon as well as their accuracy. In comparing 4-week and 12-week risk-adjusted excess return or alpha forecasts, the 13-week forecasts appear more accurate for the time period studied. Splitting the stock universe according to the magnitude of actual alpha exposes certain asymmetries in the forecasts. Using a relatively large number of observations, some preliminary conclusions are drawn. The relevance of these conclusions is not confined to neural network stock selection models.