Milkfish price-forecasting benchmarks for five Indonesian regencies: reproducibility bundle (2024–2026)

This bundle is the reproducibility archive for the companion manuscript "Naive persistence as a hard benchmark for data-scarce aquaculture price forecasting: evidence from Indonesian milkfish." It lets a reader reproduce every table and figure in that paper. The study evaluates whether statistical and machine learning forecasters — a non-seasonal ARIMA, Random Forest, XGBoost, and an LSTM — improve on 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 the official SP2KP price series from January 2024 to May 2026. Naive persistence attains the lowest error at every site, an advantage confirmed by pairwise Diebold–Mariano tests. Rounding every competing forecast to the reporting grid, under four schemes, does not overturn the result, and a leave-one-site-out transfer test trained without access to the held-out test windows does not surpass persistence at any site. Contents: the weekly price dataset for the five regencies; two notebooks, one that trains the base models and one that runs the main analysis; frozen prediction artefacts that allow the base-model numbers to be verified exactly without retraining; the forecasts of the two models re-estimated for the analysis (the ARIMA and the leakage-free transfer model); and all derived tables and figures. The naive, Random Forest, XGBoost, and LSTM forecasts are read from the frozen artefacts; only the ARIMA and the transfer model are re-estimated, and both are deterministic. The recurrent network is not retrained in this version, so the published values remain stable. To reproduce: open the main analysis notebook in Google Colab or Jupyter, make the dataset and the two frozen artefacts available in the session, and run all cells top to bottom (about two minutes on CPU; no GPU required). This is a multi-site forecasting benchmark across five regencies; it is distinct from the separately deposited single-regency constructed price series for Pati, whose focus is the construction of one series rather than a model comparison. Data source: SP2KP weekly retail price series, Indonesian Ministry of Trade (publicly available).