Trent TurnerEnvironmental Demand Signals
Testing whether weather adds predictive value beyond seasonality
- Question
- Does weather improve forward forecasts after recurring seasonality and recent demand are accounted for?
- Method
- Weather inputs, lags, rolling windows, and forecast-time availability.
- Evidence
- 2025 earlier weather-model MAE: 2.90 vs. 2.43 for the recent 4-week benchmark.
- Decision
- Exclude weather from the frozen forecast; revisit using forecast-vintage inputs.
Business Question
Does weather improve forward forecasts after recurring seasonality and recent demand are accounted for?
Why It Matters
Temperature, dew point, and demand can move together simply because they share an annual cycle. A plausible explanation must still earn its place through predictive improvement.
Data / Method
Tests covered current values, lags, rolling windows, seasonality-adjusted relationships, and chronological rolling-origin forecasts. Forecast-time availability was part of the evaluation.
Analysis
Raw relationship
Temperature and dew point shared strong seasonal movement with demand.
After seasonal adjustment
Much of the apparent independent signal weakened.
Qualitative comparison reported by the study. Paired raw and adjusted observations are not supplied, so no scatterplot or correlation coefficient is reconstructed.
Results
Weather testing identified no incremental forecast value
2025 MAE · earlier weather-model evaluation
The earlier weather-enhanced model had correlation 0.855 but higher forecast error than the recent four-week benchmark. Its first Demand Score was classified as not validated.
Interpretation
Correlation is not the same as incremental predictive value. No weather variable added enough forward value after the adaptive seasonal structure was modeled correctly.
Limitations
Realized current-week weather is not automatically available at prediction time. Future tests require forecast-vintage weather inputs. Historical visualizations used inconsistent target and model terminology; the reported evaluation is distinguished here from the retained model.
How this fits into the forecasting system
Rejected candidate features: weather remains potential explanatory or scenario context, but is not a final frozen-model component.
Next: Adaptive Demand Level