Genetic Design of Neural Networks on Stock Price Prediction

碩士論文[[abstract]]本研究探討使用遺傳演算法設計類神經網路在股價預測上的應用。遺傳演算法演化類神經網路的權重，類神經網路使用股票的技術指標作為輸入變數，而輸出則是決定該股票的買入或賣出。遺傳演算法通常使用二進位的染色體，本文使用浮點數的染色體，並用The SUnderland Genetic Algorithms Library（SUGAL）作為和類神經網路結合的軟體。浮點數的使用保有最短的界限長度（每個變數只用一個），如此可防止交換和突變對染色體的破壞。 遺傳演算法是增強式的學習方法，基於錯誤中學習，需要大量的計算，所以需要使用到模擬系統。本研究使用Windows 2000的環境，並用C語言來寫這個系統，我們使用摩根成份股來測試本系統，實驗結果驗證遺傳演算法能有效和穩定的設計類神經網路在股價預測上的應用。 關鍵詞：遺傳演算法、類神經網路、股價預測[[abstract]]This research discusses the use of genetic algorithms (GAs) to design neural networks for stock price prediction. Genetic algorithms evolve the weights for neural networks. The technical indicators of the stocks are used as neural network inputs to predict the stock should be buy or sold. The GAs usually use binary chromosomes for coding solutions. This research proposes using floating-point chromosomes to design neural networks using SUGAL, the SUnderland Genetic Algorithms Library. This coding provides schemata of the shortest defining length, one for each parameter, which prevents chromosomes from disruption by the processes of crossover and mutation. GAs are reinforcement learning techniques based on a trial-and-error, requiring heavy computation. A simulator is needed for the system. We use the C language to design the simulator, running on Windows 2000. The system is tested on MSCI Taiwan of stocks. This research reports experiments demonstrating that GAs are both effective and robust to design neural networks in stock price prediction problems