本研究嘗試利用訊號分析法 (Signal approach) 篩選重要變數後，比較Panel Logit 之固定效果與隨機效果方法與類神經模糊工具在銀行危機上之預測能力。實證結果顯示，雖然Panel Logit之固定效果與隨機效果模型皆驗證了過去理論上各變數對銀行危機之影響方向，但整體而言，其預測準確度並不比類神經模糊好；而類神經模糊除了提供更準確的樣本外預測之外，其所建構的規則庫與3D圖提供較以往更為細膩的變數關係，此關係可作為未來理論發展之基礎；整體而言銀行危機可以被有效的預測，然而單一國家的預測結果顯示不同的國家適用不同的預測工具，沒有一項工具能夠適用於所有的國家；而資料集屬性與預測工具之對應關係可作為未來之研究方向。本研究除了提供建構預警模型的另一可能途徑之外，實證結果 (沒有一項工具適用於所有的國家) 提供了異於過去的結論，可作為未來研究之參考。

This research attempts to filter the important variables through the signal approach first, then compares the prediction accuracy among the panel logit with fixed effects, panel logit with random effects, and neuro fuzzy models on the banking crises prediction. Although the panel logit models verify the effects of the explanatory variables consistently with the theory, the prediction accuracy is not as good as that of neuro fuzzy. In addition to the better out-sample prediction accuracy, neuro fuzzy also provides a more detailed and complicated relationship among the variables through the rule base and 3-dimension graphics, which could be used for further theory development or modification. Overall, the banking crises can be predictable; however, each country needs different forecasting model. The mapping between the forecasting model and data set characteristics can be further developed for the future research. This research has provided an alternative model to the banking crises warning system. Besides, the empirical results denote the contingency characteristics of the forecasting models, which can be further explored for future research.