Forecasting models generally do a very poor job in predicting future returns in financial markets which are characterized by high levels of noise. It is commonly known that predictions of typical autoregressive time series forecasting models, for example, are usually very conservative – meaning close to the mean – because of the inherent difficulty of making accurate forecasts. This situation is particularly discouraging since it is the large future values that we especially care about predicting accurately. Machine learning methods provide a potential solution to this problem by providing multiple models that each of which makes forecasts of different magnitudes. The obvious problem with this approach is overfitting, especially if the learner has no constraints on how complex its models can be. In this talk, I present the tradeoffs between increased the model complexity and forecast accuracy based on a mix of trading results and current research.