Initial predictions in learning-to-forecast experiment

In this paper we estimate the distribution of the initial predictions of the Heemeijer et al. [5] Learning-to-Forecast experiment. By design, these initial predictions were uninformed. We show that in fact they have a non-continuous distribution and that they systematically under-evaluate the fundamental price. Our conclusions are based on Diks et al. [2] test which measures the proximity of two vector sets even if their underlying distributions are non-continuous.We show how this test can be used as a fitness for Genetic Algorithm optimization procedure. The resulting methodology allows for fitting non-continuous distribution into abundant empirical data and is designed for repeated experiments.