How accurate are the forecasts?
We re-ran every model 26 times on past data (14 Oct 2024 to 14 Sept 2026), each time using only what was known then, and compared its forecast with what actually happened.
These scores are measured on the synthetic demo series. They show that the pipeline works, not how well it predicts real mandi prices. They are recomputed automatically once real data is connected.
Our model: average error
3.9%
Best simple method (ARIMA)
4.3%
Direction right, 8 weeks ahead
74%
Range contains outcome
80%
target 80%
Average error by forecast horizon
Mean absolute percentage error (lower is better).
AdikeCast modelARIMA (used in earlier studies)Same as last yearPrice stays the same
One week ahead, the simple methods are as good or better: prices rarely move far in a week. The model earns its keep over 4 to 12 weeks, which is the range the sell decision depends on.
| Model | Avg. error | 1 week | 4 weeks | 8 weeks | 12 weeks |
|---|---|---|---|---|---|
| AdikeCast model | 3.92% | 2.41%62% direction | 4.18%59% direction | 4.47%74% direction | 4.54%83% direction |
| ARIMA (used in earlier studies) | 4.31% | 2.25%52% direction | 4.12%57% direction | 5.06%57% direction | 6.00%64% direction |
| Same as last year | 5.06% | 2.75%49% direction | 5.29%47% direction | 5.78%56% direction | 6.38%59% direction |
| Price stays the same | 4.44% | 2.27% | 4.21% | 5.19% | 6.24% |