VOLATILITY FORECASTING WITHOUT DATA-SNOOPING

Data-snooping arises when the properties of a data series influence the researcher's choice of model specification. When data has been snooped, tests undertaken using the same series are likely to be misleading. This study seeks to predict equity market volatility, using daily data on U.K. stock market returns over the period 1955-1989. We find that even apparently innocuous forms of data-snooping significantly enhance reported forecast quality, and that relatively sophisticated forecasting methods operated without data-snooping often perform worse than naive benchmarks. For predicting stock market volatility, we therefore recommend two alternative models, both of which are extremely simple. © 1990.