以集成學習建構混合模型預測台灣加權股價指數之趨勢

Abstract

本研究的目標在於如何準確地預測台灣加權股價指數在數日後是否上漲至超過預設門檻，蒐集並萃取台灣加權股價指數之技術指標、其他國際重要股市指數及台灣總體經濟指標三種面向資料作為特徵值，總共有192個特徵。藉由集成學習的概念提出一個混合模型，並以單純的隨機森林模型作為標竿進行比較。因蒐集之資料皆具有時間性，故使用增長式視窗滾動法(Increasing Window Rolling)以驗證模型績效表現。結果顯示，單純的隨機森林模型雖在短天期的預測準確率高，但易受門檻標準訂定的影響，使得樣本呈現分類失衡的現象；反之在長天期的預測準確率較低，但對於不同門檻值也較為穩定，同時AUC指標也呈現較佳的表現。雖然此研究提出的混合模型並無在模型準確率上有明顯優於單純的隨機森林模型，但也觀察到混合模型的預測若能避開國際金融動盪的時期，模型表現應能不錯。

The purpose of this study is emphasized how to accurately forecast the uptrend of Taiwan Capitalization Weighted Stock Index (TAIEX) in next few days, which is required to exceed different default thresholds. The data collections in three aspects comprise technical indicators of TAIEX, other influential stock markets in the world and Taiwan’s macroeconomic indicators as model inputs. After extracting the crucial information behind these variables, there are 192 features in total.\nBy proposing a blending model based on ensemble learning, the study will present a comparison with the simple random forest model. Besides, it is worth noting that raw data is temporal ordering; therefore, “Increasing Window Rolling” will be the validation method to evaluate the performance of models. The results have shown that the simple random forest model has high predictions in short periods but prone to be affected by different default thresholds, which may make sample imbalanced. On the contrary, predictions are less accurate in long periods but more stable under different default thresholds. In addition, the AUCs are also better. Although the proposed blending model is not significantly superior to the simple random forest model, it may still provide a good performance if phase of financial crisis is disregarded.