Forecast scorecard as of 18 Nov
Last 30 days, to 18 Nov 2026
Average error (MAE)
p50 against the clearing price, every block
Skill vs naive
naive MAE ₹777/MWh
p10–p90 coverage
of blocks inside the band; 80% is calibrated
Days beating naive
lower error than the same block yesterday
Daily error, last 90 days
Worst 5 days in the last 30
| Delivery day | MAE | Naive MAE | In band | Why |
|---|---|---|---|---|
| 20 Oct 2026 | ₹2,129/MWh | ₹1,324/MWh | 6% | 55 of 96 blocks cleared at the ₹10,000/MWh price cap. |
| 22 Oct 2026 | ₹1,951/MWh | ₹776/MWh | 45% | 47 of 96 blocks cleared at the ₹10,000/MWh price cap. |
| 8 Nov 2026 | ₹1,687/MWh | ₹2,411/MWh | 42% | RAJIV GANDHI TPS unit 1 (Haryana) in forced outage, 600 MW, since 08 Nov 2026 14:20 IST: reserve shut down / standby unit |
| 21 Oct 2026 | ₹1,508/MWh | ₹1,798/MWh | 85% | 34 of 96 blocks cleared at the ₹10,000/MWh price cap. |
| 23 Oct 2026 | ₹1,185/MWh | ₹415/MWh | 100% | 43 of 96 blocks cleared at the ₹10,000/MWh price cap. |
How the scorecard is graded
What is graded. The forecast exactly as it was published, read from the vintage archive as it stood at the grading cutoff: 11:00 IST on the day before delivery, the DAM bid gate. For RTM that is the next-day forecast issued about 06:25 IST on the day before delivery, before the DAM auction. A later re-publish never improves a day's grade. The DAM model is a set of LightGBM gradient-boosted regressors for the median and the p10/p90 band, retrained weekly on the trailing 460 days of IEX prices. The RTM next-day forecast starts from the DAM forecast for the delivery day and adds an RTM–DAM spread drawn from the past year of blocks at the same time of day.
Error (MAE). The mean absolute difference between the p50 and the IEX clearing price over every 15-minute block, in ₹/MWh. Blocks that cleared at the ₹10,000 price cap are real prices and are graded like any other.
Naive benchmark and skill. For DAM the naive guess is the same block on the previous day. For RTM it is the DAM clearing price for the same block on the issue day: at the 06:25 IST issue most of that day's RTM has not cleared yet, while its DAM cleared the afternoon before. Skill is 1 minus the forecast MAE over the naive MAE: above zero beats the guess. "Days beating naive" counts the days with the lower error.
RTM before 7 Oct 2026. Until then the RTM tab graded the RTM day-ahead model at 00:00 IST on the delivery day, against the same RTM block the day before. That model was trained on cleared DAM prices but run on the DAM forecast, and it did worse than the naive guess, so it was retired. The RTM series now starts on 27 Sept 2026, the first delivery day of the next-day forecast; the old grades are kept in the archive, unchanged, and the change is in the corrections log.
Coverage. The share of blocks whose clearing price fell inside the p10–p90 band. A calibrated band covers about 80%.
Bad days stay in. A day with no forecast before the cutoff is a graded failure, listed first among the worst days. Each worst day carries a reason: the archive's own (no forecast, the 06:15 IST fallback, cap-pinned blocks, a Grid Event Ledger event) or an analyst's note.
Not the daily-average error. The 7-day tracker on IEX forecasts reports the error of each day's average price, which hides block-level misses that cancel out across the day. This scorecard grades every block.