Estimation of Prediction Intervals in Neural Network-Based Regression Models

Currently there are various methods allowing the construction of predictive models based on data. Measuring prediction uncertainty plays an essential role in fields such as medicine, physics and biology where the information about prediction accuracy can be essential. In this context only a few approaches address the question of how much the predicted values can be trusted. Neural networks are popular models, but unlike the statistical models, they do not quantify the uncertainty involved in the prediction process. In this work we investigate several regression models with a focus on estimating prediction intervals that statistical and machine learning models can provide. The analysis is conducted for a case study aiming to predict the number of crayfish in Romanian rivers based on landscape and water quality information.