This paper examines relationships between traditional summary statistics used for an economic forecast evaluation and profitability of investment decisions based on that forecast.The importance of these relationships is particularly evident for artificial neural network (ANN) supervised learning,where the process of network training is based on a chosen error function.Considering such measures as common errors,accuracy (directional),correlation (the desired and ANN output),improvement over ‘efficient prediction (e.g.last known value),their relationships with stock trading strategies profitability were found to be of complicated and non-conclusive nature.Only the degree of improvement over ‘efficient prediction shows some robust links with returns measures.