Answers are grounded in this paper and cite their sources.
Evaluates forecast accuracy of neural networks in Poland, indicating design parameters matter.
The aim of this study was to evaluate the impact of the design and training parameters of MLP neural networks on the quality of forecasts of the number of traffic accidents in Poland. The study utilized historical data on traffic accidents, including the number of fatalities, injuries, and collisions. The MLP models were analyzed in various architectural configurations (number of neurons in the hidden layer), with different activation functions, and with different proportions of the dataset allocated to the training, test, and validation sets. MAPE, RMSE, and the Theil coefficient were used as measures of forecast quality. The analysis of variance (ANOVA) revealed no statistically significant effect of the training sample size or network architecture on prediction accuracy, while the activation function showed an effect at the threshold level of statistical significance. However, Tukey’s post hoc test did not reveal any significant differences between the individual activation functions, suggesting their limited, context-dependent impact on model quality. The Pareto analysis indicated the existence of a set of non-dominated solutions, confirming the multi-criteria nature of the problem of selecting a predictive model. The results indicate that MLP models enable moderately high forecast accuracy; however, their effectiveness depends more on the trade-off between multiple quality criteria than on individual design parameters. The effect of diminishing returns as the complexity of the neural network increases was also noted.
No takes yet. Share an insight, caveat, or question.
Piotr Gorzelańczyk (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: