Stock Return Outliers
17 Pages Posted: 13 Nov 2017
Date Written: November 9, 2017
Abstract
Standard deviations and market-betas based on winsorized rates of return predict their own future realizations better than equivalents based on unwinsorized rates of returns. A good prescription is to winsorize rates of return around plus and minus 10- 15%, especially for samples of all CRSP stocks. There is no meaningful predictive gain using stocks’ own means and standard deviations and/or contemporaneous market rates of return; and eliminating or zeroing returns rather than winsorizing them. Winsorizing is better than not winsorizing.
This note then investigates two simple research questions using treated returns:
1. Winsorized log-volatility estimates perform much better than unwinsorized log- volatility estimates in forecasting, and they add marginal predictive power to option-implied volatility forecasts.
2. Annual average stock return predictions with past averages can be very sensitive to winsorization choices (including not winsorizing, which is also implicit in common monthly rates of returns). This evidence suggests that the seemingly second-order innocuous choice of where to winsorize (or not) can unfortunately have important inference consequences.
Suggested Citation: Suggested Citation