Prescriptive Analytics
Also known as: decision optimization, optimization analytics, operations research, decision support analytics, prescriptive modeling
Prescriptive analytics recommends specific actions to achieve desired outcomes, going beyond descriptive analytics (what happened), diagnostic analytics (why it happened), and predictive analytics (what will happen) to answer 'what should we do?' Prescriptive methods include optimization (finding the mathematically optimal allocation of resources given constraints), simulation (modeling the consequences of different action choices), and decision trees (structuring multi-step decision problems with probabilistic outcomes).
Prescriptive analytics is embedded in many operational systems: GPS navigation recommending the optimal route, supply chain optimization allocating inventory across distribution centers, and pricing algorithms recommending real-time prices for airline seats.
Prescriptive Analytics, explained at three levels
At home
Prescriptive analytics doesn't just tell you what will happen -- it tells you what to do about it. GPS navigation is prescriptive: it analyzes your destination, current traffic (predictive), and possible routes, then recommends the specific route that will get you there fastest (prescriptive). A personal finance app that tells you your account balance (descriptive) and predicts you'll be short on rent next month (predictive) becomes prescriptive when it recommends which specific expenses to cut to ensure you make rent. Prescriptive analytics is the step from insight to action recommendation.
At work
Prescriptive analytics recommends specific business actions: an airline pricing system recommends exactly what price to charge for each remaining seat on each flight (dynamic pricing optimization); a supply chain system recommends exactly how much inventory to order for each product at each location (inventory optimization); a marketing system recommends which specific promotion to send to which specific customer (personalization at scale). These systems embed business rules, cost structures, and prediction models to recommend actions that maximize business objectives (revenue, margin, service level) subject to operational constraints.
At home
The core prescriptive analytics methods: linear programming (LP) -- finding the best allocation of limited resources when the objective and constraints are linear (maximize revenue from a product mix given production capacity); integer programming -- same as LP but with discrete (whole number) decisions; simulation -- running many scenarios under different assumptions to understand the distribution of outcomes under different action choices; reinforcement learning -- an AI approach where a system learns to take actions through trial and error, optimizing for a reward signal. Most business prescriptive analytics uses LP/integer programming or simple decision rules derived from predictive model outputs.
At work
Prescriptive analytics in workforce scheduling: hospital staffing, call center workforce management, and retail scheduling all face the same problem -- how many people of which skill types to schedule across which time slots to meet predicted demand while minimizing labor cost. This is an integer programming problem: the decision variables are staff schedule assignments, the objective is minimize cost, and the constraints include demand coverage requirements, labor regulations (maximum consecutive hours, minimum rest time), staff skill availability, and fairness requirements (equitable distribution of desirable vs. undesirable shifts). Purpose-built workforce optimization software (Kronos, NICE, Verint) solves these problems with custom OR engines.
At home
Stochastic optimization vs. deterministic optimization: deterministic optimization assumes all inputs are known with certainty (demand, cost, availability). Real decisions involve uncertainty: demand is a forecast (with error), supply availability is uncertain, competitor actions are unknown. Stochastic optimization incorporates uncertainty explicitly -- optimizing over the distribution of possible outcomes rather than a single point estimate. Stochastic programming and robust optimization are mathematical frameworks that find solutions that perform well across a range of scenarios rather than optimizing for a single assumed future. These methods are appropriate when the variability in outcomes is high and decisions cannot be easily reversed.
At work
Combining predictive and prescriptive analytics in closed-loop systems: the most powerful analytical systems combine prediction and prescription in a continuous feedback loop. A dynamic pricing system predicts demand at each possible price (predictive model), then optimizes the price to maximize expected revenue (prescriptive optimization), then observes actual demand at the chosen price (outcome data), then retrains the predictive model with this new data point (learning). This closed loop means the system continuously improves as it accumulates more data about how demand responds to different prices. Reinforcement learning formalizes this feedback loop mathematically, enabling systems that learn optimal decision policies from experience without explicit human specification of the optimal policy.
Real-world examples
At home
Marcus was planning his personal investment allocations across retirement accounts. He input his constraints: target retirement date 25 years away, risk tolerance (medium), current account balances, and monthly contribution amounts. A financial planning tool used optimization to recommend an exact allocation across stock funds, bond funds, and international funds that maximized expected returns for his risk level over 25 years. The tool also simulated 1,000 market scenarios to show the probability distribution of portfolio values at retirement -- giving Marcus both a recommended action (specific allocation) and an understanding of the range of outcomes.
At work
Nadia ran operations for a national courier company. Her team implemented a route optimization system: every morning, the system ingested that day's delivery orders, truck availability, driver schedules, and real-time traffic data. The optimization engine solved a vehicle routing problem -- assigning deliveries to specific trucks and sequencing stops to minimize total fuel cost and delivery time while meeting time-window requirements for each delivery. The system generated complete route plans for each truck in minutes. The optimization reduced fuel costs by 12% and improved on-time delivery rates compared to manually-planned routes.
Common confusions
- Many people assume prescriptive analytics is the 'highest level' of analytics and should always be the goal — Prescriptive analytics is only valuable when the business has the operational infrastructure to act on its recommendations at the required speed and scale. A recommendation engine that recommends personalized prices for 10 million customers in real-time only creates value if the e-commerce platform can actually implement those prices for each customer session. If the organization cannot operationalize the recommendations -- because of system limitations, organizational culture, or implementation cost -- prescriptive analytics creates no value regardless of its analytical sophistication. Descriptive analytics that drives better human decisions is often more valuable than prescriptive analytics that the organization cannot operationalize.
- Some assume optimization always finds the globally best solution, making the answer unambiguous — Optimization finds the best solution given the objective function and constraints that were specified -- but the specification of objectives and constraints is a human design choice that significantly affects the output. Different objective functions (maximize profit vs. maximize revenue vs. maximize customer satisfaction) produce different optimal solutions for the same problem. Constraints encode business rules and values choices (maximum overtime hours, minimum service levels for certain customer types) that affect what the optimizer recommends. The optimizer is only as good as the problem formulation: a technically optimal solution to a wrongly-specified problem can be operationally disastrous.
For analysts
Prescriptive analytics engagements are typically more complex and higher-value than descriptive or predictive projects because they require not just analysis but a formal model of the decision problem: what are we optimizing? What are the constraints? What are the decision variables we can control? For consulting practice, the most common prescriptive analytics applications are pricing optimization (what price maximizes revenue or margin?), resource allocation (how to distribute budget, staff, or inventory across competing uses?), and process scheduling (what sequence of operations minimizes time or cost?). The practical starting point is decision structuring -- before any optimization, mapping the decision precisely: what can be controlled, what are the objectives, and what are the constraints.
History & origin
Prescriptive analytics has roots in operations research (OR), which emerged as a formal discipline during World War II (optimizing logistics, scheduling, and resource allocation for military operations). Linear programming (George Dantzig, 1947) provided a foundational optimization method. Post-war, operations research techniques spread to industrial applications: production scheduling, transportation routing, supply chain optimization, financial portfolio optimization (Harry Markowitz, 1952). Decision analysis (Howard Raiffa, 1968) formalized how to structure multi-stage decisions under uncertainty. The term 'prescriptive analytics' emerged in the 2010s as a complement to descriptive and predictive analytics, popularized by consulting firms and analytics platform vendors. Modern prescriptive systems often combine ML-based predictive models (forecasting demand, predicting failures) with optimization engines (recommending resource allocations given the forecasts).