Technical report
Detailed notes, interpretation, and operational boundaries
This section documents what the published evidence means, how it was produced, and where the interface deliberately avoids making a stronger claim than the data supports.
How the production method is selected
Candidate methods are compared on chronological folds. A later time holdout is reported separately and is not an input to the automated selector.
- Validated recent-level + lead-aware blend is the current production point method. The best lead-aware candidate improved pooled selection MAE by 2.4%; the gate requires 1%.
- The candidate won 3 of 3 usable selection folds; the configured stability gate requires 66.6%.
- Lag, rolling, calendar, market, crop, freshness, arrival-coverage, and volatility features are computed only from information available at each forecast origin.
How the seven-day point path is produced
The production blend keeps 65% of the stable five-report moving average, adds 35% of a lead-aware tree estimate, and applies a 7.5% damped recent-level adjustment per lead day.
- The automated search selected the tree and level-drift weights from 90 parameter pairs using only chronological selection folds; holdout values are computed afterward.
- Lead day, target weekday, target week, target month, recent momentum, volatility, arrivals, and market context can now change the central estimate from one target date to the next.
- No cosmetic noise is added. A few genuinely stable series may still round to the same rupee on adjacent days, while their uncertainty ranges remain separately calibrated by horizon.
Locked holdout performance
The final holdout runs from 2026-07-17 to 2026-09-14 and contains 1,96,317 forecast examples.
- Mean absolute error is ₹532 per quintal and WAPE is 10.2%.
- Directional accuracy is 47.4%; this is contextual evidence, not a guarantee for a particular market or date.
- Aggregate metrics can hide weak commodities or markets, so the published evaluation retains segmented horizon, commodity, state, market, and coverage results.
Uncertainty and safe interpretation
Intervals are asymmetric residual ranges calibrated on earlier out-of-time errors with hierarchical fallback when a narrow segment has too few examples.
- The nominal interval target is 80%; empirical locked-holdout coverage is 72.3%.
- Coverage below the nominal target is visible because hiding it would overstate reliability. Wider intervals should be read as less precision, not as a larger expected price move.
- Forecasts are secondary to observed reports and are information only, not trading, procurement, or financial advice.