Seasonal time-series evaluation · Sri Lanka

Rice production forecast
backtest

Expanding-window evaluation of four target-history models. The target is seasonal rice production in thousand metric tonnes (not yield). Forecast inputs use observed production history and the future season label; realized test-period weather, acreage and macroeconomic values are excluded.

149seasonal observations · Yala 1950–Yala 2024

33forecast origins per horizon

1–6season-ahead horizons

4models compared on common origins

MAE by forecast horizon

SARIMAXSARIMAX + RF residual

Horizon-specific scores

Lower MAE, RMSE, WAPE and MASE are better. Scores aggregate overlapping rolling forecasts and are descriptive, not independent confidence estimates.

Model	Horizon	Origins	MAE	RMSE	WAPE	MASE
Direct RF (lags)	1	33	246.4	321.4	19.41%	1.744
SARIMAX	1	33	252.3	385.0	19.87%	1.738
SARIMAX + RF residual	1	33	245.4	352.8	19.33%	1.713
Seasonal naive	1	33	303.4	413.3	23.90%	2.111
Direct RF (lags)	2	33	390.0	477.4	18.72%	2.727
SARIMAX	2	33	333.7	507.0	16.01%	2.260
SARIMAX + RF residual	2	33	328.8	449.8	15.78%	2.331
Seasonal naive	2	33	335.8	478.2	16.11%	2.317
Direct RF (lags)	3	33	234.0	313.1	17.87%	1.646
SARIMAX	3	33	254.0	331.8	19.39%	1.793
SARIMAX + RF residual	3	33	253.7	323.6	19.37%	1.839
Seasonal naive	3	33	264.5	350.0	20.19%	1.891
Direct RF (lags)	4	33	374.9	488.0	17.61%	2.673
SARIMAX	4	33	402.7	558.0	18.91%	2.756
SARIMAX + RF residual	4	33	370.1	504.4	17.38%	2.580
Seasonal naive	4	33	376.7	552.3	17.69%	2.603
Direct RF (lags)	5	33	252.5	349.9	18.84%	1.810
SARIMAX	5	33	285.6	370.9	21.31%	2.022
SARIMAX + RF residual	5	33	269.0	347.3	20.08%	1.951
Seasonal naive	5	33	260.2	355.9	19.42%	1.893
Direct RF (lags)	6	33	373.8	478.8	17.24%	2.764
SARIMAX	6	33	366.5	521.7	16.90%	2.474
SARIMAX + RF residual	6	33	345.5	478.4	15.93%	2.414
Seasonal naive	6	33	326.8	474.4	15.07%	2.276

Interpretation and limits. This is a historical rolling-origin benchmark, not a guarantee of future performance. The available sample is small; the last historical years also participate in the rolling evaluation, so retain a final untouched prospective evaluation for any formal model-selection claim. The hybrid has no published prediction interval in this release. Climate-shock attribution is not established by these aggregate scores.

Generated from outputs/backtest_summary.csv. Rebuild with python scripts/build_dashboard.py. See the methodology and data dictionary.

Rice production forecastbacktest

MAE by forecast horizon

Horizon-specific scores

Rice production forecast
backtest