利用倒傳遞類神經網路探討波段與價格行為的相關性

Abstract

近年來巨量資料與機器學習的應用普及之後，宣告著資料分析、關聯與行為發現(Pattern Discovering)成為顯學的時代來臨，意味著機器學習的技術，將成為解決問題的基本技能，加上演算交易的成熟，使得先進國家對此方面重視度提高。 金融交易市場中，獲利之主因來自一個有正向報酬的交易，即進出間的價差。意即掌握相對高低點進出，不僅可使交易風險降低，亦可簡化模型分析的複雜度。因此本研究嘗試以技術分析，整理波段的趨式、動態、形態結合倒傳遞類神經網路學習、訓練並預測其漲跌幅，分析其中之物理現象。 文中嘗試將波段分成Not Well-formed及Well-formed兩種波的形態進行訓練。Not Well-formed在波段發展過程中只有波段臨界點而沒有均線交叉點，但Well-formed波態兩者都有。其中波段臨界點因為發生的時間點較均線交叉點早，因此可能因為時間價值而有較高的報酬；然後Well-formed中的均線交叉點其物理力量的資訊較完整，因此可能使得整體模型判斷力提升。因此設計交易策略X(參考完整波段中的兩時間點)、交易策略Y(僅參考Not Well-formed部分的資訊)，比較其績效表現。 研究結果顯示，波段臨界點與均線交叉點在準確率上，因為有倒傳遞類神經網路的學習，使波段臨界點的準確率提升，與一般認為較早交易的時間點，面臨反轉的機會較高的風險降低。因此，在實驗中此兩點的準確率已無明顯之差異。若在波段臨界點與均線交叉點在準確率上無明顯差異的前提下，觀察此兩點的年化報酬率、平均獲利點數及交易次數，發明波段臨界點的表現均較均線交叉點優。表示經倒傳遞類神經網路的學習，屏蔽早期不確定因素，可以使交易時間點提早，以享受波段帶來的獲利。

In financial markets, the profit gained from a positive reward of trading. Base on this concept, if the mode focus only on the spreads, not only can reduce the risk of the transaction (from spreads itself), can also simplify the complexity of the model. Therefore, this study attempts to gather the information by using the technical analysis, combining form back propagation neural network learning, and training and predict to find the worded-trade points. This paper attempts to classify the price wave into two categories- the Not Well-formed and the Well-formed. The Not Well-formed wave during its life cycle only has the wave critical point and has no moving average crossing point, but the Well-formed wave has both. Because of the wave critical point happened earlier than the moving average crossover point, we may think that the time value of the wave critical point will bring more profit than the moving average crossover point. On the other hands, the moving average crossing point in the Well-formed will have more complete information in trend, momentum, and pattern, thus improving the forecasting accuracy. Therefore, designing the two trading strategy -the strategy X (including two time points), and the strategy Y (including Not Well-formed parts only), and compare both strategies performance. The results show that the accuracy rate between the two points has no significant which breaks the time value loss hypothesis. This is because the mode has learned from the patterns and improved the accuracy rate on the moving average crossover points. Basing on this conclusion, checking on the annualized rate of return, the average gained point per trade, and the number of transactions between two trading strategies, the result shows that the wave critical point has better performance than the moving average crossing point, which means that the BPNN can learn and shield the uncertainties at earlier trade point, thus can gain more profit in earlier trade point safely.