Predicting Volatility : Getting the Most Out of Return Data Sampled at Different Frequencies

We use the MIDAS (Mixed Data Sampling) approach to study regressions of future realized volatility at low-frequency horizons (one to four weeks) on lagged daily and intra-daily (1) squared returns, (2) absolute returns, (3) realized volatility, (4) realized power and (5) return ranges. We document first of all that daily realized power and daily range are surprisingly good predictors of future realized volatility and outperform models based on realized volatility. Moreover, MIDAS models with daily data - range, realized power, realized volatility - require a polynomial with at least 30 days. We document that high-frequency absolute returns are also better at forecasting future low frequency realized volatility than high-frequency squared returns. We also discuss many issues that are encountered in practice, such as long memory and seasonality.All the results are based on a commonly used FX data set