Laws 紋理增強 GAF-CNN 之股市趨勢預測

Abstract

本論文針對金融時間序列預測中模型訓練提出一整合框架。透過格拉姆角場（Gramian angular field, GAF）將股價序列影像化，並結合 Laws 紋理能量測量（Laws' texture energy measures, Laws' TEM）提取影像局部特徵。實驗結果顯示，在 6 至 16 天之不同視窗長度測試下，本論文之 Laws-CNN 模型於標普 500 指數與台灣加權指數等標的中，在中位數準確率與模型穩定性上優於基準模型。

This thesis proposes an integrated framework for model training in financial time series prediction. By transforming stock price sequences into images via Gramian angular field (GAF) and incorporating Laws' texture energy measures (Laws' TEM) to extract local image features, the proposed framework effectively enhances model robustness. Experimental results indicate that under window sizes ranging from 6 to 16 days, the proposed Laws-CNN model outperforms the baseline models in terms of median accuracy and model stability across representative benchmarks, including the S\&P 500 and the Taiwan Capitalization Weighted Stock Index (TWII).