In this paper, we propose an alternative method to predict the Gonadosomatic Index (GSI), based on a technique known as computational neural networks (CNNs), with two main objectives: (1) to develop a quick and reliable method for the prediction of the fish reproductive period under variable environmental conditions, and thus (2) to reduce the field sampling and laboratory efforts. Three different neural architectures (5-6-6-1, 5-8-8-1 and 5-10-10-1), whose 'training' was carried out controlling three threshold determinism coefficients (Rt2: 0.7, 0.8 and 0.9), were trained to estimate the GSI of an introduced pumpkinseed (Lepomis gibbosus) population inhabiting a highly fluctuating Mediterranean stream in southern Spain. This GSI estimate was made using several easily measured fish and environmental variables. The correlation (R) between the GSI observed (GSIr) and the GSI predicted by the CNN (GSIe) was very high (>0.8 in all cases). The optimal CNN structure was the 5-6-6-1 with \( R^{2}_{t} = 0.8 \) because it produced the best generalization of the confidence limits of GSIe with respect to GSIr. To compare with traditional multiple regression analysis, we submitted the data to the same process as with CNN. The validation of the regression model produced a much lower correlation (R) than the CNN models. As an example of the predictive capacities of the CNN models, we predict the hypothetical pumpkinseed reproductive cycle of our population but under the environmental conditions found in the Camargue marshes (South France).