Forecasting annual excess stock returns via an adaptive network‐based fuzzy inference system

In this study, an adaptive network‐based fuzzy inference system (ANFIS) and a neural network were tested for the ability of these techniques to forecast the annual excess returns of three large publicly traded companies from a time series of said returns. The predictive ability of these techniques was compared with that of an autoregressive moving average (ARMA) model. The Fair–Shiller test was used in the comparisons in order to obtain results that were not subjective and so that conclusions could be made regarding the information used by the techniques in the generation of their forecasts. Since predictive ability does not translate to profitability, a simple trading strategy was used to determine the ability to generate profits from trading upon the forecasts of the respective techniques. As hypothesized, the ANFIS and neural network techniques are able to generate forecasts with significant predictive ability. However, neither technique dominates the other or the ARMA model. In tests of the ability of the techniques to generate profits from their forecasts, a simple trading strategy was used (trading on the predicted sign of the return). The ANFIS and the neural network generated profits in all of the trading scenarios. However, neither technique dominated the other, nor did they consistently outperform the traditional and naive models (strategies). The mixed results in the predictive ability tests and the profitability tests indicate that the conclusions from the study differ based upon the context in which the forecasts are used. Copyright © 2005 John Wiley & Sons, Ltd.