Forecastability of Intentional Homicide Patterns in Ecuador (2014–2026): A Rolling-Origin Evaluation of Statistical Forecasting Algorithms

Ecuador has experienced one of the fastest escalations of lethal violence in Latin America, with annual intentional homicides rising from about 1000 in 2014–2019 to 9283 in 2025. This study evaluates how well such patterns can be forecast, and what governs that forecastability, using the complete official open dataset of 43,976 victim-level records covering January 2014–June 2026 (portal update 15 July 2026). Incidents are aggregated into four monthly pattern series—case volume, firearm share, mean victim age and geographic concentration in Guayas—and fifteen specifications—classical, automatically configured, and state-of-the-art neural and gradient-boosted baselines—are compared under a rolling-origin protocol with 31 evaluation origins and a 12-month horizon, with every configuration decision taken inside each training window. Accuracy is reported as the mean and standard deviation of MAE, RMSE, MAPE, MASE and R2 across origins, interval quality as empirical coverage and the mean scaled interval score, and differences are tested with Diebold–Mariano statistics under Holm correction. No pairwise comparison among the 303 tested survives multiplicity control, and in all four series, the best and second-best specifications are statistically indistinguishable: algorithm choice is not what these data can resolve, and the result holds with current deep-learning architectures in the candidate set rather than in their absence. The test is conservative and its power is limited, so this is non-detectability at this sample size rather than evidence that the differences are zero. What does determine it are diagnosable properties of the series. Bounded, low-variability series are forecast with mean MASE between 0.58 and 1.13 across the full candidate set, and below 1.00 by twelve of the fifteen specifications on the firearm share, whereas unbounded count series under regime change exceed an MASE of 2.09 for every specification tested, including simple benchmarks. No series requires a seasonal difference—the OCSB test returns D=0 in all four and the automatic search never selects one—although weak seasonal autoregressive terms are retained at some origins, and the firearm share shows a small but significant monthly effect. A genuine external validation against the observed 18 months of 2025–2026, unseen by any model, reproduces the same asymmetry (firearm share, MAPE 2.3% with 100% interval coverage; volume, bias −14.0%). Splitting the evaluation by regime shows that the crisis made count series roughly twice as unpredictable and bounded series roughly twice as predictable: the escalation destabilised the level of the phenomenon while stabilising its composition.