以機器學習模型預測臺灣上市櫃公司月營收公告之訊息

Abstract

依據《證券交易法》，臺灣發行有價證券之公司應於每月十日以前 公布上月之營收，此規範使財務報表使用者得透過月營收資訊較頻繁 地了解公司的營運情形，亦即強化會計資訊品質特性中的時效性。本 論文以過去幾月營收資訊，進行預測當期月營收是否為好消息或壞消 息之二元分類，研究樣本為台灣上市櫃公司之合併月營收資訊，研究 期間為 2015 年 2 月至 2020 年 1 月。 本論文欲挑選具備最佳預測月營收好壞消息能力的機器學習模型， 在同時考慮資料範圍大小、採用之演算法，以及資料前處理方法的多 種情況下，探討此三種有關模型設定的因素如何影響模型的預測成 效。此外，本論文亦設計兩種增設虛擬變數的方法，以討論本論文所 提及之相關月份對預測月營收好壞消息之影響。兩類虛擬變數分別 為有關陽曆月份的標示和農曆新年期間的標示。本論文實證結果顯 示，將包含較多產業類別的大資料集，進行前處理後，以梯度提升樹 (Gradient Boosting Decision Tree)訓練，其模型預測準確率顯著較高。 此外，若在資料集中加入標示陽曆月份的虛擬變數或標示是否為農曆 新年期間的虛擬變數，皆能提高模型預測營收好壞消息的能力，驗證 在預測月營收資訊時月份標示之重要性。

Public announcement of monthly revenues is required for issuers in Tai- wan pursuant to the Securities and Exchange Act, which reports the operating status of companies for users of financial statements on a frequent basis. That is, announcements of monthly revenues enhance the timeliness of accounting information. In this study, I predict the signals of monthly revenues—to fore- see if it is good or bad news—based on the historical revenues. Samples of this study are consolidated monthly revenues from listed companies in Tai- wan; the study period is from February 2015 to January 2022. This study aims to find out the best model setting to predict signals of revenues, considering three key factors: the scope of data, the algorithms, and preprocessing techniques. Furthermore, I design two dummy variable generating methods to specialize the importance of calendar months and Chi- nese New Year when making predictions. The findings conclude that gradient boosting tree-based models fitted by a larger preprocessed dataset achieve a significantly higher accuracy. Besides, a dataset with dummy variables in- dicating the calendar months or Chinese New Year reaches the greater pre- dictability of signals of monthly revenues.