Prescriptive Analytics

Prescriptive analytics recommends actions, such as prices, customer contacts, stock orders, or work schedules. A defensible recommendation needs an objective, constraints, and a valid model of the consequences of available actions.

An objective someone has stated

Maximizing revenue, margin, conversion, retention, or lifetime value will produce different recommendations from identical data. There is no neutral choice, and no analysis can make it — it is a business decision about what the organization is trying to do.

A system that recommends without a declared objective has one regardless, chosen implicitly by whoever implemented the scoring. That is the common case, and it surfaces later as an argument about why the model keeps suggesting discounts: it was optimizing conversion, and nobody said margin mattered.

Where several goals compete, the honest form is a stated trade-off — maximize margin subject to not letting conversion fall below a floor — rather than a weighted blend whose weights nobody can interpret.

Constraints that reflect reality

A recommendation the organization cannot execute is not an answer. Constraints worth writing down explicitly include budget and capacity, contractual and regulatory limits, fairness requirements, and the plain question of what the people receiving the recommendation are able and willing to do.

For example, a proposed route that ignores a field engineer’s shift or a local access restriction may be impossible to follow. Record overrides and their reasons so reported gains reflect what was actually executed.

A valid model of action outcomes

Some decisions use known physical or contractual relationships. An optimizer can allocate work against specified costs and capacities without a separate statistical study of causal effects; the Google OR-Tools linear-programming example illustrates objectives and constraints of this kind. When a decision depends on how an intervention changes human outcomes, causal identification matters. Recommending a discount for its effect on purchases requires more than observing what selected discounted customers did.

Hernán puts the methodological asymmetry plainly: “Automatic variable selection procedures may work for prediction, but not necessarily for causal inference.” A model can be highly accurate at predicting churn because it learned that customers who contact support three times are about to leave, and the action read out of it — reduce support contacts — is worse than doing nothing.

For that intervention question, estimate effects using experiments when appropriate or causal inference with stated assumptions. Prediction can supply useful inputs, but the allocation decision also depends on effect, value, cost, and constraints. Check the validity of the outcome model appropriate to the task rather than requiring every optimization problem to begin with a causal experiment.

When to automate, and what automation does to your data

Numerous, individually small, reversible decisions can be candidates for automation, but these properties are not sufficient. Check authority, legal obligations, cumulative consequences, monitoring, exception handling, and recovery. For decisions requiring human judgment, provide a recommendation with its reasoning and give the reviewer enough time, evidence, and authority to act.

When recommendations drive actions and the resulting observations feed future training, the data reflects the policy’s choices. This feedback can limit what is learned about alternatives, and it can occur with human decisions as well as automation. Account for the selection process when evaluating or retraining the system.

Where safe, lawful, and feasible, randomized holdouts or experiments can estimate policy effects after launch. Otherwise use an evaluation design with explicit assumptions and limits, alongside operational monitoring. Effects established in one setting may not transfer to another; Kohavi and colleagues’ web-experiment findings illustrate why the conditions of the evaluation matter.

In a randomized policy evaluation, compare outcomes by grouping all units according to their original randomized assignment, including those who did not follow the recommendation: the intention-to-treat comparison. Track uptake separately as a process measure. Comparing only followers with the holdout selects on behavior after assignment and can introduce bias. Claims about treatment received or effects among compliers need their own estimand, assumptions, and analysis. Model accuracy alone does not show whether the policy improves outcomes.

How prescription relates to descriptive, diagnostic, and predictive questions is worked through in What Kind of Question Is It?.

References:

• Hernán, The C-Word, American Journal of Public Health 108(5), 2018
• Kohavi, Deng, Longbotham & Xu, Seven Rules of Thumb for Web Site Experimenters, KDD 2014
• Hyndman & Athanasopoulos, Forecasting: Principles and Practice (3rd ed.), §1.1
• Google OR-Tools, Solving a Linear Programming Problem