Earnings per share forecasts: linear and neural networks models

In this study, we compare different input datasets and forecast methodologies to predict earnings right before they get announced. We start with a time-series approach using only 5 past earnings per firm to avoid survivorship bias. Then we add in turn, analysts’ forecasts, market and macro-economic data, and firm specific data to the predictors list. We finally test all these datasets together in a kitchen sink approach. We compare our forecast errors with the simple time-series approach for both linear method and neural network to pick up any potential non-linearities. We find that the best prediction is provided by linear methods and that the analysts' forecasts dataset adds the most predicting power as analysts incorporate in their estimates all information available at the market, economic and firm level. Because models tend to overfit, we observe that neural networks and large predictors datasets are outperformed by simpler models with fewer predictors