AUGUR_

Day-ahead solar, forecast with 64% lower error
than Transelectrica's own forecast.

Root-mean-square error (RMSE) against Transelectrica's own published day-ahead forecast, same delivery slots, same information cutoff. Public data. Leak-safe, vintage-controlled — the weather each forecast used was frozen to what was knowable at gate time, never the realized weather. Out-of-sample test window, 2026. Losses shown, not hidden.


The result — skill vs Transelectrica day-ahead (RMSE)

Positive = we beat Transelectrica's own published forecast for the same delivery slots, at the same information cutoff. Lead is how far ahead the forecast is issued.

target1h2h3h4h8h16h24h (day-ahead)
solar+70%+66%+65%+64%+64%+64%+65%
wind onshore+30%+19%+12%+9%+5%+2%-1%
large win (≥40%) solid win (10–40%) ~parity (0–10%) Transelectrica wins (<0)

Source manifest: experiments/x02/asof_engines_ro.json (LightGBM residual, ENTSO-E actuals, ICON-EU + ECMWF-IFS + GFS vintage weather, n ≈ 15,500 test points/cell).

Why it works — and why we trust it

The mechanism. Transelectrica's day-ahead solar forecast over-forecasts by ~35% at midday (mean day-ahead 1,194 MW vs 883 MW actually delivered; peak 3,457 MW vs 2,600 MW). That is a large, persistent, structural bias. A residual model that learns to correct it wins by a wide margin — and keeps winning no matter how far ahead the forecast is issued.

Solar skill is flat across every horizon — ~+70% at 1 hour, still ~+64% at 24 hours. That flatness is the proof it is bias-correction, not weather leakage. Leaked future weather would inflate the long horizons and then collapse under a leak-safe rebuild — which is exactly what happened to the German and Belgian solar numbers when we ran the same audit on them. Romania's solar survived it intact, because the edge never came from weather at all; it came from correcting a broken incumbent forecast.

We checked the obvious artifact too: solar is zero at night, which is trivially easy to predict. Restricting the score to daylight-only slots gives the same +64% — the skill is real daytime accuracy, not a night-zeros illusion.

What we don't claim

Wind day-ahead is honest parity. Romanian onshore wind shows a real short-horizon nowcasting edge (~+31% one hour ahead) that decays to roughly zero by the day-ahead horizon — Transelectrica's wind forecast is decent, so there is no standing bias to exploit. We report it at parity and do not claim a day-ahead wind win. (This is the mirror image of Belgium, where wind survived the audit and solar did not.)

Not yet live on a Romanian market. This is a backtest on public data — rigorous, reproducible, leak-safe, but a backtest. Our live track record is on Belgian solar, where we submit hourly probabilistic forecasts to Elia's Predico platform and have ranked as high as 4th of 30 forecasters. A live Romanian feed is the natural next step.

Built by Eloquentix Inc. — the engineering firm behind the E.ON Virtual Power Plant and the Sonnen/Shell battery-storage platform. Evaluation: strict temporal split (train < 2026-01-01 ≤ test), no lookahead, public data only (energy-charts.info / ENTSO-E for Transelectrica actuals + forecasts, Open-Meteo archived forecast runs for weather).  ·  ← all markets