The INFORMS ENRE 2026–27 challenge asks for a 48-hour day-ahead LMP forecast with a 90% interval. This dashboard covers the baselines' rolling-origin backtest. Each forecast day o uses prices and fuel through o, plus load/solar for o+1 and o+2. Extreme hours follow the challenge's definition: above the 95th or below the 5th percentile of the preceding week.
A saved run keeps the full CSV, so anyone with this page can open it again. Scores are 48 h over the run's last 182 forecast days.
Each point is one forecast day's MAE (thin line: focus model). Thick lines are 7-day rolling means. Click a day to open its 48-hour path below.
How the baseline intervals are built. Each model gets its own interval, using empirical residual quantiles (split conformal) kept separately for every (lead, hour) cell.
At forecast day o, take that model's errors actual − forecast from the previous 91 forecast days.
Origin o−1 is skipped, because its day-2 target is not realized yet.
The 5th and 95th percentiles of those 91 errors are added to today's point forecast to give lo and hi. The first 28 days with too few past errors are dropped.
No future information enters: the interval is built only from errors that were observable at o.
Coverage @ 90 is the share of hours in the window where lo ≤ actual ≤ hi (target 0.90).
Winkler is (hi − lo) + 20·(lo − actual) when the actual falls below the band, + 20·(actual − hi) when above, and the width alone otherwise. It is averaged over hours, and lower is better.
For an uploaded CSV, both are computed from its <model>_lo / _hi columns. Models without them are left out of this section.