Loading the latest numbers…

Backtest: v2-tuning

Model v2 replayed over 44 days (2026-08-12 to 2026-09-24). These are the days the configuration was chosen on, so the tests below are optimistic. For the first days the weight on our model is zero, until seven days of history exist.

How it compares

Our forecast: 318 MW
IESO: 399 MW
IESO area sum: 366 MW
IESO + recent bias: 341 MW

Mean absolute error over 44 days (2026-08-12 to 2026-09-24).

Pairwise comparisons
Method	Against	Lower error	95% range of the gain	p-value	Days better	Verdict
Our forecast	IESO	20.3%	46 MW to 122 MW	< 0.001	75%	Better
IESO area sum	IESO	8.3%	26 MW to 41 MW	< 0.001	86%	Better
IESO + recent bias	IESO	14.4%	16 MW to 115 MW	0.054	57%	No clear difference
Our forecast	IESO + recent bias	6.8%	-1 MW to 40 MW	0.058	59%	No clear difference

Mean error by hour of day

IESO
IESO area sum
Our forecast

IESO misses most at 6 am: 571 MW on average, with a mean error of −524 MW.

Show the numbers

Mean absolute error by hour ending
Hour ending	Ours	IESO	Area sum	IESO mean error
1 am	228 MW	295 MW	295 MW	−278 MW
2 am	256 MW	334 MW	334 MW	−278 MW
3 am	256 MW	307 MW	307 MW	−231 MW
4 am	270 MW	317 MW	316 MW	−232 MW
5 am	303 MW	474 MW	339 MW	−425 MW
6 am	289 MW	571 MW	317 MW	−524 MW
7 am	321 MW	443 MW	337 MW	−354 MW
8 am	276 MW	343 MW	269 MW	−221 MW
9 am	250 MW	298 MW	253 MW	−127 MW
10 am	324 MW	333 MW	294 MW	−124 MW
11 am	350 MW	329 MW	321 MW	−141 MW
noon	341 MW	365 MW	343 MW	−141 MW
1 pm	344 MW	370 MW	366 MW	−159 MW
2 pm	356 MW	432 MW	432 MW	−248 MW
3 pm	408 MW	556 MW	504 MW	−391 MW
4 pm	423 MW	568 MW	517 MW	−415 MW
5 pm	401 MW	485 MW	485 MW	−328 MW
6 pm	401 MW	433 MW	432 MW	−312 MW
7 pm	332 MW	382 MW	383 MW	−305 MW
8 pm	294 MW

Who was closer, day by day

Hover or tab through the days for details.

Frozen configuration

Frozen at (UTC): 2026-09-26T08:34:13
Base forecast: Sum of IESO's four area forecasts, last version before the 10:00 EPT cutoff
Own forecast: Kumo, zero-shot, predicting the load level; context from the same season in earlier years (±45 days) and the last 60 days, up to 10000 rows
Combination: Weight on our model chosen each day on days already known, needs at least 7 days
Research record: notebooks/load_error/reports (E01 to E06), frozen/v2.yaml