摘要：Generation error is one of the shortcomings ofthe classical mean-variance portfolio selection model whichwould result in unstable performance in out-of-sample datasets. Machine learning provides several potential solutionsfor reducing the generalization error, such as adding penaltyterms and random sampling. Due to some significant advantages of machine learning algorithms, we implement BlackLitterman (BL for short) series portfolio models integratedwith quantitative opinions generating from machine learningalgorithms in this paper. And, the fundamental factors in theFama-French model form the basis of quantitative opinions.Considering the non-linearity among fundamental factors, weconstruct high-order terms and cross terms of basic featuresby the approach of dimension-increasing transformation. Amulti-period portfolio strategy is designed in our work for thetimeliness of quantitative opinions, the experimental resultsreveal that the optimal BL model with investor opiniongenerating from Random Forest gained over 20% averageannual return and a Sharpe Ratio of 1.25. By comparison,S&P 500 index gained about 14.94% annual return and aSharpe Ratio of 1.00, the 1/N strategy gained about 15.92%annual return and the Sharpe Ratio of 0.99. Moreover, theBL series models are more diversified and robust, about 30%to 60% of assets are selected to construct the portfolio. Evenwhen the transaction cost is taken into account, our proposedmodels still obtained higher cumulative returns than S&P 500if the transaction cost is lower than 30%%.
