Naive persistence as a hard benchmark: A multi-site forecasting reality check on data-scarce retail price series

Reproducibility bundle for the manuscript "Naive persistence as a hard benchmark for milkfish retail price forecasting in data-scarce Indonesian markets" The study evaluates whether sophisticated statistical and machine learning forecasters (SARIMA, Random Forest, XGBoost, and LSTM) improve upon a naive persistence baseline for one-step-ahead weekly retail milkfish (Chanos chanos) prices across five Indonesian regencies (Pati, Lamongan, Gresik, Maros, and Pinrang), using official SP2KP price series. Under a single one-step-ahead protocol, naive persistence attains the lowest RMSE and the lowest mean absolute scaled error at every site, an advantage confirmed by pairwise Diebold-Mariano tests and unchanged by a leave-one-site-out transfer test. The bundle reproduces every table and figure in the paper. Contents: the weekly price dataset; the complete analysis notebook; frozen prediction artifacts that allow exact verification of the published numbers without retraining; and all derived tables and figures. To reproduce: open the notebook in Google Colab, make the dataset available in the session, and run all cells top to bottom. Because the LSTM is nondeterministic across random seeds (largest variance at Maros), re-running may shift the reported LSTM RMSE slightly; the frozen artifacts hold the canonical run. Data source: SP2KP weekly retail price series, Indonesian Ministry of Trade (publicly available).