Beating a Random Walk

• In Event: Forecasting Profitability

Abstract

Until recently, the accounting and finance literatures generally used either a random
walk or analysts' estimates as forecasts of accounting earnings. Hou et al. [2012]
suggested the use of a cross-sectional forecasting model and several studies immediately used this model because of the obvious advantage that forecasts can be formed
for a sample that is much greater than the sample of firms covered by analysts. Unfortunately,
subsequent studies have shown that the Hou et al. [2012] forecasts are
not significantly better than random walk forecasts. We present a simple modification
of Hou et al. [2012] - the use of quantile rather than OLS regressions in the prediction
model - that produces forecasts signicantly better than random walk forecasts.
Quantile regressions are intuitively appealing because these regressions minimize the
sum of the absolute deviations, which is consistent with the use of the absolute difference between forecast and actual as the indicator of accuracy. OLS regressions, on the
other hand, minimize the sum of squared deviations, which is not consistent with this
indicator.

Authors

• Peter Easton, University of Notre Dame

• Peter Kelly, University of Notre Dame

• Andreas Neuhierl, University of Notre Dame