Ensemble forecasting means blending the output of multiple forecasting models into one combined prediction. If one model forecasts 100 units and another forecasts 150, a genuine ensemble doesn't pick the model it trusts more. It combines them, weighting each model's contribution based on how reliable that model has been for that specific item.
That blend still has to earn its place. Horizon evaluates it against every other candidate model on backtested accuracy, the same way it evaluates a plain statistical or machine learning model, and it doesn't win every time. Sometimes a basic model outperforms the blend for a given SKU, and Horizon uses that model instead.
No, not by itself. Letting single models compete for a win, with no genuine blended candidate in the competition, is model selection, not ensemble forecasting. Many vendors market that as "ensemble" anyway, letting several single models compete on a SKU or category and picking whichever one wins, without ever building an actual blended candidate into the competition.
Real ensemble forecasting means a genuine blend, a weighted combination of models, competes as one of the candidates. It's judged on the same backtested accuracy as everything else in the pool, and it can lose.
If your platform says it ran an "ensemble" and the result traces back to one model's raw output, that's model selection, not blending. A genuine ensemble result is a weighted combination of several models, even though it still shows up as a single forecast number.
Ensemble forecasting isn't something Horizon applies to every item by default, and it isn't a fallback the system reaches for only when nothing else stands out. It's a standard, well-defined candidate: a weighted combination of the model families Horizon evaluates, built the same way every time.
For a given planning item, Horizon weighs how well a blend would fit using signals like historical accuracy, bias, volatility, intermittency, seasonality, causal factors, and new-product behaviour, then compares that blend's backtested performance against every other candidate model for that item. A stable, low-volatility item and a highly seasonal item can end up with different winners, a single model for one, a blend for the other, because those signals favour different candidates for each.
Planners still see the reasoning behind whichever candidate wins. Horizon can show why a model, blended or single, was selected: seasonality, intermittent demand, promotional behaviour, lifecycle stage, or holdout performance. Planners don't lose that visibility just because the forecast might come from a combination of models rather than one.
Most portfolios don't have a single demand pattern. A typical mix includes stable products, seasonal products, intermittent products, fast-growing products, promotional products, and new products with little history.
No single model handles all of that well. A model tuned for stable, high-volume items will misread a seasonal spike. A model built for intermittent demand won't serve a fast mover. A blend that draws on several models at once has more chances to be the best fit somewhere in that mix than any one model does on its own.
That's why ensemble forecasting tends to win more often as a portfolio gets more diverse. A narrow, homogeneous product line gives it fewer openings; single models are more likely to fit well already. A mixed CPG or industrial portfolio, with seasonal SKUs sitting next to promotional items and new launches, gives a genuine blend the most room to earn the win.
Ensemble forecasting is one part of how Horizon approaches demand planning as a whole.
A single best-fit model still misses the parts of demand it isn't built to catch. Where one model underreacts to a promotional spike, another model that's more sensitive to short-term signal can offset it in a blended forecast. When that blend captures more of an item's demand pattern than any single candidate, it wins the backtest and gets used. When it doesn't, the system uses whichever single model does the job better.
Single-method forecasting breaks quietly. A method that fit a SKU's demand pattern six months ago can drift out of fit as that pattern changes, and nothing flags it until accuracy has already slipped. Someone has to notice, investigate, and reconfigure.
Horizon's model-selection process removes that dependency, for every candidate, not just the blend. Every model, single or ensemble, gets backtested and re-evaluated on an ongoing basis, so a shift in fit gets caught automatically rather than requiring a planner to notice and intervene.
Demand patterns don't stay put. A product that sells at a steady, predictable rate can become seasonal once a retailer expands distribution, or intermittent once a category matures and volume drops. A static setup needs someone to catch that shift and manually swap in a different method.
Horizon doesn't wait for that. As a pattern changes, the candidates get re-tested, and whichever one now fits best, a single model or the blend, takes over, sometimes as a shift in a blend's weighting, sometimes as a full changeover from one candidate to another.
New products don't have a history to blend from. Ensemble forecasting, in the strict sense, needs training data, so it can't forecast a SKU with no track record the same way it forecasts an established one, and most single models can't either.
Horizon handles this with separate methods built for that gap: matching a new item to similar products with known demand patterns, applying an expected lifecycle curve with adjustments, and incorporating structured planner input where market knowledge matters more than history. These run alongside the standard model-selection process rather than replacing it and phase out as the new item builds enough history for Horizon to backtest it like anything else in the portfolio.
Not every platform that says "ensemble forecasting" means the same thing. Before taking the label at face value, check:
