Stock Return Prediction by History Mapping

Stock markets proved to be statistically predictable on an economically interesting scale over the past decade by fully data driven automatically constructed maps that associate to a set of new factor values a return prediction that is the average of historically observed returns for an area in factor space around the new values. Partitioning of the factor space in suitable areas is easily done by regression trees. Annual updating improves performance.Such an updated model with five value and two momentum factors on World Industrials showed out-of-sample over the past decade and after transaction costs more than 90% probability of best to worst quintile outperformance within six months, and 90% probability of more than 15% annual outperformance over ten years. Such models deliver excellent return/risk tradeoff and lower risk by internal diversification of alpha sources, exploiting linear as well as non-linear multifactor anomalies.The choice of factors requires economic expertise and insight, both enhanced by single factor maps and projections of regression tree models that allow to interpret the contents of the noisy and extensive tree models. Such maps as function of time show for instance that momentum factors ceased to work over the past decade while value factors proved reliable