國立台灣科技大學 資訊工程系所
智慧型系統實驗室 研究論文
Intelligent System Laboratory Paper

94級畢業碩士 李大聖 發表論文

金融時間序列預測之調適性模型選擇

摘要

    金融時間序列是一種不具線性(non-linear)與平穩性質(non-stationary)且包含了大量雜訊的時間序列。當我們要正確的預測金融時間序列時必須同時考慮上述三個性質的影響，並且要選擇一個恰當的預測模型。因此，預測模型的選擇可說是正確預測金融時間序列的關鍵點之一。但是模型選擇一直以來存在著精確性與處理時效不可兼得的問題，尤其是在金融時間序列這類複雜度高的資料。因此，如何在一個及時(real-time)的金融預測模型選擇方法中兼顧精確性與處理時效是研究上的一大挑戰。
    根據艾略特於西元1938年所提出來的波浪理論，他將市場上的價格趨勢，歸納出幾個不斷反覆出現的型態。因此，過去的歷史資料分析能夠協助我們選擇預測模型的對應參數值。我們根據波浪理論提出了一個以重覆性特徵為基礎的模型選擇方法，並將此方法與移動視窗法結合。使得預測模型可視金融時間序列在不同時間的特性變化作最佳化的調適，這樣的方法也同時減少了非平穩性質對於預測模型的不良影響。最後，實驗證實了當決策支援系統搭配我們所建議的模型選擇與金融時間序列預測方法，能在真正的股票市場中獲利。

Adaptive switching on model selection of financial time series

Abstract

    Financial time series are non-linear, non-stationary data. They also contain a huge amount of noise. To obtain high accuracy prediction results, those properties must be concerned simultaneously; also a proper financial prediction model must be selected carefully. However, the model selection is a trade-off between the time complexity and the accuracy. Therefore, to balance the process time and the accuracy is a challenge for the read-time financial model selection.
    The cycle of stock fluctuated trends theoretically repeats themselves according to the wave theory (Nalph Nelson Eilliott, 1938). Based on this, we propose a repeat-pattern based financial model selection method. The model selection method can cooperate with the moving-window based training algorithm, and thus to accomplish an adaptive model selection method. The adaptive on model selection method is able to make a financial prediction model more stable. That results in our proposed method based decision support system can be applied to the real stock market efficiently and obtain good performance, as indicated from the experiment results.