random forest

Random Forest

GPU-accelerated Random Forest ensemble trained via cuML, producing one model per forecast horizon (6 total). Excels at short-term predictions (Lead 1–2).

Architecture

Parameter	Value
Library	cuML (NVIDIA RAPIDS)
Trees per model	50
Models	6 (one per Lead 1–6)
Input	Flat feature vector from Panel Construction
Training	GPU-accelerated

Unlike the sequence models (LSTM, Transformer), each RF model is trained independently on a specific lead target, receiving the full 18-feature lag vector as a flat input.

Strengths

• Best short-term F1: 0.97 at Lead 1
• Highest recall across short horizons — catches nearly all genuine surges
• Robust: No hyperparameter sensitivity, handles missing features gracefully
• Fast training: GPU acceleration via cuML

Weaknesses

• RMSE degrades at Lead 4–6 (long-horizon patterns are harder for tree ensembles)
• No sequence modeling: Treats lag features independently rather than as a temporal sequence
• 6 separate models: No parameter sharing across horizons

Role in Ensemble

The Horizon-Aware Ensemble gives RF the highest weight at short horizons:

Lead	RF Weight
1	60%
2	50%
3	30%
4	20%
5	10%
6	5%

Performance

Lead	F1	Precision	Recall	RMSE
1	0.97	0.96	0.97	Best at lead 1
6	~0.82	~0.78	~0.87	Degrades

See Model Performance for full comparison.

Artifacts

src/models/trained_models/
├── rf_lead_1.joblib
├── rf_lead_2.joblib
├── rf_lead_3.joblib
├── rf_lead_4.joblib
├── rf_lead_5.joblib
└── rf_lead_6.joblib

See Also

• Horizon-Aware Ensemble — How RF is weighted in production
• LSTM — Mid-horizon alternative
• Transformer — Long-horizon alternative
• Training Pipeline — Training process
• GPU Acceleration — cuML infrastructure