運用類神經網路與田口法預測台灣ETF指數

Using Neural Network and Taguchi Method to Predict Taiwan ETF-50 Stock Index Price

陳俊諺(Chin-Yen Chen)

指導教授 ： 李維平

中原大學/商學院/資訊管理研究所/碩士(2018年)

https://doi.org/10.6840/cycu201800631

摘要

台灣目前的經濟成長大不如從前，除了正確的管理財務之外，適當的投資也成為理財的重要環節之一。如今科技的發展迅速，人工智慧和大數據的崛起，也漸漸影響到了金融領域，預測股價已成為這方面熱門的探討話題。
以往研究上的實驗設計通常分為試誤法、一次一因子實驗法、全因子實驗法等，其中試誤法較常被使用於實驗設計裡，缺點在於其並非一種有系統性的方法，且過於依賴個人經驗，當實驗因子變多時，實驗的次數也將以倍增的方式成長，耗時又耗力，因此本研究透過使用田口式直交表實驗法作為本研究之實驗設計方法。藉由部分因子實驗，讓實驗次數大幅降低，以節省時間與成本，且可有效找出最佳參數組合。
實驗結果表示，利用田口法取代試誤法，透過直交表所規劃的9組實驗，省去了更多試誤法所不必要的實驗，讓實驗更有效率。另外在篩選方法上，則以皮爾森相關係數更優於其它的變數篩選方式，最終模型誤差=0.15，漲跌方向準確率=68%，成功降低以往模型的誤差，並提高模型的穩定性。

關鍵字

倒傳遞類神經網路 ； 股票股市 ； 皮爾森相關係數 ； 灰關聯 ； 逐步迴歸 ； 田口法

並列摘要

Taiwan's current economic growth is not as good as before. In addition to proper financial management making right investments has also become very important. Nowadays, with rapid development in science and technology, and the rise of artificial intelligence and big data have gradually affected the financial sector. Prediction stock prices has become a hot topic in this regard.
The experimental designs in past researches were usually trial-and-error, one-factor-at-a-time, and full-factor method. The trial-and-error method was more oftenly used in experiments over others. However, the disadvantage was that it wasn't a systematic method, and too much reliance on personal experience. When the factors became more, the number of experiments would also grow exponentially, which was time consuming and laborious. Therefore, this study used the Taguchi orthogonal test method as the experimental design method of this study. By using partial-factor experiments, the number of experiments was greatly reduced, thus saving time and budget, and effectively finding the best combination of parameters.
The results showed that by replacing the trial and error method with the Taguchi method, only nine sets of experiments were planned by the orthogonal table, eliminating unnecessary experiments from trial and error, making the entire experiment more efficient. In addition, in the screening method, the Pearson correlation coefficient was better than other variable screening methods. The final model error was 0.15, and the direction accuracy was 68%. The error of past models were successfully reduced and the stability of the model had improved.

並列關鍵字

BPN ； Stock Market ； Pearson Correlation ； Grey Relational Analysis ； Multiple Regression ； Taguchi Method

參考文獻

[1] De Villiers, J., & Barnard, E. (1993). Backpropagation neural nets with one and two hidden layers. IEEE Transactions on Neural Networks, 4(1), 136-141.

[2] Zhang, Y., & Wu, L. (2009). Stock market prediction of S&P 500 via combination of improved BCO approach and BP neural network. Expert systems with applications, 36(5), 8849-8854.

[3] Hsieh, L. F., Hsieh, S. C., & Tai, P. H. (2011). Enhanced stock price variation prediction via DOE and BPNN-based optimization. Expert Systems with Applications, 38(11), 14178-14184.

[4] Amin Hedayati Moghaddam, Moein Hedayati Moghaddam, Morteza Esfandyari, Stock market index prediction using artificial neural network, Journal of Economics, Finance and Administrative Science, Volume 21, Issue 41, 2016, Pages 89-93, ISSN 2077-1886。

[5] Aparna Nayak, M. M. Manohara Pai, Radhika M. Pai, Prediction Models for Indian Stock Market, Procedia Computer Science, Volume 89, 2016, Pages 441-449, ISSN 1877-0509。