Commodity factor investing via machine learning

We investigate the factor investing in Chinese commodities markets following two steps. The first step is to find profitable characteristics. We find that some technical characteristics can produce a comparable out-of-sample performance to the fundamental characteristics. The second step is to integrate various commodity characteristics to generate a composite signal. We apply the naïve equal-weighted model, three linear models and four tree-ensemble nonlinear models for style integration. The empirical results show that the four nonlinear machine learning integration models produce better out-of-sample performance than the linear models. Meanwhile, among the four tree-ensemble algorithms, the XGBoost algorithm performs best with control of the overfitting problem. •We investigate the factor investing in Chinese commodities markets.•Nonlinear integration is useful to improve the investment performance when combining characteristics.•The XGBoost algorithm performs best with control of the overfitting problem.