基因類神經網路在臺股指數期貨的預測與蝶式交易策略研究

An Empirical Application of Genetic Algorithm with Artificial Neural Networks to Forecast the Price of Taiwan Index Futures and the Simulation of Butterfly Trading Strategies

陳適宜(SHIH-YI CHEN)

指導教授 ： 古永嘉

共同指導教授 ： 徐純慧(CHUN-HUI HSU)

國立臺北大學/商學院/企業管理學系/碩士(2010年)

摘要

期貨交易具有相當的風險性，蝶式交易策略則為降低期貨交易風險的重要手段。以往學者大多利用迴歸分析或時間數列模型來進行股價的預測。本研究試圖整合基因演算及類神經網路模型，對不同到期月份台指期進行價格預測，並運用蝶式交易的原理來模擬投資損益。
本研究以2007年9月5日到2008年11月19日為區間，共計300筆日資料。依變數為近月到遠月期貨計五種期貨報酬率，自變數則採用成交量、未平倉契約口數、五日K值、十日K值、五日D值、十日D值、五日乖離率及十日乖離率等8種指標，各6期之落差期數，故共計54個變數。運用逐步迴歸進行初步變數篩選，再利用本研究自行開發之基因演算及類神經網路程式，使用SAS-IML程式語言，運用以上九種研究變數做為本研究之交易資訊歷史資料，進行類神經網路訓練。採用移動窗格法，即根據第1天至第30天的訓練結果來預測第31天的5個不同到期月份的台指期價格，以交易策略1而言，藉由買進1口漲幅最大的到期月份台指期及賣出1口漲幅最小的到期月份台指期，以此蝶式交易進行套利；其後根據第2天至第31天的訓練結果來預測第32天的5個不同到期月份的台指期價格，藉由買進1口漲幅最大的到期月份台指期及賣出1口漲幅最小的到期月份台指期，來進行套利。而由於每個交易日其漲幅最大的到期月份及漲幅最小的到期月份可能會不同，因此每個投資預測交易日的蝶式交易組合也就不同。本研究每日進行交易並當日沖銷，然後依此類推以進行第33天以後之預測及交易。
本研究自行開發之基因演算及類神經網路程式，在90天之訓練基期下其5個不同到期月份台指期的命中率介於50％~57％，而考慮交易成本後之年化報酬率在交易策略1為140.13％，在交易策略2則為75.82％，可見以基因演算及類神經網路來進行台指期報酬率的預測可獲得良好的效果。

關鍵字

基因演算法,移動窗格法,逐步迴歸,蝶式交易策略,類神經網路

並列摘要

Butterfly trading strategy is one of the most important method to reduce the risk of the considerably high-risk futures exchange. Regression analysis or time series model utilized to achieve this purpose by most researchers in the conventional way. In this thesis, we attempt to integrate genetic algorithm with artificial neural network to forecast the price of Taiwan index futures（TX） of various expiry months. Moreover, we apply the principle of butterfly trading strategy to simulate the investment gains and losses.
The period of study spans from 2007/09/05 to 2008/11/19 and the study includes 300 pieces of daily data. The dependent variables are rate of return of five different futures from nearby month to far month. There are 54 variables are composes of six lag lengths of nine indices, which include rate of return, trading volume, the volume of open interest, 5-day K index、10-day K index、5-day D index、10-day D index、5-day bias & 10-day bias. The preliminary variable selection is done by stepwise regression. The training of genetic algorithm and artificial neural network developed in this study is completed via SAS-IML programming language and the trading information collected from the nine indices listed above. With moving window method, we are able to forecast the rate of return of Taiwan index futures of five expiry months on the thirty-first day based on the training result from the first day to the thirtieth day. We then exploit butterfly trading strategy to conduct an arbitrage by buying in the Taiwan index futures with the highest rising percentage of its expiry month and selling out the one with the lowest rising percentage of its expiry month. Via the same method, we forecast the rate of return of Taiwan index futures of five expiry months on the thirty-second day based on the training result from the second day to the thirty-first day. We then exploit butterfly trading strategy to conduct an arbitrage by buying in the Taiwan index futures with the highest rising percentage of its expiry month and selling out the one with the lowest rising percentage of its expiry month. The combinations of butterfly trading strategy are different on every business day predicted to invest because the expiry months of the highest and lowest rising percentages are not identical on every business day. In this study, we trade every day and day trade on that day. The prediction and trade on and after thirty-third day can be deduced by the same analogy.
The accuracy rates are between 50%~57% when the base period is 90 days and the annualized return considered with transaction cost reaches 140.13% in trading strategy 1，and 75.82% in trading strategy 2. According to the data presented in this study, the effect of the genetic algorithm and artificial neural network is significant in the prediction for the rate of return of Taiwan index futures.

並列關鍵字

Butterfly trading, genetic algorithm, artificial neural network, stepwise regression, moving window method

參考文獻

Back, B, T. Laitinen, and K. Sere（1996）, Neural Network and Genetic Algorithms for Bankruptcy Predictions, Expert System Application, 11(4).

Bergerson, K. and D.C. Wunsch, （1991）, A Commodity Trading Model Based on a Neural Network-Expert System Hybrid, IJCNN-91, I, 289-293.

Holland, J.H. (1992), Adaptation in Natural and Artificial Systems, the MIT Press, Cambridge.

Kimoto, T., and K. Asakawa,（1991）, Stock Market Prediction System with Modular Networks. IEEE International Joint Conference on Neural Networks, 1, 1-6.

Lam M.（2004）, Neural Network Techniques for Financial Performance Prediction: Integrating Fundamental and Technical Analysis, Decision Support Systems, 567-581.