透過機器學習與深度學習預測月營收並形成套利投資組合

Abstract

本論文使用公司過去發布之月營收資訊，搭配兩種預測模式與兩種資料特徵前處理方法，並分別使用隨機森林模型、極限梯度提升模型、多層感知機模型、Transformer 模型，並結合迴歸器與分類器，用於預測下一個月的月營收。此外，本論文使用預測之月營收與基準相比作為預測消息好壞之基礎，建立不同的投資組合，實證套利之可能性。 經過模型和結果實證，即使在有新公報與疫情影響之下，本論文得到以下三個結論：第一，證明可將機器學習與深度學習應用於預測未來月營收，在迴歸器的部分，機器學習模型預測能力較深度學習模型佳，在分類器的部分則是極限梯度提升模營與多層感知機預測能力較佳。第二，預測型態與月營收前處理對於機器學習與深度學習模型預測能力皆有影響，其中又以月對月預測搭配原始月營收資料，預測偏誤較大，而經過去趨勢化之月營收資料，可以消除季節性因素對於模型預測能力之影響。最後，本論文發現在假設市場完美、沒有交易成本與摩擦之情況下，在迴歸器中，機器學習與深度學習搭配去趨勢化之月營收皆能產生相對較高且穩定之正向套利報酬，其中僅少數年份報酬為負，並且在四個模型中，以多層感知機套利結果最為穩定。在分類模型中，本論文發現極限梯度提升與多層感知機，可以得到預測準確度高且套利報酬穩定之結果。

This paper uses historical monthly revenue information released by companies, combined with two forecasting patterns and two feature preprocessing methods, and employs four machine learning models. It integrates regressors and classifiers to predict the next month's revenue. Additionally, this paper uses the predicted monthly revenue compared to a benchmark as the basis for determining the message of monthly revenue released is positive or negative and establishes different investment portfolios to empirically test the possibility of arbitrage. Based on the models and empirical results, despite the impact of new International Accounting Standard and COVID-19, this paper draws the following three conclusions: First, it is demonstrated that machine learning and deep learning can be applied to forecast future monthly revenue. In terms of regressors, machine learning models have better predictive ability than deep learning models, while for classifiers, eXtreme Gradient Boosting and Multilayer Perceptron show better predictive ability. Second, forecasting patterns and monthly revenue preprocessing impact the predictive ability of both machine learning and deep learning models. Specifically, month-to-month forecasting pattern with raw monthly revenue data leads to larger prediction biases, whereas detrended monthly revenue data can eliminate the influence of seasonal factors on the model's predictive ability. Finally, this paper finds that under the assumption of a perfect market with no transaction costs or frictions, in the case of regressors, both machine learning and deep learning models combined with detrended monthly revenue data can generate relatively high and stable positive arbitrage returns, with only a few years showing negative returns. Among the four models, the Multilayer Perceptron yields the most stable arbitrage results. In the case of classifiers, this paper finds that Extreme Gradient Boosting and Multilayer Perceptron achieve high prediction accuracy and stable arbitrage returns.