註釋 Many published studies of corporate failure prediction claim a high degree of accuracy, often over 90%, in predicting failure on the basis of only a small number of financial ratios. This study uses a uniquely large sample to determine how dramatically increased sample size, allowing better estimates of accuracy and more thorough out of sample validation, effects these results. Models such as Altman's Z score are found to perform poorly on the large sample. Significant improvements are possible through the introduction of new data. This study includes payment behaviour in several models, and this is shown to have a strong positive effect. Neural networks are relatively new in this area. Some comparative studies have been made, with conflicting results. This study looks in detail at their performance relative to accepted methods such as logistic regression. Neural networks are shown to have some powerful properties, but their use in failure prediction seems to offer no improvement over the conventional methods, at least using the methodologies tested here. Further research isjudged necessary. Finally, the study examines the form of the financial data used in traditional models. Constructing trend data is shown to be useful, and different forms of this are examined. The transformation of data is examined in some detail. Various transformations are discussed, including a new function, the hyperbolic tangent or tanh. Transformation of data is found to be very effective in improving a model. 