What is prescriptive analytics?
Prescriptive analytics is the stage of analysis that goes beyond forecasting an outcome to recommending what to do about it. It takes the predictions from predictive analytics, adds the business’s objectives and constraints, and uses optimization, simulation, or rules to find the action that scores best. A demand forecast says a warehouse will need 1,200 units next month; prescriptive logic says order 900 now from the primary supplier and 300 from the backup, given lead times, truck capacity, and the cash the finance team has released.
The output is a decision or a ranked set of options, often with the expected result of each. Scenario planning tools that let a manager test what happens if a supplier is a week late belong to the same family.
How prescriptive analytics works
Three ingredients are needed. The first is a model of the system, which may be a forecast, a set of equations, or a simulation. The second is an objective, such as minimizing total cost or maximizing on-time deliveries. The third is a set of constraints that encode the rules the business will not break, such as no overtime beyond a limit and no schedule that violates a maintenance window. An optimization solver then searches the allowed options for the best one. Monte Carlo simulation is often used to test how the recommended action holds up when the predictions turn out to be wrong.
Prescriptive analytics in practice
A building products distributor with eleven branches runs a nightly replenishment recommendation. The forecast model predicts demand per item per branch. The prescriptive layer then decides transfer quantities between branches and purchase quantities from suppliers, subject to truck capacity, minimum order quantities, and a working capital ceiling. A buyer reviews the recommendations each morning, accepts most, and overrides a few based on information the model does not have, such as a contractor’s verbal commitment to a large job. The override rate is tracked, and a rising rate is the signal to revisit the model.
What to watch for
- Constraints are where the real business knowledge lives, and a recommendation that ignores an unwritten rule will be rejected on the first day.
- Keep a human approval step wherever the recommended action changes a system of record, at least until the override rate shows the logic is trusted.
- Prescriptive output inherits the errors of the predictions beneath it, so test recommendations under pessimistic forecasts.
- Ask whether the objective function reflects what the business actually wants, since optimizing cost alone will quietly sacrifice service.