本文嘗試利用「理解神經網路系統」(reasoning neural networks),來學習臺灣發行量加權股價指數漲跌與相關經濟變數間之關係,並進行分析及預測。由於理解神經網路系統可決定適當的隱藏層節點數,並因達成完全學習的效果而不致受到局部極小值的困擾,使其在此類課題的研究上明顯優於常用的倒傳遞網路(back propagation networks)。然而,網路系統本身無法提供使用者預測訊息背後的意義,因此本文一方面使用敏感度分析,從系統外部來瞭解輸出入變數間的關係;另一方面則從系統內部著手,對系統的隱藏層輸出值進行模糊聚類分析(fuzzy cluster analysis),藉由對其內在表示的探討,來瞭解系統所學習到的內涵。結果顯示,臺灣股市以月為投資期間的市場效率無法被質疑,股價與總體經濟變數間之關係亦十分合理。
This study employs reasoning neural networks to learn the nonlinear regularities both in Taiwan's value-weighted Stock Index series itself and inrelation ships among the Index and relevant economic variable series. Forecasting and analyzing are then conducted. Due to the merits of being ableto determine appropriate number of nodes in hidden layers and not having local minimum problems, reasoning neural networks are superior to frequently used back propagation networks for the topic examined. Since the network system does not reveal information learned and utilized directly, sensitivity analysis and fuzzy cluster analysis are employed to uncover the contents and meanings of the information hold by the system. The results show that Taiwan's Stock Index variations could be reasonably explained by relevant macroeconomic variables. And no arbitrage profits could be made through monthly stock trading.