Bangladesh Food Price Forecaster
P10 / P50 / P90 forecasts of monthly food prices for 110 Bangladeshi markets and 4 commodities (lentils, oil, rice, wheat flour) at horizons of 1, 3, 6 and 12 months.
Method
- Target is the cumulative log return
log(P_t+h / P_t), never the price level. Level targets inflate R² and break tree models on unseen periods, which cannot extrapolate a rising series. - LightGBM quantile regression over 29 features: lagged returns, rolling moments, national and regional co-movement, intra-month range, seasonality, trust score, coordinates.
- Split-conformal calibration, Mondrian-conditioned on commodity x recent volatility.
- Evaluated by rolling-origin backtest (expanding window, quarterly origins from 2015), not a single fixed split.
Honest performance note
Against a random-walk baseline, skill is regime-dependent: positive when prices are moving, negative in calm periods, with skill correlating around +0.4 to +0.6 with realized price movement. Point accuracy is not the selling point; calibrated intervals are.
Files
onnx/h{h}_q{q}.onnx— standard ONNXtrees/h{h}_q{q}.json.gz— compact trees for browser inference (bit-exact)manifest.json— feature order and conformal tables
Data: Andree, B. P. J. (2021). Monthly food price estimates by product and market. World Bank Microdata Library.