Ensemble forecasting combines multiple forecasting methods — typically statistical methods (exponential smoothing, ARIMA, Holt-Winters, Croston) and machine learning methods (gradient boosting, neural networks, Prophet variations) — to produce more accurate forecasts than any single method delivers across diverse SKU portfolios. The combination matters more than any individual method: different demand patterns are best forecast by different methods, and ensemble approaches with automatic per-SKU model selection pick the right method for each SKU.
This article covers the mechanics, when ensemble outperforms single methods, when single methods are sufficient, and what to look for in ensemble implementations.
Horizon's ensemble approach includes statistical methods (exponential smoothing variants, ARIMA family, Croston, Syntetos-Boylan-Approximation for intermittent demand), ML methods (gradient boosting, Prophet variations), and causal methods incorporating external factors. Automatic per-SKU model selection happens through holdout accuracy evaluation with stability checks to avoid noise-driven method flipping. Planners can see which method was selected for each SKU and why. NPI methods (analog-based, lifecycle curve, structured judgment) integrate alongside the standard methods.
This sized for mid-market portfolios ($100M-$3B revenue, 500-5,000 SKUs) with the breadth typical of mature mid-market platforms. Enterprise platforms (o9, Kinaxis, SAP IBP, Blue Yonder) include broader ensemble implementations matched to enterprise complexity. Specialists (ToolsGroup, Flowlity) focus on specific aspects (probabilistic methods, intermittent demand) with depth in those areas.
Real-world portfolios contain diverse demand patterns. Some SKUs have stable demand fit best by exponential smoothing. Some have seasonal patterns fit best by Holt-Winters or seasonal ARIMA. Some have intermittent demand best fit by Croston or Syntetos-Boylan-Approximation. Some have promotion-driven demand fit best by causal regression or ML methods. Some have new-product launch patterns requiring analog-based or judgmental forecasting.
Forcing a single method across this diverse portfolio produces systematic accuracy problems — some SKUs forecast well, others poorly, average accuracy below what's achievable. Ensemble methods address this by selecting appropriate methods per SKU pattern, capturing the strengths of multiple approaches.
Modern ensemble approaches typically include 8-20 methods across categories. Statistical methods include: simple moving average (baseline for stable demand), exponential smoothing variants (single, double, triple/Holt-Winters for different patterns), ARIMA and SARIMA (autoregressive integrated moving average, with seasonal variants), Croston's method (intermittent demand), Syntetos-Boylan-Approximation (improved intermittent demand handling). Machine learning methods include: gradient boosting (XGBoost, LightGBM variants), random forests, neural network approaches (LSTM for time series), Prophet (Facebook's open-source time series approach). Each method has demand patterns it handles well and patterns it handles poorly.
For each SKU, ensemble platforms typically: train each method on historical data, evaluate each method on holdout data (recent history reserved from training), select the method with best holdout accuracy, sometimes use multiple methods in weighted combination if no single method clearly dominates. The selection happens automatically — planners don't choose methods per SKU manually.
The selection logic typically uses metrics like MAPE, WMAPE, or MAD on holdout data. Some implementations also consider stability (method consistency across multiple holdout periods) to avoid selecting methods that look great on one period but vary widely.
Demand patterns change. A SKU best forecast by exponential smoothing today may shift to seasonal patterns over 6 months, making Holt-Winters more appropriate. Ensemble platforms re-evaluate method selection periodically (typically monthly or quarterly), automatically switching methods when better fits emerge. This continuous re-evaluation is one of the major advantages over single-method platforms — adaptation happens automatically rather than requiring manual reconfiguration.
Ensemble methods can't forecast products without history through standard training. New product introduction (NPI) forecasting requires different approaches: analog-based methods (find similar products, apply their patterns), judgmental forecasting frameworks (structured planner input based on market knowledge), expected lifecycle curves (assumed launch pattern with adjustments). Mature ensemble platforms typically include these NPI-specific approaches alongside the standard methods.
Pure time-series methods forecast based on historical patterns alone. Ensemble approaches often include causal methods that incorporate external factors: pricing, promotional activity, marketing campaigns, weather, economic indicators, market events. These methods can deliver substantial accuracy gains for SKUs with significant causal influences. They require more data infrastructure (the causal factors must be captured and clean) but deliver value when conditions are right.
The strongest ensemble case: portfolios with diverse demand patterns. CPG manufacturers with mix of fast-moving baseline products, seasonal SKUs, promotional items, and new product launches. Industrial manufacturers with mix of standard MTS products, customer-specific configurations, and aftermarket parts. Each pattern benefits from different methods; ensemble captures the variety.
If your current forecasting uses a single method but your portfolio has patterns that method doesn't handle well, ensemble delivers substantial improvement. Example: pure exponential smoothing across a portfolio with significant seasonal SKUs — those seasonal SKUs forecast poorly until methods like Holt-Winters get applied.
Demand patterns evolve. Products launch, mature, decline. Markets shift. Continuous re-evaluation of method fit captures these changes. Single-method approaches require manual reconfiguration when patterns shift, which usually doesn't happen in practice.
Not every situation requires ensemble. Cases where single methods deliver adequate accuracy: homogeneous portfolios (e.g., pure capital equipment with consistent demand patterns), small SKU counts where per-SKU manual method selection is practical, organizations with deep forecasting analytical capability that can manage method selection thoughtfully. The honest qualifier: most mid-market portfolios are diverse enough to benefit from ensemble, but small or homogeneous operations may not.
Not all ensemble forecasting is created equal. Quality differentiators: number of methods included (8-20 typical for mature implementations, fewer suggests limited capability), automation depth (fully automatic per-SKU selection versus partial automation requiring planner choices), transparency (planners can see which method was selected and why, for trust and refinement), stability handling (avoiding method-flipping where selection changes month-to-month based on noise), causal factor integration (ability to incorporate external factors when relevant), NPI handling (specific methods for new product forecasting), continuous re-evaluation cadence (monthly typical for active markets, quarterly for stable markets).
