利用機器學習預測下一期股價報酬－從基金投資組合來看

Abstract

績效好的基金經理人在選股時的想法為何？一直都是學術界與投資人好奇的問題。本文透過Classification Tree模型的學習方法，使用財務領域中四個面向之數據—包括財務報表科目（包含ROE杜邦分析）、股票估值、公司之盈餘管理相關因子與財務報表外之公司軟資訊（Soft Information），分析基金經理人使用公司各種面向資訊之程度與其預測能力。進一步探討不同投資風格之基金經理人，在選股時所參考的決策因子與路徑是否有所不同？本研究發現對於基金經理人來說，最為重要的選股決策因子為營運之現金流量。此外，基金經理人有更多的機會接觸到軟資訊，而實證結果Classification Tree模型中Jones模型相關變數也有在節點中，並且發現在市值較低的公司中，財務報表盈餘管理之情況，會影響基金經理人選股的判斷，進而影響公司的股票報酬率。更為有趣的是，Classification Tree模型除了可以避免掉變數共線性問題等的模型限制，並且相較於迴歸模型可以得知基金經理人選股是有前後重要順序的，透過Classification Tree節點中專案討論，確定此模型結果與實務上是一致的，不同投資風格的基金經理人—包括投資風格較廣、投資於特定產業或投資於特定公司規模—在選股時所參考之決策因子與路徑相當不同。最後，在資料經過GLM、Classification Tree或SVM之訓練後，模型預測財務報表公布後之個股報酬率方向可達到六成準確率，而將公司資料堆疊訓練之強韌性分析，預測準確率更可達六成五。

What is the decision-making by high-performing mutual fund managers when they select stocks? It has always been a question that investors are curious about. The purpose of this study is to analyze the key factors and the decision-making paths of top-performing fund managers by utilizing the learning method of Classification Tree model. In particular, this study uses the holding portfolio data of top-performing fund managers, and collect the stock characteristics of top 5% holdings selected by these fund managers. Four attributes of stock characteristics are considered: (1) financial performance collected from a firm’s financial statements (including ROE and elements in DuPont analysis), (2) stock evaluation information collected from a firm’s stock performance, (3) information associated with a firm’s earnings management, and (4) soft information, which is related to a firm’s private information. Since there are many fund styles, this study further explores whether fund managers with different investment styles have different decision factors (tree nodes) and tree paths when they select stocks. There are several findings. First, we find that for top-performing fund managers, the most important factor of selecting stocks is a firm’s operating cash flow. Second, we document that for companies with low market capitalization, earnings management variable (from Jones’s model) is selected as the tree node in the Classification Tree model, indicating that a firm’s earnings management behavior will become one of the key factors when fund managers selecting stocks for small cap funds and hence affect the company's stock return. Third, one advantage of our results is that the order of key stock-selection factors by top-performing fund manager can be analyzed in the tree paths and tree nodes. We further collect firm information from newspapers and document that for a particular firm, the key factors selected in tree nodes are consistent with the factors discussed in the newspapers for the same firm. In addition, we find that the key factors (tree nodes) and tree paths (order of tree nodes) are quite different for fund managers with different investment styles (i.e., broad market funds, sector funds, or funds with different market caps). Finally, when the data is trained by GLM, Classification Tree, or SVM, the accuracy rate of model predicting the trend of stock returns after the financial statements are published can reach 60%. For robustness analysis, such as stacking company data for training, the prediction accuracy rate are as high as almost 65%.