Model Architecture
Optim1.1 uses a four-model ensemble with Bayesian model averaging. Each algorithm contributes based on its domain expertise and historical calibration accuracy. The pipeline from raw data to forecast output is fully automated.
Prediction Pipeline
Ensemble Components
Long Short-Term Memory network fine-tuned with attention gates. Captures long-range sequential dependencies across multi-year time series. Especially effective for economic cycles and climate seasonality.
Global Feature Importance
Averaged across all models and domains. Values represent normalised Shapley (SHAP) contribution scores.
Confidence Calibration
Confidence scores are derived from ensemble disagreement, historical calibration error, and signal-to-noise ratio of live feeds. Lower confidence does not mean the forecast is wrong — it indicates wider uncertainty bounds.