Researchers often try to forecast corporate earnings by feeding historical data about metrics such as earnings, sales, and tax rates into mathematical models. Two Booth researchers—Associate Professor Joseph Gerakos and Assistant Professor Robert B. Gramacy—believe that researchers could get better results by simplifying their models and returning to methods popular 20 years ago.
In a working paper called “Regression-Based Earnings Forecasts,” Gerakos and Gramacy compare forecasting models that incorporate 40 years worth of historical data about a company’s activities and performance. To allow the models to be evaluated on a level playing field, they adjust the data to account for conditions, such as inflation, that influenced corporate results. They then compare the models’ predictions to actual corporate earnings to see which gives the most accurate forecasts.
They find that simpler models outperform more complex ones, even when more variables seem necessary to examine a study subject. Their findings are at odds with the current trend in forecasting, which calls for more complicated models. The researchers consider their paper a roadmap for people creating new forecasting applications.