Answers are grounded in this paper and cite their sources.
Comparative forecasting analysis demonstrates naive persistence outperforms complex machine learning models in data-scarce aquaculture retail markets, highlighting the impact of reporting granularity.
Machine and deep learning are increasingly used to forecast food commodity prices where data are scarce, yet these methods are rarely tested against the simplest benchmark: that next week’s price equals this week’s. We test this hypothesis on milkfish prices (Chanos chanos), a staple of Indonesian aquaculture, using weekly retail prices from five sites between January 2024 and May 2026. Against naive persistence we benchmark an ARIMA, Random Forest, XGBoost, and a ten-seed LSTM under a one-step-ahead protocol. Naive persistence attains the lowest error at every site, and Diebold–Mariano tests confirm its advantage over the machine learning models under absolute-error loss. The explanation most likely lies in the reporting system rather than the market: prices are published on a coarse rounded grid, so the reported price is unchanged in 80%–91% of weeks, and only a no-change forecast reproduces it exactly. This ceiling is hard to escape: rounding every rival forecast onto the same grid narrows the gap without closing it significantly, and pooling across sites does not surpass persistence. We argue that reporting granularity is a first-order feature of the forecasting problem, common in data-scarce settings, and offer a preliminary diagnostic for practitioners with comparable data.
No takes yet. Share an insight, caveat, or question.
Umam et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: