In this paper, an empirical comparison of four micro differential evolution variants to solve partial instances of the Full Model Selection (FMS) problem on time-series databases is presented. Two smoothing methods, three time-series representations, and one classifier, all of them with their corresponding parameters, are optimized by the evolutionary search. The study focuses on the effects of the crossover operator and the type of base vector used in the differential mutation. The four variants are tested in three experiments to analyze the final statistical results, the convergence behavior and the type of optimized model obtained. Six temporal databases with different features are solved. The results, which are statistically validated by using non-parametric tests, suggest that the four variants are able to provide competitive results. However, using the best vector in the current population, coupled with a binomial crossover, allow the micro differential evolution to reach better final results, while the opposite (base vector chosen at random and exponential crossover) help to find a good model with less solution evaluations in temporal databases whose models are very expensive to evaluate.