Exploring Linkages in Asset Returns Using Machine Learning Techniques

I examine how linkages in asset returns can be used to forecast company returns. I expand on the methodology employed by Rapach et al. (2015) and predict next month industry returns or individual stock returns using returns of various asset classes. With industries as explanatory variables, a portfolio that goes long (short) in companies with high (low) predicted standardized return has a significant annualized alpha of 9.35 percent if one controls for common equity risk factors. The top bucket accounts for most of the generated alpha, because the training window of a company contains relatively more upwards events as defaults cannot occur in the training data. For industry-to-industry predictions, the annual alpha is 9.49% and significant with both buckets contributing equally well in terms of alpha. If one takes industry-to-industry predictions as extra factor in the factor model, industryto-company predictions remain significant. However, the reverse does not hold. Finally, I find that equal weighted industries have more explanatory power than value weighted industries.