Evaluating Methods for Forecasting Earnings Per Share

Abstract

In this study, the relative accuracy of four well known methods for forecasting are compared The methods are applied to the time series of earnings per share for a random sample of United States corporations over a lengthy period of time. All the time series exhibit both period‐to‐period movements and seasonal fluctuation. The four models are, (1) Holt‐Winters multiplicative exponential smoothing model, (2) univariate Box‐Jenkins model, (3) linear autoregression of data seasonally adjusted by the Census II–XII method, and (4) linear autoregression of the data seasonally adjusted by the X11‐ARIMA method. The study of financial data of this type is important because (1) these data exhibit time series properties of trend, seasonality, and cycle, (2) earnings per share forecasts are important for purposes of financial planning and investment; and (3) previous studies of this nature were not as exhaustive in terms of the statistical analysis of the results