Leakage-Free Spatiotemporal Count Forecasting: Benchmarking Sixteen Machine Learning Architectures with Conformal Uncertainty and Consensus Attribution

Forecasting zero-inflated count series over large, heterogeneous spatio-temporal panels is complicated by a constraint machine learning practice often overlooks: the most recent observations of the target variable are not yet available when a forecast must be provided. A model trained without enforcing this constraint learns from information that does not exist at inference time, and reports accuracy it cannot reproduce in deployment. This study presents a leakage-free forecasting framework and benchmarks nineteen trained model variants spanning statistical baselines, classical machine learning, gradient boosting and deep sequence models, of which sixteen are retained after three linear-family regressors are excluded for severe overfitting. A 44-feature matrix organized into six blocks is built under a programmatically asserted admissibility rule that accepts a feature only if it is measurable from information available at least one period before the forecast origin. Evaluation uses a real panel of 355,026 municipality-week observations from 474 Brazilian municipalities over fourteen years, partitioned strictly in time. Gradient boosting dominates, with a two-stage occurrence-magnitude model attaining a log-scale R² of 0.8426, and LightGBM attains 0.8416 at 23.9 ms per 1,000 inferences, 18.9 times faster than the strongest recurrent alternative. Split-conformal calibration yields intervals whose empirical coverage is 90.0% against a 90% target. Friedman and pairwise Wilcoxon testing with Cliff's delta finds no effect size larger than small, so model selection follows inference latency and uncertainty support rather than point accuracy. Relaxing the admissibility rule inflates apparent accuracy by 0.679, whereas a four-fold reduction in training history costs 0.001 and 10% input noise costs 0.692. Data quality, not model capacity, binds deployed performance.