Multicollinearity - the presence of high correlation between predictor variables, time varying volatility, and leptokurticity - the tendency of extremes to appear more often than would be observed from the normal distribution are empirical properties of financial time series. In time series regression multicollinearity, time varying volatility, and leptokurticity of predictors can lead to bias and/or lower precision of forecasts. Multicollinearity is typically addressed by constructing factor variables from correlated predictors. It is shown that factors derived from a variant of the Partial Least Squares method can yield higher precision forecasts than forecasts based on the common Principal Component dynamic factor framework. Additionally, a method of applying a fractional power transformation to predictor variables is developed to mitigate the influence of time varying volatility and extreme values on regressors and estimated coefficients. Applications are made to asset return and realized volatility/correlation forecasting, with extensions to dynamic hedging.