Distinguishing Descriptive, Diagnostic, and Prescriptive Problems
You'll learn how to classify data science problems into four fundamental types based on what question they answer.
Before diving into modeling, you need to identify what kind of problem you're solving. Data science problems fall into distinct categories that require different approaches:
The Four Problem Types
Descriptive problems answer "What happened?" They summarize past data without explaining why. Example: "Sales dropped 15% last quarter" or "Our website had 50,000 visitors this month." These use statistics, dashboards, and visualizations to characterize patterns.
Diagnostic problems answer "Why did it happen?" They dig into causes and relationships. Example: "Sales dropped because our main competitor launched a promotion" or "Traffic increased due to our email campaign." These require correlation analysis, hypothesis testing, and causal investigation.
Predictive problems answer "What will happen?" They forecast future outcomes based on patterns. Example: "We expect 20% customer churn next month" or "This patient has high risk of readmission." These use machine learning models trained on historical data.
Prescriptive problems answer "What should we do?" They recommend specific actions. Example: "Offer a 10% discount to these at-risk customers" or "Stock 500 units of Product X next week." These combine predictions with optimization and decision frameworks.
Why This Matters
Misclassifying your problem leads to wasted effort. If stakeholders want to know why sales fell (diagnostic), delivering a prediction model (predictive) won't satisfy them. If they need a recommendation (prescriptive), just describing the current state (descriptive) leaves them without actionable guidance.
Think of it like a doctor: describing symptoms, diagnosing the disease, predicting recovery time, and prescribing treatment are four separate (though related) tasks requiring different skills.
Key Takeaway: Clearly classify whether your problem needs description, explanation, prediction, or recommendation—each requires distinct analytical approaches and delivers different value to stakeholders.