使用 10-K 報告中的公司關係建構時序圖神經網路以預測股票回報率

Abstract

股市預測一直以來都是熱門的研究議題，準確的預測能夠幫助投資者作出投資決策，進而最大化獲利。這類的研究過去多以時間序列模型將股票預測視為回歸問題（預測股價），或是將其視作分類問題（預測股價漲跌），並把每支股票當作是一個獨立的時間序列處理，沒有考慮到股票（公司）與股票（公司）之間會互相影響。除了時間序列相關的深度學習，讓模型考慮自身的時序相關外，隨著圖神經網路近期的發展，我們可以透過加上股票之間的連結，讓模型也考慮股票與股票間的互相影響：我們將股票視為圖上的點，過去的股價相關資料和財務數據當作點的特徵，透過10-K 報告的內容形成圖網路的連結，結合長短期記憶、圖注意力網路和排序學習，試圖去預測隔日股票投報率的排名，進而形成最大化獲利的投資組合。相較於現有使用靜態圖神經網路預測股市的研究，我們使用的是 spatial-based 的圖神經網路，而不是基於 spectral-based 的神經網路。此一設計可以讓我們處理動態圖問題，搭配每年發佈的 10-K 報告，我們可以及時地更新圖上的連結，進而做出更反映現況的預測。 在衡量排序的指標和累計投資回報率上，我們的模型相較於各種基本策略、基準模型和大盤指數在多數測試年中勝出。我們的模型在 2018、2019跟2020 年的累計投資回報率分別取得17.20%、26.01% 和105.26% 的優異表現，而S&P 500 市場指數的累計投資回報率則分別是 -5.82％、26.16% 和 21.05%。

Stock market prediction is a widely researched topic. Precise prediction can assist investors in making better investment decisions and maximizing profits. In the past, most studies in this area have used time series models for stock prediction, either as a regression problem (predicting prices) or a classification problem (predicting price increases or decreases). However, there has been a tendency to overlook the interactions between stocks. In addition to the time series-related deep learning, which allows the model to consider its temporal correlation. With the recent development of graph neural networks, we can add connections between stocks so that the model can also consider the interactions between stocks: we treat stocks as nodes in a graph, utilizing historical price-related data and financial figures as node features, and we establish connections in the graph network through the content of 10-K reports. We aim to predict the ranking of next-day stock returns and subsequently construct an investment portfolio for maximizing profits. To achieve this, we integrate Long Short-Term Memory (LSTM), Graph Attention Networks (GATs), and learning to rank. In contrast to existing studies that employ static graph neural networks for stock market prediction, our approach employs spatial-based graph neural networks instead of spectral-based ones. This design enables us to address dynamic graph problems, and in combination with annually released 10-K reports, we can promptly update the graph's connections, resulting in more updated and reflective predictions. Regarding ranking metrics and cumulative investment return ratios, our model outperforms various baseline strategies, benchmark models, and market indices in most testing years. Our model achieved outstanding cumulative investment return ratios of 17.20%, 26.01%, and 105.26% in 2018, 2019, and 2020, respectively, while the S&P 500 market index had cumulative investment return ratios of -5.82%, 26.16%, and 21.05% for the same years.