Datasets:
| model,author,family,inputs,enabled,reference,description | |
| chronos2,Foretec,chronos,history,True,False,"Amazon Chronos-2 (120M), zero-shot on the target's own history, native quantile head." | |
| chronos2_cov,Foretec,chronos,"history,nwp,load_forecast,fuel",True,False,"Chronos-2 (120M) zero-shot with covariates. Known-future NWP ensemble means and spreads (wind 100 m, irradiance, cloud, temperature), ENTSO-E load forecast (price only) and fuel cost, over the history and the delivery day." | |
| chronos_bolt,Foretec,chronos,history,True,False,"Amazon Chronos-Bolt base (205M), zero-shot, last 2048 quarter-hours of history." | |
| climatology,Foretec,baseline,history,True,False,"Multi-year climatology. Mean and 10-90th percentiles of all values at the same CET/CEST time of day within +/-14 calendar days, 2023-01-01 to 2026-08-31 (before the backtest window). No recent information at all." | |
| lgbm_cov,Foretec,nwp,"history,nwp,load_forecast,fuel,calendar",True,False,"LightGBM per series on the NWP ensemble (ICON-EU, GFS, IFS at capacity-weighted centroids, ensemble mean/spread/disagreement), ENTSO-E load forecast (price only) and calendar; for price also wind/solar in all three zones and fuel cost. Trained on the 28-day window, one model per quantile." | |
| lgbm_lags,Foretec,baseline,"history,calendar",True,False,"LightGBM per series on the series' own history, trained on the last two years and retrained monthly (through the end of the previous month). Direct day-ahead: one row per quarter-hour from the history cut-off to the end of D, with the same quarter-hour 1 to 7 days earlier (only what is public at the cut-off), the latest level and its deviation from yesterday fading over the horizon, steps ahead, time-of-day Fourier terms and day of week. Intervals from held-out residuals per hour of day. Same pipeline as lin_lags; only the regressor differs. Before 14 Oct 2026: recursive LightGBM on two days of lags, trained on the 28-day window." | |
| lgbm_prod,Foretec,nwp,"history,nwp,load_forecast,fuel,calendar",True,False,"LightGBM per series on the NWP ensemble (ICON-EU, GFS, IFS at capacity-weighted centroids, ensemble mean/spread/disagreement), the ENTSO-E load forecast (price only), fuel and calendar, set up the way the backtest ablation (Oct 2026) found to matter: two years of history, recency-weighted (half-life 120 days), one LightGBM per quantile; solar trained on its capacity factor (installed capacity from the Marktstammdatenregister, Elia and Energinet for DE-LU, BE, DK1, DK2; a recent-peak proxy for FR and NL); price and load retrained every night on data through the day before, wind and solar monthly. Ablation: daily retraining cut price and load error by 15-23 % against monthly, the capacity factor cut solar error by 12-27 %, tuning the LightGBM parameters gained nothing." | |
| lin_cov,Foretec,nwp,"history,nwp,load_forecast,fuel,calendar",True,False,"Linear regression (ridge, penalty by cross-validation) on exactly the lgbm_cov inputs: the NWP ensemble (ICON-EU, GFS, IFS at capacity-weighted centroids, ensemble mean/spread/disagreement), ENTSO-E load forecast (price only) and calendar; for price also wind/solar in all three zones, fuel cost and yesterday's price. Trained on the 28-day window like lgbm_cov; intervals from the residuals of the last 7 days per hour of day." | |
| lin_cov_2y,Foretec,nwp,"history,nwp,load_forecast,fuel,calendar",True,False,"The two-year variant of lin_cov: Linear regression (ridge, penalty by cross-validation) on exactly the lgbm_cov inputs and training: the NWP ensemble, ENTSO-E load forecast (price only), calendar (for price also wind/solar in all zones, fuel and yesterday's price), the last two years, retrained monthly. Intervals from the residuals of the last 120 training days per hour of day." | |
| lin_lags,Foretec,baseline,"history,calendar",True,False,"Linear regression (ridge, penalty by cross-validation) on exactly the lgbm_lags pipeline: the series' own history over the last two years, retrained monthly; the same quarter-hour 1 to 7 days earlier, the latest level and its fading deviation from yesterday, steps ahead, time-of-day Fourier terms and day of week. The classical autoregression-with-seasonal-regressors benchmark, in place of ARIMA and ETS, which do not suit 15-minute data with a 96-step daily cycle." | |
| moirai2,Foretec,moirai,history,True,False,"Salesforce Moirai 2.0 small, zero-shot quantile forecasts." | |
| moirai2_cov,Foretec,moirai,"history,nwp,load_forecast,fuel",False,False,Moirai 2.0 small zero-shot with known-future covariates as feat_dynamic_real. | |
| naive_daily,Foretec,baseline,history,True,False,Same quarter-hour on the last available day. | |
| naive_weekly,Foretec,baseline,history,True,False,Same quarter-hour one week earlier. The baseline every entry is measured against. | |
| sundial,Foretec,sundial,history,True,False,"Tsinghua Sundial base (128M), zero-shot generative, 50 sample paths." | |
| t0_alpha,The Forecasting Company,t0,history,True,False,"The Forecasting Company t0-alpha (102M parameters), zero-shot on the target's own history." | |
| t0_alpha_cov,The Forecasting Company,t0,"history,nwp,load_forecast,fuel",True,False,"The Forecasting Company t0-alpha, zero-shot with known-future covariates (NWP ensemble means and spreads, ENTSO-E load forecast (price only), fuel cost) over the history and the delivery day." | |
| t0_beta,The Forecasting Company,t0,history,True,False,"The Forecasting Company t0-beta (256M parameters), zero-shot on the target's own history." | |
| t0_beta_cov,The Forecasting Company,t0,"history,nwp,load_forecast,fuel",True,False,"The Forecasting Company t0-beta, zero-shot with known-future covariates (NWP ensemble means and spreads, ENTSO-E load forecast (price only), fuel cost) over the history and the delivery day." | |
| t0_gamma,The Forecasting Company,t0,history,True,False,"The Forecasting Company t0-gamma (about 256M parameters, third public t0 release, October 2026), zero-shot on the target's own history. Runs on tfc-t0 0.6.0, which its weights require." | |
| t0_gamma_cov,The Forecasting Company,t0,"history,nwp,load_forecast,fuel",True,False,"The Forecasting Company t0-gamma, zero-shot with known-future covariates (NWP ensemble means and spreads, ENTSO-E load forecast (price only), fuel cost) over the history and the delivery day. Runs on tfc-t0 0.6.0, which its weights require." | |
| tabby,Huawei Noah's Ark Lab (Paris),tabby,history,True,False,"Tabby (paris-noah/Tabby, September 2026, CC BY-NC 4.0): a PatchTST foundation model (147M parameters) whose frozen pretrained backbone carries a prompt post-trained on GIFT-Eval training tasks; zero-shot on the target's own history, 99 quantiles. Each series is forecast on its own (no covariates)." | |
| tabby_cov,Foretec,tabby,"history,nwp,load_forecast,fuel",True,False,"Tabby (paris-noah/Tabby) with known-future covariates in TimesFM's ""xreg + timesfm"" style: a ridge on the NWP ensemble means/spreads, ENTSO-E load forecast (price only) and fuel cost explains what the covariates can, Tabby forecasts the residual zero-shot (it has no covariate input of its own), and the ridge forecast is added back to every quantile. Load is forecast by Tabby alone: on weather only, the ridge stood in for the daily cycle and cost more than it explained (NL backtest Rel. MAE 1.9 against 0.65)." | |
| tabby_pretrain,Huawei Noah's Ark Lab (Paris),tabby,history,True,False,"Tabby-Pretrain (paris-noah/Tabby-Pretrain, CC BY-NC 4.0): the pretrained PatchTST backbone of Tabby without the post-trained prompt (146M parameters), zero-shot on the target's own history through the tabby-tsfm package, 99 quantiles." | |
| tabby_pretrain_cov,Foretec,tabby,"history,nwp,load_forecast,fuel",True,False,"Tabby-Pretrain (paris-noah/Tabby-Pretrain) with known-future covariates in TimesFM's ""xreg + timesfm"" style: a ridge on the NWP ensemble means/spreads, ENTSO-E load forecast (price only) and fuel cost explains what the covariates can, the backbone forecasts the residual zero-shot, and the ridge forecast is added back to every quantile. Load is forecast by the backbone alone: on weather only, the ridge stood in for the daily cycle and cost more than it explained (NL backtest Rel. MAE 1.9 against 0.65)." | |
| timesfm25,Foretec,timesfm,history,True,False,"Google TimesFM 2.5 (200M), zero-shot, continuous quantile head." | |
| timesfm25_cov,Foretec,timesfm,"history,nwp,load_forecast,fuel",True,False,"TimesFM 2.5 with in-context linear regression on the covariates (xreg + timesfm): a ridge fit on the history explains what the covariates can, TimesFM forecasts the rest." | |
| timesfm3,Foretec,timesfm,history,True,False,"Google TimesFM 3.0, zero-shot, target history only." | |
| timesfm3_cov,Foretec,timesfm,"history,nwp,load_forecast,fuel",True,False,"TimesFM 3.0 zero-shot with its native known-future covariates (NWP ensemble means/spreads, ENTSO-E load forecast (price only), fuel cost)." | |
| timestral,Mistral AI,timestral,history,True,False,"Timestral-209M (mistralai/Timestral-209M, October 2026, Modified MIT weights): a causal patch transformer with decoupled forecast tokens and covariate-aware attention, zero-shot on the target's own history, 21 quantiles." | |
| timestral_cov,Mistral AI,timestral,"history,nwp,load_forecast,fuel",True,False,"Timestral-209M zero-shot with its native known-future covariates (NWP ensemble means/spreads, ENTSO-E load forecast, fuel cost) over the history and the delivery day." | |
| tirex,Foretec,tirex,history,True,False,"NX-AI TiRex (35M, xLSTM), zero-shot quantile forecasts." | |
| tirex2,Foretec,tirex,history,True,False,"NX-AI TiRex-2 (xLSTM, 38M active parameters in univariate mode, Apache-2.0, NeurIPS 2026), zero-shot quantile forecasts from the target's own history." | |
| tirex2_cov,Foretec,tirex,"history,nwp,load_forecast,fuel",True,False,"NX-AI TiRex-2 with known-future covariates: the NWP ensemble (ICON-EU, GFS, IFS) at capacity-weighted centroids, the ENTSO-E load forecast (price only), fuel cost and calendar, the same compact feature set as chronos2_cov, standardised on their history. Zero-shot." | |
| toto,Foretec,toto,history,True,False,"Datadog Toto Open Base 1.0 (151M), zero-shot, 128 sample paths." | |
| toto_cov,Foretec,toto,"history,nwp,load_forecast,fuel",False,False,"Toto Open Base 1.0 zero-shot, covariates as extra variates with their known future values injected while decoding." | |
| tso_cutoff,"Elia, NED.nl",tso,tso_forecast,True,False,"The grid operators' own day-ahead wind, solar and load forecasts as they stood before the gate. BE: Elia's forecast published at 11:00 on D-1 (Elia Open Data). NL: the NED.nl national wind (onshore + offshore) and solar forecast (Nationaal Energie Dashboard, backed by TenneT), fetched live at the cut-off, the same quantity as the NL actuals (NED's national estimate). Load, BE and FR: the TSO's day-ahead total load forecast on ENTSO-E (A65), which must be published two hours before the gate. NL load: NED.nl's national load forecast, fetched live at the cut-off, the same gross consumption as the NL load actuals (TenneT's ENTSO-E forecast covers a different quantity). FR wind and solar: none, because RTE publishes its next-day wind and solar forecast at 16:15 on D-1, after the gate, and offers no earlier (D-2) vintage for wind or solar. Live only: what was published before the gate cannot be verified after the fact, so there is no backtest. Point forecasts." | |
| tso_final,"RTE, Elia, NED.nl, ENTSO-E / TSOs",tso,tso_forecast,True,True,"Reference, not ranked: the TSOs' final day-ahead wind and solar forecast, published after the 12:00 gate. FR: RTE's day-ahead forecast published at 16:15 on D-1 (RTE Generation Forecast API); BE: Elia's day-ahead forecast published at 18:00 on D-1 (Elia Open Data, 15 minutes; ENTSO-E only has it hourly). DE-LU, DK1, DK2: the TSOs' day-ahead forecast on ENTSO-E (A69), due by 18:00 on D-1. NL: NED.nl's national wind and solar forecast as it stands at the evening scoring run (after 18:00 on D-1), the same quantity as the NL actuals; NED overwrites its forecast, so NL is live only (TenneT's ENTSO-E forecasts cover a narrower scope). Fetched at the evening scoring run, once published; backtest uses the published version." | |
| ttm_r2,Foretec,ttm,history,True,False,"IBM Granite TinyTimeMixer r2 (~1-5M), zero-shot point forecast, 1536-step context." | |