| This study evaluates the relative performance of four nonlinear methods using linear regression as a benchmark in forecasting the following “dependent” variables: next quarter's earnings surprises, abnormal returns on the earnings announcement day, and the risk-adjusted 20 day return. The four nonlinear methods are the artificial neural network (ANN), genetic algorithm (GA), classification and regression tree (CART), and naïve bayesian method (NB).; We use two measures to compare each method's accuracy. The first measure is the total number of times that a method gives significant forecasts. The second measure is the total number of times that a method receives the highest evaluation score in its forecasts. GA has the highest accuracy based on these two measures. NB ranks second. We provide possible explanations for these findings.; The research also addresses a number of issues that are not adequately covered by previous event studies, namely, small universe, short time period, survivorship bias, no out-of-sample test, and assumption of linearity.; The research uses several types of factors to forecast the dependent variables, namely, past earnings forecast errors, firm size, expectation measures such as earnings revision/estimates, and value measures such as cash flow based return on investment. One of our findings is that smaller companies tend to have larger earnings surprises, either positive or negative, than those of bigger companies. This is in accordance with prior research which has proposed that the amount of private predisclosure information production and dissemination is an increasing function of firm size and the amount of “unexpected” information conveyed to the market by actual earnings reports is inversely related to firm size. Furthermore, we confirmed that earnings forecast errors show consistently strong serial correlation. Again, this is in accordance with prior research which states that analysts underestimate the persistence of earnings forecast errors when revising their earnings forecasts.; Finally, we applied the most accurate prediction model, that based on GA, to implement a systematic trading strategy. Since the GA's patterns are rules, it is possible to interpret them. This is a major advantage of GA over other methods. Furthermore, we compared the 20-day return of the combined long rule from GA with the benchmark, 20-day return of SP500 index. The statistical results reject the null hypothesis that the 20-day return of the long strategies is NOT better than the benchmark.; We conclude our analysis with a discussion on the possible explanations for the superior results with the GA, and more generally, discuss how it can be a valuable tool for theory building. |
