使用卷積視覺模型進行交易特徵的分類與物件偵測

Abstract

近年來，機器學習技術在金融交易中的應用越來越普遍，尤其是深度學習模型。本文旨在通過基於視覺的深度學習技術，幫助技術分析交易員更好地識別和客製化技術交易特徵。我們提出了一個端到端的解決方案，使交易員能夠獨立標記數據並訓練模型，從而解決了傳統基於規則的交易特徵編碼的低效率問題，並且消除了與第三方溝通的需要。在本研究中，我們採用了GAF時間序列編碼方法，並使用了基於視覺的卷積神經網絡（CNN）進行訓練。我們證實了GAF時間序列編碼非常適合用於金融交易數據，即使在每個類別僅有100個樣本的極小數據集下，我們的訓練方法仍然能夠維持出色的訓練效果。此外，我們還將實驗從靜態分類擴展到動態物體檢測，並使用了YOLO version 1模型進行修改，訓練結果遠高於用作比較基準的KNN-DTW模型。最後，我們採用了著名的道氏理論，展示了從標記和檢測到策略結合和儀表板的完整流程。總的來說，我們展示了即使在非常少量的數據和低模型參數下，我們的方法仍然能夠維持高精度和穩定性，非常適合金融交易行業。

In recent years, machine learning techniques have been increasingly used in financial trading, especially the deep learning model (1; 16; 39). This paper aims to aid technical analysis traders in better identifying and customizing candlestick patterns using vision-based deep learning techniques. We propose an end-to-end solution that enables traders to label data and train models independently, thus resolving the efficiency issues of traditional rule-based trading pattern coding and eliminating the need for communication with third parties. In this study, we adopted the Gramian Angular Field (GAF) (46) time-series encoding method and trained it with a vision-based convolutional neural network (CNN) (23). We show that the GAF encoding method is highly suitable for financial trading data, and even with a tiny amount of data with only 100 samples per class, our training method could still maintain outstanding training advantages. In addition, we extended the experiment from static classification to dynamic object detection by modifying the YOLO version 1 (33) model, which display a superior performance than the typical KNN-DTW (36; 47) model used as a comparison baseline. Finally, we used the famous Dow theory (9; 15) to show the complete process from labeling and detection to strategy combination. Overall, we demonstrated that our method can maintain high accuracy and stability even with very little data, which is highly suitable for the financial trading industry.