利用隨機森林模型建構台灣指數期貨交易策略

Constructing a TAIEX Futures Trading Strategy Using Random Forest

鄭仁杰(Cheng Jen-Chieh)

指導教授 ： 江彌修

國立政治大學/商學院/金融學系/碩士(2018年)

https://doi.org/10.6814/THE.NCCU.MB.017.2018.F06

摘要

過去幾十年以來，預測金融商品價格走勢一直都是被熱烈討論的研究領域，但由於各種不同面向的因素交互影響，導致市場特性總是複雜且波動，價格走勢預測更加困難。普遍而言，錯誤率可被視為交易策略風險的指標，因此必須極小化錯誤率才能讓每單位風險享有更高獲利。為了使問題更單純，本研究將價格走勢預測視為分類問題，而本篇文章會使用機器學習（Machine Learning）來預測類別。在眾多演算法中，本研究選用多數決學習（Ensemble Learning）中具有許多良好特性的隨機森林（Random Forest）為本次交易策略建構的基礎架構。
　　本研究選用技術面與籌碼面指標作為訓練模型的特徵，建構兩個交易策略，而分析預測結果的方法除了隨機森林的袋外錯誤率（OOB Error Rate）以外，本研究會更著重在績效表現，以更接近交易策略的本質。由於台股期貨報酬不符合常態，本研究引入一種更為直覺的指標－卡馬比率（Calmar Ratio）作為評估績效的主要標準，另外再加入多種績效指標來提升績效評估的穩健性。
　　本研究透過不同切入角度測試策略績效與穩健性，結果也一再顯示，扣除手續費後兩個策略的績效確實遠遠勝過大盤，且擊敗大盤的情況不僅存在於訓練區間，更能延伸到測試區間。除此之外，本研究透過種種數據驗證測試區間屬於單一景氣循環的上漲區段，而處於這樣的背景間接降低了測試區間內兩個交易策略績效的鑒別度，使交易策略雖然勝過大盤績效，但差異性不大，而從袋外錯誤率的角度可以發現，交易策略確實具有足夠的穩健性。

關鍵字

機器學習 ； 多數決學習 ； 隨機森林 ； 交易策略 ； 台灣加權股價指數期貨 ； 袋外錯誤率 ； 卡馬比率

並列摘要

Over the past decades, predicting trends in financial products prices has been an area of interest, but due to the interaction effects of different factors from all sides, the nature of market is always complex and dynamic. In general, error rate is seen as a proxy of risk of trading strategy, and it needs to be minimized to improve strategy effectiveness. To simplify the problem, the forecasting problem in our research is treated as a classification problem, and Machine Learning is used to solve it. Because of some attractive characteristics, our research used one of Ensemble Learning, which is Random Forest, to construct trading strategies.
Our research selected technical and chip indicators as the features to train model, and the ways to analyze predictions contained OOB error rate, which derived from Random Forest, and the performance indicators. Because TAIEX Futures historical returns are non-normal distribution, our research introduced an intuitive performance indicator- Calmar Ratio as the evaluation criteria, and the other performance indicators have been added to improve the robustness.
Our research have tested the performance of strategies and the robustness from different angle, and the result shows that our strategies truly beat the benchmark in whole period, not just training period. Besides, there is a lot of evidence that testing period in our research was in recovery to the peak, and this will lower the discrimination between strategies and benchmark performance. However, from the point of view of OOB error rate, our strategies are truly sufficiently robust.

並列關鍵字

Machine learning ； Ensemble learning ； Random forest ； Trading strategy ； TAIEX futures ； OOB error rate ； Calmar ratio

參考文獻

1. Ballings, M., Van den Poel, D., Hespeels, N. and Gryp, R. (2015), “Evaluating Multiple Classifiers for Stock Price Direction Prediction,” Expert Systems with Applications, Vol. 42, No. 20, pp.7046-7056.

2. Blume, L., Easley, D. and O′Hara, M. (1994), ”Market Statistics and Technical Analysis: The Role of Volume,” Journal of Finance, Vol. 49, No. 1, pp.153-181.

3. Brieman, L. (2001), “Random Forests”, Machine Learning, Vol. 45, No. 1, pp.5-32.

4. Brieman, L., Friedman, J. H., Olshen, R. A. and Stone, C. J. (1984), “Classification and Regression Tree”, Wadsworth.

5. Brooks, C., A. Rew, and S. Ritson (2001), “A Trading Strategy Based on the Lead-Lag Relationship between the Spot Index and Futures Contract for the FTSE 100,” International Journal of Forecasting, Vol. 17, pp.31-44.