Ensemble Forecasting

Ensemble methods directly address the fundamental bias-variance decomposition of prediction error. High-bias models (e.g., linear regression) systematically underfit complex demand patterns, while high-variance models (e.g., deep neural networks) overfit noise in training data. By combining models with uncorrelated error structures, ensembles reduce variance without introducing additional bias.

• Bagging (Bootstrap Aggregating) reduces variance by training identical model architectures on different bootstrap samples of the data and averaging predictions
• Boosting reduces bias by sequentially training models where each new model focuses on the errors of its predecessors
• Stacking uses a meta-learner to optimally weight the outputs of heterogeneous base models
• The error reduction is proportional to the average error of individual models minus the diversity of their predictions

The effectiveness of an ensemble depends critically on model diversity—the degree to which constituent models make different errors on the same data points. Correlated errors provide no benefit; uncorrelated errors cancel out. Diversity is engineered through multiple orthogonal strategies.

• Algorithmic diversity: Combining fundamentally different model families (ARIMA, LSTM, Gradient Boosting, Temporal Fusion Transformer) that capture different structural patterns
• Data diversity: Training identical architectures on different subsets, time windows, or resampled versions of the training data
• Feature diversity: Providing different subsets of exogenous variables to each model, forcing them to learn from different information signals
• Hyperparameter diversity: Varying learning rates, regularization strengths, and architecture depths across otherwise identical models
• Horizon-specific specialization: Training separate models optimized for short-term (1-7 day) versus long-term (30-90 day) forecast horizons

Aggregating point forecasts is straightforward (weighted average), but ensemble forecasting in supply chains requires probabilistic aggregation to preserve uncertainty quantification. The method of combining predictive distributions determines the quality of the final prediction intervals.

• Linear opinion pool: A weighted mixture of individual predictive distributions, preserving multi-modality when models disagree on likely outcomes
• Bayesian model averaging: Weights each model by its posterior probability of being the true data-generating process, naturally penalizing overfit models
• Quantile averaging: Independently averages each quantile across models, ensuring the aggregated 95th percentile is the mean of individual 95th percentiles
• Vincentization: A non-parametric method that averages the quantile functions themselves, producing a coherent aggregated distribution
• CRPS-weighted stacking: Optimizes model weights by minimizing the Continuous Ranked Probability Score on a validation set

A critical operational advantage of ensembles in production supply chain systems is graceful degradation under model failure. When a single forecasting model encounters a regime shift, data pipeline error, or concept drift, its predictions can become catastrophically wrong. An ensemble provides inherent fault tolerance.

• If one constituent model produces nonsensical forecasts (e.g., negative demand), outlier detection can automatically down-weight or exclude it from aggregation
• Diversity acts as implicit regularization, preventing the ensemble from over-indexing on spurious patterns that any single model might latch onto
• During cold-start scenarios for new products, ensembles can weight transfer-learning models more heavily while gradually incorporating models trained on accumulating history
• The ensemble's prediction interval naturally widens when constituent models disagree, providing an automatic signal of elevated uncertainty that can trigger human review

Ensemble forecasting operates on a compute-accuracy Pareto frontier. Each additional model improves forecast accuracy with diminishing returns while linearly increasing inference latency and infrastructure cost. Production systems must navigate this tradeoff.

• Inference latency: A 10-model ensemble requires 10 forward passes, potentially violating sub-second response SLAs for real-time demand sensing applications
• Parallelization: Since constituent models are independent at inference time, batched execution across GPU clusters can maintain latency while scaling ensemble size
• Pruning strategies: Recursive feature elimination can identify and remove models that contribute minimal marginal improvement to CRPS, yielding a minimal viable ensemble
• Tiered ensembles: A lightweight 3-model ensemble serves real-time operational decisions, while a heavyweight 15-model ensemble runs nightly for strategic planning horizons
• Typical enterprise deployments balance 5-8 diverse models, achieving 15-25% CRPS improvement over the single best model

A probabilistically calibrated ensemble produces prediction intervals that contain the true realized value at the advertised frequency. A well-calibrated 90% prediction interval should contain the actual demand 90% of the time. Calibration and sharpness are independently assessed.

• Probability Integral Transform (PIT) histogram: A uniform PIT distribution indicates perfect calibration; U-shaped or hump-shaped histograms reveal systematic overconfidence or underconfidence
• Reliability diagrams: Plot observed frequency against forecast probability, where deviation from the diagonal identity line indicates miscalibration
• Sharpness: Measures the concentration of predictive distributions—narrower intervals are sharper but only valuable if calibrated
• Isotonic regression recalibration: A post-hoc non-parametric method that adjusts the ensemble's output probabilities to achieve calibration without retraining constituent models
• Quantile calibration: Verifies that each forecast quantile (e.g., 10th, 50th, 90th) matches the empirical quantile of observed outcomes in backtesting