Data-Built Fuzzy Rule Bases Versus Count Regression for the Prediction Component of Road Network Screening: A Monte Carlo Benchmark for Sustainable Road Safety Management

Road network screening decides which segments receive limited safety budgets, so its ranking quality bears on the road-safety targets of the 2030 Agenda for Sustainable Development. Fuzzy rule bases built from crash data are often proposed as readable screening aids but are usually tested on field data whose true risk is unknown. This study benchmarks a data-built, grid-based Mamdani rule base and its zero-order Takagi–Sugeno counterpart against Poisson, negative binomial and interaction regression, gradient boosting and an oracle regression. Six data-generating mechanisms, including a rule-structured and a smooth non-linear one, were crossed with three training sizes (250, 800 and 4000), with 200 replicates each, and every method was scored against the known mean. At 800 records, the Mamdani rule base ranked segments with a Spearman correlation of 0.33 to 0.51 under the log-linear mechanisms, against 0.78 to 0.93 for Poisson regression, and placed 25 to 36% of the riskiest decile in its own top decile, against 46 to 76%. Under the rule-structured surface, it drew level with Poisson regression only at 4000 records. A paired ablation traced most of the gap to noisy cell consequents rather than to the grid itself, and roughly one in five neighbouring rule pairs pointed against a known effect direction. In this prediction step, which excludes the empirical Bayes adjustment, count regression ranked most reliably; a small, cross-validated and audited rule base can serve as an interpretability aid.