The stock returns and ROE are meaningful to the shareholders to realize the level of investment feedbacks and companies’ profitability. The accurate forecasts for both factors thus can be very important to the investors. Instead of consulting to the financial experts, this study proposes an approach by decoding artificial neural networks (ANN) to automatically construct a rule base for performing forecasts. The ANN being implemented is the so-called back-propagation neural network. The algorithm known as TREPAN is introduced to uncover the hidden knowledge from ANN for building the relationship between company’s current financial indices and the probable performance in the next season. The study uses Taiwan stock market electronic companies in the time period from years 2000 to 2005 as a basis for carrying out experiments. The inputs for the ANN in this preliminary study are only concerned with the fundamental factors. It is expected that, through this empirical study, one may accelerate the rule base construction for the financial expert systems and to provide the more clear traces to improve the diagnosis to the companies. The results reveal that, using fundamental factors as inputs, the ANN can perform up to 70.68% accuracy in the experiments. In terms of TREPAN algorithm, the knowledge of companies’ financial performance can be successfully extracted from ANN, though the minor error may occur. Some interesting discoveries are also addressed.