Leakage-free geospatial market panel and analysis code for market-level food-price surge prediction in Indonesia, 2007–2025

Data and analysis code supporting the article "When Earth-Observation Signals Fail to Transfer: Leakage-Free Spatiotemporal Validation of Market-Level Food-Price Prediction in Indonesia" (GIScience & Remote Sensing, under review).

The dataset is a monthly panel of 222 Indonesian retail markets across 34 provinces, January 2007 to December 2025 (228 months, 50,616 market-months, 133 variables). Each market-month pairs a composite retail food-price index over nine staple commodities with market-centred MODIS land-surface temperature, vegetation indices and evapotranspiration; CHIRPS precipitation; ERA5-Land temperature and soil water; NASA POWER; ESA WorldCover cropland share; travel-time accessibility; and global food and energy price indicators.

Every transformation is leakage-free: anomaly moments, imputation values, winsorisation limits, seasonal means and surge thresholds are estimated on training data only and applied forward, so the panel supports honest temporal and spatiotemporal validation. The period column carries the development (2007–2018), validation (2019–2021) and locked test (2022–2025) splits. Model inputs are the *_z_lf columns; the *_z columns are full-sample standardised anomalies retained for comparison only.

The accompanying R pipeline reproduces every reported result: the validation-design ladder (random rows, whole-month blocks, spatial blocks, leave-region-out, locked temporal and combined spatiotemporal), model-block ablation with distributed Earth-observation lags, paired whole-month bootstrap inference, ex-ante food-energy stress specifications, calibration selection and decision-curve analysis. Precomputed outputs are included so results can be inspected without re-running the pipeline.

Contents: data.zip (analysis-ready panel, data dictionary, preview file, source provenance and licences); code.zip (master pipeline, Step 01–09 scripts, precomputed results, publication figures, result registry); README.md; CITATION.cff; licence files; CHECKSUMS.sha256.

Licensing: the data and documentation are released under CC BY 4.0; the R source code is released under the MIT Licence. Upstream data providers retain their own terms — see SOURCES_AND_LICENCES.md inside data.zip, which includes the mandatory Copernicus attribution and disclaimer.