Demand forecasting that starts with the best model for every pattern

Demand forecasting software builds a statistical or machine-learning baseline of what a business will need to sell, stock, or produce, then keeps that baseline current as new demand data comes in. Horizon's version doesn't run one model against every product. It runs statistical, machine learning, intermittent, causal, and Horizon-built models against each demand pattern separately, and publishes whichever one actually wins for that specific SKU-location combination.

Horizon runs five categories of forecasting method: statistical models for stable, well-understood demand; machine learning models for complex or high-volume patterns; intermittent-demand models for sparse, irregular SKUs; causal models that factor in external drivers like price or promotions; and Horizon-built models for edge cases the standard library handles poorly, like a brand-new launch with no history. Instead of picking one method upfront, the system runs all of them against every pattern and keeps whichever one actually performs best for that specific case.

AI improves demand forecasting accuracy mainly by matching the model to the pattern instead of forcing one method on every product. A steady, high-volume item and a brand-new launch with zero history need fundamentally different math. Horizon's auto-selection tests model fit by accuracy, bias, and stability for each pattern, and explainability shows which features drove the pick, so a planner can see the reasoning behind the number, not just the number itself.

Horizon groups products and customers into segments based on volatility, growth, seasonality, and value, so each pattern gets analyzed at the right level instead of one-size-fits-all. When demand moves, it pinpoints which factor, price, an event, a customer, the market, or supply, is actually behind it, and separates a genuine change from noise, a data issue, or a one-off event. That distinction matters because reacting to noise wastes planning effort, while missing a real shift creates stockouts or excess inventory.

Yes. Horizon's AI includes a conversational assistant that answers plain-language questions about demand movement, such as which factors are driving it or how two forecast versions differ, instead of requiring a dashboard query or filter setup. It also ranks recommendations by impact and priority so the biggest risks surface first, and suggests the next action based on how a specific segment is behaving right now. Over time, it improves by feeding outcomes back into the model.

Yes. Planners can adjust the forecast directly, and the system requires a reason code on every override. That reason code exists specifically to stop an override from turning into a second, invisible forecast nobody can audit later. Before committing to a change, a planner can also run a what-if scenario against the baseline and see the impact immediately.

Yes. Any two saved sources, planner input, the forecast engine's output, or an earlier version, can be compared across a custom date range. Rather than scanning every line for what changed, a threshold filter surfaces only the items that moved by a meaningful amount, and the largest delta gets called out automatically.

Yes. Multi-level aggregation and disaggregation across SKU, location, and customer runs without performance drops, even at full hierarchy depth.

Yes. Supply planning modules are available, but none of them are mandatory, you can license just the Demand Forecasting engine on its own instead of taking the full suite.