v4.1 — Quantile Loss + Weather Interactions

Date: March 2, 2026

What Changed

Two improvements merged into a single release: a loss function change that corrects systematic underprediction, and 15 new features that encode physical relationships between weather and electricity price drivers. All 4 forecasting approaches (hourly, expanded, pure15, hybrid15) were retrained and backtested.

Quantile Loss (q=0.55)

Spanish electricity prices are right-skewed: bounded near zero, with occasional spikes above 200 EUR/MWh. Standard loss functions have a structural problem:

• MSE targets the conditional mean — sits below the median on skewed data
• MAE targets the conditional median — closer, but still below the mean

Quantile loss at q=0.55 targets the 55th percentile, slightly above the median. This directly corrects the underprediction bias without distorting the price forecast shape.

Model	Loss Configuration
HistGBT	loss="quantile", quantile=0.55
LightGBM	objective="quantile", alpha=0.55
XGBoost	objective="reg:quantileerror", quantile_alpha=0.55

The optimal quantile was set through cross-validation over [0.50, 0.52, 0.54, 0.56, 0.58].

Weather Interaction Features

Raw weather variables (temperature, wind speed, cloud cover) have no inherent price meaning. A 30 km/h wind tells the model nothing about prices unless it knows how much of the generation mix depends on wind. These 15 interaction features encode the physical relationships that matter:

Origin-Time Features (at forecast time)

Feature	Formula	Why It Matters
weather_cold_x_demand	max(0, 15-temp) x demand	Cold weather + high demand = expensive heating load
weather_temp_deviation	temp - 15	Distance from mild baseline
weather_heating_degree_days	max(0, 18-temp)	Standard heating load proxy
weather_cooling_degree_days	max(0, temp-24)	Cooling load (summer AC)
weather_temp_deviation_sq	(temp-15)^2	Non-linear: extremes cost more than mild deviations
weather_wind_x_wind_share	wind_speed x wind_share	Wind blowing fast AND turbines are a large generation share
weather_precip_x_hydro	precip x hydro/demand	Rain boosting hydro reservoir output
weather_cloud_x_solar	cloud_cover x solar_share	Clouds suppressing solar when solar is a large share
weather_sunshine_x_solar	sunshine x solar_share	Clear skies amplifying solar output
weather_ghi	direct + diffuse radiation	Total solar energy reaching the surface

Target-Time Features (at predicted hour)

Feature	Formula	Why It Matters
target_weather_temp_deviation	temp - 15	Expected temperature comfort deviation
target_weather_heating_dd	max(0, 18-temp)	Expected heating load
target_weather_cooling_dd	max(0, temp-24)	Expected cooling load
target_weather_wind_x_wind_share	wind x wind_share	Expected wind displacement of expensive generators
target_weather_cloud_x_solar	cloud x solar_share	Expected solar suppression
target_weather_ghi_x_solar	radiation x solar_share	Expected solar generation value

These features give the model pre-computed domain knowledge that would otherwise require many tree splits to approximate. For example, weather_wind_x_wind_share directly encodes “wind is blowing fast AND wind is a large share of generation” — something a tree model would need to learn by splitting on both variables independently.

Configuration

Parameter	Value
Loss function	Quantile 0.55
Ensemble	3 models (HistGBT, LightGBM, XGBoost)
Features	76–78 (57 base + 15 weather interactions + weather raw)
Feature selection	No
Peak-split	No

Multi-Approach Training

All 4 forecasting approaches were retrained from scratch with quantile loss + weather features:

Approach	Models Trained	Training Data
Hourly	6	3+ years hourly (32K rows)
Expanded	Uses hourly models	Hourly models expanded to 15-min
Pure15	6	5 months genuine 15-min (14K rows)
Hybrid15	6	3+ years expanded + 5 months genuine (146K rows)

Total: 18 models (3 model types x 2 run modes x 3 approaches requiring training).

Bug Fixes

• LazySection spacing: Lazy-loaded dashboard sections now use flex + gap: 1.5rem, matching parent dashboard spacing
• Batch evaluation tag injection: /evaluation/summary/batch and /evaluation/timeline/batch now correctly translate tagged model names (e.g., ensemble_backtest + tag=v4.1 -> ensemble_v4.1_backtest)
• Backtest --approach all: New flag runs all 4 approaches in a single command

Results

149-day backtest (Oct 2025 – Feb 2026), ensemble models at hourly resolution:

Approach	Run Mode	Baseline MAE	v4.1 MAE	Change	Baseline Bias	v4.1 Bias
Hourly	Dayahead	16.40	13.42	-18.2%	-13.77	-8.68
Hourly	Strategic	20.42	20.47	+0.2%	-17.38	-16.93
Expanded	Dayahead	17.28	14.00	-19.0%	-14.34	-8.96
Expanded	Strategic	21.25	20.99	-1.2%	-17.78	-17.22
Hybrid15	Dayahead	16.63	15.04	-9.6%	-14.19	-10.24
Hybrid15	Strategic	21.18	23.27	+9.9%	-17.68	-20.37
Pure15	Dayahead	14.09	15.00	+6.5%	-13.46	-9.85
Pure15	Strategic	26.31	28.87	+9.7%	-25.24	-28.43

Summary: Dayahead MAE improved 18-19% for the main hourly and expanded approaches, with bias reduced by 37%. Hourly and expanded strategic forecasts (D+2-D+7) are essentially flat — expected, since weather interaction features encode local conditions that lose predictive value over multi-day horizons. Pure15 and hybrid15 strategic regressed (~+10% MAE) due to limited training data combined with 15 additional features increasing overfitting risk. Pure15 dayahead also regressed slightly (+6.5%) for the same reason.

Implementation Notes

• Quantile target configured as QUANTILE_TARGET = 0.55
• Weather interaction features computed during both training and prediction
• Backtest supports --approach all to run all 4 approaches in a single command