Forecasting Equity Returns and Volatility with Regime-Switching Partial Least Squares

A Regime Switching Partial Least Squares Approach to Forecasting Industry Stock Returns: Using monthly stock returns on 16 industry portfolios next to the S&P500 com- posite stock return index as forecasting targets, this paper shows that a data reduction technique which incorporates regime dependent forecasting power of various macroeconomic and  nancial predictors produces positive and signi cant (at 0:05 and 0:01 levels) as well as economically valuable OOS results for a majority of the 16 portfolios, as well as for the composite stock index. Direct tests show that incorpo- rating regime dependent predictability produces signi cantly better results relative to linear models. Furthermore, it is shown that macroeconomic predictors contain forecasting information above and beyond the information contained in  nancial predictors alone. In fact, in the employed setup, macroeconomic predictors forecast well on their own and even better in conjunction with  nancial variables. This is true for symmetric preferences in the form of a mean-variance investor and also when asymmetric preferences in the form of a CRRA investor are employed.

A Regime Switching Partial Least Squares Approach to Forecasting Realized Industry Equity Volatility: Using monthly industry-level realized stock volatilities on 16 industry portfolios next to the S&P500 composite index as forecasting targets, this paper shows that a dimension reduction technique that incorporates regime dependent forecasting power of  nancial predictors produces positive and highly signi cant OOS results relative to an AR(6)-benchmark for most of the 16 industries, as well as for the composite index. This is not the case when regime dependent predictability is not accounted for. In addition to regime switching, which is estimated using a Bayesian approach, the auto-regressive structure of realized volatility is incorporated directly into the dimension reduction technique. Results are robust with respect to estimation window size and OOS validation period. However, individual industries seem to di er from each other with respect to their underlying data generating processes. Furthermore, it is shown that in the current setup macroeconomic predictors contain little forecasting information when used individually and worsen results when used jointly with  nancial predictors.