類神經網路應用於財務性時間序列之圖形辨識

Abstract

本論文研究財務性時間序列，將之視為圖形辨識問題，以類神經網路 作為辨識模型。以價格波動訊號之技術指標作為辨識模型使用之輸入特徵 。所採用的原始價格波動訊號需經過無相移濾波器與線性迴歸分析，以設 定辨識模型之目標輸出向量。透過辨識模型進行學習，以求得所需要的操 作訊號。藉由模擬結果，驗證我們系統的操作績效。實証後發現操作的績 效在每區段操作正確率與報酬率並不穩定，顯示價格波動的結構隨時間變 動。操作正確率不到50%，但報酬率達30%，顯示在操作正確的報酬率高過 操作失敗的損失。 This thesis is aimed at the pattern recognition of financial time series. We use a neural network for the model of pattern recognition. Before processing pattern recognition, we filter out the noise on the time series by wavelet transform to prepare the target vectors of recognition model and adopt the technical indexes as the features of the time series. By recognition model, we recognize the trading signal of time series. The performance of the system is demonstrated by a series of simulation. We find that the gain rate and the profit are not stable each period. It shows that the structure of time series varies with time. The gain rate is less than 50%, but the profit is over than 30%. It means that the gain in trading period is higher than the loss.