The predict-then-optimize paradigm
Why AInventory first predicts demand and then optimizes the order decision, instead of applying a single formula to every product.
The problem with the one-size-fits-all formula
Traditional inventory management relies on fixed rules: a reorder point calculated from average demand, an assumed standard deviation, and a standard safety-stock factor. That approach carries a silent assumption: that all products share the same statistical behavior. In practice, a steady-selling product and a highly variable seasonal product do not have the same uncertainty signature, and treating them identically produces surplus on some SKUs and stockouts on others.
AInventory breaks that assumption by implementing the predict-then-optimize paradigm: first it projects future demand while respecting each SKU's own behavior, then it translates that projection into an order decision that balances the cost of running short against the cost of holding excess inventory.
The three-step flow
1. Listen to the product
Before forecasting, the engine analyzes the historical forecast-error record of each SKU and identifies which statistical pattern best describes it. Each product has its own uncertainty signature: the distribution of its past errors reveals whether demand is stable, intermittent, seasonal, skewed, or heavy-tailed. This step does not impose a model from the outside; it lets the time series reveal its own nature.
2. Simulate the future
With the pattern identified, the engine generates thousands of demand scenarios consistent with that SKU's real uncertainty. This simulation replaces the rigidity of a closed-form formula (which assumes a normal distribution and fixed parameters) with a flexible representation of what could happen. The result is a cloud of possible futures, not a single point estimate.
For the technical detail on how these scenarios are constructed, see Monte Carlo Simulation.
3. Choose the best decision
Across those same simulated scenarios, the engine evaluates different replenishment policies and selects the one that minimizes total expected cost, weighing the cost of surplus against the cost of shortage. The final decision comes with a full justification: which statistical pattern was identified, which policy won, and why it outperformed the alternatives.
To understand the available policies and how they are evaluated, see Replenishment policies.
Why separating the two steps matters
Separating prediction from optimization has a key advantage: each step can improve independently. A more accurate forecast model feeds the same optimization logic and yields better orders without changing the structure. And if business conditions change (a new storage cost, a shift in the target service level), the optimization policy recalibrates without needing to retrain the demand model.
When a Demand Planner manually adjusts a forecast, that adjustment enters the optimization step with the same logic. The system always optimizes over whatever demand is presented to it, whether model-generated or manually modified by the team.
Relationship with the glossary
The terms predict-then-optimize, uncertainty signature, replenishment policy, and shortage cost have entries in the bilingual glossary.