機器學習於財務異常報酬之應用

Abstract

本文主要利用財務實證上已經被證實有效的異常報酬因子(Anomalies)，試圖一次將多個因子同時納入機器學習模型(machine learning model)的考量中，並用此模型的預測建構一個本金中立的投資組合(dollar-neutral portfolio)，並且使用交易實務上會用到的衡量指標去評估策略的績效和穩定性，並且嘗試使用多種不同的參數組合和模型去做比較。

此實驗的發現主要有以下三點: 1.財務上已被實證過有效的異常報酬因子可以當作一種對於一間公司的歷史資料有效的特徵轉換(feature transform)，用這樣的方式去建構出來的策略在歷史回測中可以賺取巨額的超額獲利。 2.多因子模型建構出來的投資策略績效比單一因子更好並且更穩定。 3.提供異常報酬確實存在的現象可以給投資人、公司以及監管單位決策時做參考。

This paper mainly uses the Anomalies, which has been proved to be effective in financial evidence. Try to use the machine learning model with multiple factors at the same time and construct a Dollar-neutral portfolio by the prediction of the model. Also use the index used in trading practices to measure the performance and stability of the strategy, and try to compare the performance with many different combinations of parameters and model.

The main findings of this experiment are the following three points: 1.Anomalies that has been proved to be valid in financial can be an effective feature transform for a company's historical data. Construct the strategy in this way can earn huge excess profits in the history. 2.The model constructed with multi-factor is better and more stable than the model constructed with single factor. 3.The fact that anomalies is truly exists can be provided to investors, companies, and regulators to reference when they making the decision.