Chapter 5. The Four Pillars of Analytics: Descriptive to Prescriptive

Analytics is not a single activity—it's a spectrum of approaches, each answering different questions and requiring different levels of sophistication. Understanding this spectrum is crucial because the type of analytics you choose determines the value you deliver .

Many organizations get stuck at the descriptive level ("What happened?") when the real business value lies in predictive and prescriptive analytics ("What will happen?" and "What should we do?"). As an analyst, your job is to recognize which type of analytics a problem requires and guide stakeholders toward the most valuable approach.

This chapter introduces the four pillars of analytics :

1. Descriptive Analytics : What happened?
2. Diagnostic Analytics : Why did it happen?
3. Predictive Analytics : What will happen?
4. Prescriptive Analytics : What should we do?

Each pillar builds on the previous one, increasing in complexity and business value.

5.1 The Analytics Spectrum: Descriptive, Diagnostic, Predictive, Prescriptive

The Four Questions

Every business problem can be framed as one of four questions:

The progression of value:

• Descriptive  analytics tells you there's a problem
• Diagnostic  analytics tells you what caused it
• Predictive  analytics tells you what's coming
• Prescriptive  analytics tells you what to do about it

Most organizations spend 80% of their analytics effort on descriptive analytics, which delivers the least value. The goal is to move up the spectrum.

Characteristics of Each Type

1. Descriptive Analytics

• Focus : Historical data, reporting, summarization
• Techniques : Aggregation, visualization, dashboards, KPIs
• Output : "Revenue was $2.3M last month"
• Skill level : Basic (SQL, Excel, BI tools)
• Time orientation : Past

2. Diagnostic Analytics

• Focus : Understanding causes and relationships
• Techniques : Drill-down, segmentation, correlation, root cause analysis
• Output : "Revenue dropped because customer acquisition cost increased 40%"
• Skill level : Intermediate (statistical thinking, business knowledge)
• Time orientation : Past

3. Predictive Analytics

• Focus : Forecasting future outcomes
• Techniques : Regression, time series, machine learning, probability models
• Output : "Revenue will likely be $2.1M next month (±$150K)"
• Skill level : Advanced (statistics, modeling, programming)
• Time orientation : Future

4. Prescriptive Analytics

• Focus : Recommending optimal actions
• Techniques : Optimization, simulation, decision analysis, A/B testing
• Output : "To maximize revenue, allocate 60% of budget to Channel A, 30% to Channel B, 10% to Channel C"
• Skill level : Expert (optimization, domain expertise, business judgment)
• Time orientation : Future + Action

The Maturity Curve

Organizations typically evolve through these stages:

Stage 1: Reactive (Descriptive)

• "What were last month's sales?"
• Reporting is manual, backward-looking
• Decisions are based on intuition + basic reports

Stage 2: Investigative (Diagnostic)

• "Why did sales drop?"
• Analysts dig into data to find causes
• Decisions are based on understanding patterns

Stage 3: Anticipatory (Predictive)

• "What will sales be next month?"
• Models forecast future outcomes
• Decisions are proactive, not reactive

Stage 4: Optimized (Prescriptive)

• "How should we allocate resources to maximize sales?"
• Systems recommend optimal actions
• Decisions are data-driven and automated where possible

Your role as an analyst : Help your organization move up this curve. Don't just answer "what happened?"—push toward "what should we do?"

5.2 Descriptive Analytics

Descriptive analytics is the foundation. It answers "What happened?"  by summarizing historical data into understandable formats.

5.2.1 Summaries, Dashboards, and Basic Reporting

Purpose : Make data accessible and understandable to decision-makers.

Common outputs:

• Reports : Monthly sales reports, quarterly performance reviews
• Dashboards : Real-time KPI tracking (revenue, conversion rates, customer counts)
• Scorecards : Performance against targets (actual vs. budget)

Key principles for effective descriptive analytics:

1. Know your audience

• Executives need high-level summaries (one number, one chart)
• Managers need operational details (daily metrics, trends)
• Analysts need granular data (raw numbers, ability to drill down)

2. Choose the right visualization

• Trends over time : Line charts
• Comparisons : Bar charts
• Proportions : Pie charts (use sparingly) or stacked bars
• Distributions : Histograms
• Relationships : Scatter plots

3. Highlight what matters

• Use color to draw attention to problems or opportunities
• Show variance from targets or previous periods
• Include context (is 5% growth good or bad?)

4. Make it actionable

• Don't just show numbers—show what's changed and why it matters
• Include thresholds or targets
• Link metrics to business goals

Example: Sales Dashboard

A good sales dashboard might include:

Top-level metrics (for executives):

• Total revenue: $2.3M (↓ 8% vs. last month)
• New customers: 450 (↑ 12% vs. last month)
• Average deal size: $5,111 (↓ 18% vs. last month)

Trend analysis (for managers):

• Revenue by month (last 12 months) → shows seasonal patterns
• Revenue by product line → shows which products are growing/declining
• Revenue by region → shows geographic performance

Operational details (for analysts):

• Conversion rates by channel
• Sales cycle length by segment
• Win/loss rates by competitor

The mistake most people make : Creating dashboards with 50 metrics that no one looks at. Better approach : 5-7 key metrics that drive decisions, with the ability to drill down for details.

5.2.2 Common Descriptive Techniques (Aggregation, Grouping, Cohort Analysis)

1. Aggregation

Summarizing data at different levels of granularity.

Examples:

• Daily sales → Weekly sales → Monthly sales
• Individual transactions → Customer lifetime value
• Product-level revenue → Category-level revenue

When to use : When raw data is too detailed to be useful.

Caution : Aggregation hides variation. Always check if the average is representative or if there are important subgroups.

2. Grouping and Segmentation

Breaking data into meaningful categories.

Examples:

• Customers by region, industry, size
• Products by category, price tier, margin
• Transactions by channel, time of day, payment method

Why it matters : Averages lie. "Average customer spends $100" might hide that:

• Enterprise customers spend $10,000
• Small businesses spend $500
• Individuals spend $50

These three groups need completely different strategies.

3. Cohort Analysis

Tracking groups of users/customers who share a common characteristic over time.

Example: Customer Retention by Signup Month

Insights:

• Retention is improving for newer cohorts (70% vs. 65% in Month 1)
• About 50% of customers churn within 3 months
• Long-term retention (12 months) is around 28%

Why cohort analysis is powerful : It separates growth from retention. You might see "total active users" growing, but cohort analysis reveals that you're losing customers as fast as you acquire them.

Common applications:

• Customer retention and churn
• Product adoption and engagement
• Marketing campaign effectiveness over time

Prompt to AI for Cohort Analysis:

I have a dataset with customer_id, signup_date, and activity_date.

Create a cohort retention analysis showing:

1. Cohorts by signup month

2. Retention rate for each cohort at 1, 3, 6, 12 months

3. Visualization of retention curves

5.3 Diagnostic Analytics

Diagnostic analytics answers "Why did it happen?"  It goes beyond reporting to uncover causes and relationships.

5.3.1 Root Cause Analysis and Drill-Down Techniques

The Problem with Descriptive Analytics Alone:

"Sales decreased 15% last quarter."

This tells you there's a problem, but not what to do about it. You need to understand why .

Root Cause Analysis Framework

Step 1: Decompose the metric

Break down the high-level metric into its components.

Example: Revenue Decomposition

Revenue = Customers × Average Order Value

If revenue is down, is it because:

- Fewer customers? (acquisition problem)

- Lower order value? (pricing or product mix problem)

Further decomposition:

Customers = Traffic × Conversion Rate

Average Order Value = Items per Order × Price per Item

Now you have specific areas to investigate.

Step 2: Segment the data

Look at the metric across different dimensions:

• Time : When did the change occur? (specific week, month, season)
• Geography : Which regions are affected?
• Product : Which products or categories?
• Customer : Which customer segments?
• Channel : Which acquisition or sales channels?

Example:

"Sales decreased 15% overall, but:

• Region A: ↓ 30%
• Region B: ↓ 5%
• Region C: ↑ 10%

This is a Region A problem, not a company-wide problem."

Step 3: Look for correlations and patterns

What else changed at the same time?

• Did we change pricing?
• Did a competitor launch a new product?
• Was there a marketing campaign change?
• Did we have operational issues (stockouts, shipping delays)?

Caution : Correlation ≠ causation. Just because two things happened at the same time doesn't mean one caused the other. But correlations give you hypotheses to test.

Step 4: Form and test hypotheses

Based on your analysis, form specific hypotheses:

Hypothesis : "Sales decreased in Region A because our main competitor lowered prices by 20%."

Test :

• Check competitor pricing data
• Look at win/loss rates against that competitor
• Survey lost customers
• Analyze price sensitivity in Region A vs. other regions

Step 5: Validate with additional data

Don't rely on a single data source. Triangulate:

• Quantitative data (sales numbers, web analytics)
• Qualitative data (customer feedback, sales team input)
• External data (market trends, competitor actions)

The "5 Whys" Technique

Keep asking "why" until you reach the root cause.

Example:

1. Why did revenue decrease?  → Fewer customers
2. Why fewer customers?  → Lower conversion rate
3. Why lower conversion rate?  → Higher bounce rate on product pages
4. Why higher bounce rate?  → Page load time increased from 2s to 8s
5. Why did load time increase?  → New image gallery feature wasn't optimized

Root cause : Unoptimized image gallery feature.

Action : Optimize images or remove feature.

5.3.2 Attribution Analysis in Marketing and Operations

Attribution  answers: "Which actions or channels contributed to this outcome?"

This is critical in marketing (which campaigns drove sales?) and operations (which factors caused the delay?).

Marketing Attribution Example

A customer's journey before purchase:

1. Saw Facebook ad (Day 1)
2. Clicked Google search ad (Day 5)
3. Visited website directly (Day 7)
4. Received email (Day 10)
5. Clicked email and purchased (Day 12)

Question : Which channel gets credit for the sale?

Attribution Models:

1. Last-touch attribution

• Email gets 100% credit
• Problem : Ignores all the earlier touchpoints that built awareness

2. First-touch attribution

• Facebook gets 100% credit
• Problem : Ignores the touchpoints that closed the sale

3. Linear attribution

• Each touchpoint gets 20% credit
• Problem : Treats all touchpoints as equally important (they're not)

4. Time-decay attribution

• More recent touchpoints get more credit
• Email: 40%, Direct: 30%, Google: 20%, Facebook: 10%
• Better : Recognizes that closing touchpoints matter more

5. Data-driven attribution

• Use statistical models to determine actual contribution of each channel
• Requires significant data and sophisticated analysis
• Best : Based on actual behavior, not assumptions

Why attribution matters:

Without proper attribution, you might:

• Over-invest in last-touch channels (email, direct) that get credit but don't drive new customers
• Under-invest in awareness channels (social, display) that don't get credit but are essential
• Make decisions based on incomplete information

The reality : Most conversions involve multiple touchpoints. Understanding the full customer journey is essential for optimizing marketing spend.

Operations Attribution Example

Problem : Project was delivered 3 weeks late.

Attribution analysis : Which factors contributed to the delay?

Insight : Scope creep is the biggest issue. Future projects need better scope management and change control processes.

5.4 Predictive Analytics

Predictive analytics answers "What will happen?"  It uses historical data to forecast future outcomes.

5.4.1 Forecasting Outcomes and Probabilities

The shift from descriptive to predictive:

• Descriptive : "We sold 10,000 units last month"
• Predictive : "We'll sell 11,500 units next month (±1,000)"

Why prediction matters:

• Planning : How much inventory to stock, how many staff to hire
• Budgeting : Revenue and cost forecasts
• Risk management : Probability of customer churn, equipment failure, project delays
• Opportunity identification : Which customers are likely to buy, which leads to prioritize

Types of Predictions

1. Point forecasts

• Single number: "Sales will be $2.3M next month"
• Problem : Doesn't capture uncertainty

2. Interval forecasts

• Range: "Sales will be $2.1M to $2.5M next month (95% confidence)"
• Better : Communicates uncertainty

3. Probability forecasts

• Distribution: "30% chance sales exceed $2.5M, 50% chance between $2.1M-$2.5M, 20% chance below $2.1M"
• Best : Enables risk-based decision making

Common Forecasting Techniques

1. Time Series Forecasting

Use historical patterns to predict future values.

Techniques:

• Moving averages : Simple, good for stable trends
• Exponential smoothing : Weights recent data more heavily
• ARIMA : Captures complex patterns (trend, seasonality, autocorrelation)
• Prophet  (Facebook): Handles seasonality and holidays automatically

When to use : When you have regular time-series data (daily sales, monthly revenue, hourly traffic)

Example : Forecasting next quarter's revenue based on last 3 years of quarterly data.

2. Regression Models

Predict an outcome based on relationships with other variables.

Example : Predict house price based on:

• Square footage
• Number of bedrooms
• Location
• Age of house

When to use : When you have explanatory variables that influence the outcome.

3. Classification Models

Predict which category something belongs to.

Examples:

• Will this customer churn? (Yes/No)
• Will this loan default? (Yes/No)
• Which product will this customer buy? (Product A/B/C)

Techniques:

• Logistic regression
• Decision trees
• Random forests
• Neural networks

When to use : When the outcome is categorical, not continuous.

4. Machine Learning Models

More sophisticated techniques that can capture complex, non-linear relationships.

Examples:

• Gradient boosting (XGBoost, LightGBM)
• Neural networks
• Ensemble methods

When to use : When you have large datasets and complex relationships that simpler models can't capture.

Caution : More complex ≠ better. Start simple, add complexity only if needed.

Evaluating Forecast Accuracy

How do you know if your forecast is good?

Common metrics:

1. Mean Absolute Error (MAE)

• Average absolute difference between forecast and actual
• Easy to interpret: "On average, we're off by $50K"

2. Mean Absolute Percentage Error (MAPE)

• Average percentage error
• Good for comparing across different scales: "On average, we're off by 8%"

3. Root Mean Squared Error (RMSE)

• Penalizes large errors more heavily
• Useful when big misses are particularly costly

The key question : Is the forecast accurate enough for the decision you need to make?

• Forecasting next year's revenue for budgeting: ±10% might be fine
• Forecasting tomorrow's demand for perishable inventory: ±5% might be necessary

5.4.2 From Explanatory to Predictive Modeling

Explanatory models  help you understand relationships:

• "How does price affect demand?"
• "What factors influence customer churn?"

Predictive models  help you forecast outcomes:

• "What will demand be next month?"
• "Which customers will churn?"

Key differences:

Example: Customer Churn

Explanatory approach:  "We want to understand what causes churn."

• Use logistic regression
• Interpret coefficients: "Each additional support ticket increases churn probability by 15%"
• Focus on statistical significance and causation
• Output : Understanding of drivers

Predictive approach:  "We want to identify which customers will churn next month."

• Try multiple models (logistic regression, random forest, XGBoost)
• Choose the one with best out-of-sample accuracy
• Don't worry if it's a "black box"
• Output : List of high-risk customers to target with retention campaigns

The Prediction Workflow

1. Define the prediction target

• What exactly are you predicting?
• What time horizon? (next week, next month, next year)
• What level of accuracy is needed?

2. Gather and prepare data

• Historical data on the outcome
• Potential predictors (features)
• Clean, transform, handle missing values

3. Split data

• Training set : Build the model (typically 70-80%)
• Validation set : Tune the model (10-15%)
• Test set : Final evaluation (10-15%)

Never  evaluate on the same data you trained on—that's cheating!

4. Build and compare models

• Start simple (linear regression, logistic regression)
• Try more complex models if needed
• Compare performance on validation set

5. Evaluate on test set

• Final check: How well does it perform on completely unseen data?
• This is your honest estimate of real-world performance

6. Deploy and monitor

• Put the model into production
• Track actual performance
• Retrain periodically as patterns change

Common Pitfalls in Predictive Modeling

1. Overfitting

• Model is too complex, memorizes training data
• Performs great on training data, terrible on new data
• Solution : Use simpler models, regularization, cross-validation

2. Data leakage

• Using information that wouldn't be available at prediction time
• Example: Predicting customer churn using "cancellation date" as a feature
• Solution : Carefully think through what data would actually be available

3. Ignoring business context

• Building a technically accurate model that's useless in practice
• Example: Churn model that requires data not available until after customer churns
• Solution : Involve business stakeholders throughout

4. Not updating models

• Patterns change over time (customer behavior, market conditions)
• Yesterday's model becomes less accurate
• Solution : Monitor performance, retrain regularly

5.5 Prescriptive Analytics

Prescriptive analytics answers "What should we do?"  It goes beyond prediction to recommend optimal actions.

This is the highest value form of analytics, but also the most complex.

5.5.1 Optimization Models for Decision Support

Optimization  finds the best solution among many possibilities, subject to constraints.

Structure of an optimization problem:

1. Decision variables : What can you control? (prices, quantities, schedules, allocations)
2. Objective function : What are you trying to maximize or minimize? (profit, cost, time, risk)
3. Constraints : What are the limits? (budget, capacity, regulations, physical limits)

Example 1: Product Mix Optimization

Scenario : A factory produces three products. How many of each should you make to maximize profit?

Decision variables:

• x₁ = units of Product A
• x₂ = units of Product B
• x₃ = units of Product C

Objective function  (maximize):

Profit = 50x₁ + 40x₂ + 60x₃

Constraints:

Labor hours:     2x₁ + 3x₂ + 4x₃ ≤ 1000 hours available

Machine time:    4x₁ + 2x₂ + 3x₃ ≤ 1200 hours available

Raw materials:   3x₁ + 3x₂ + 2x₃ ≤ 900 units available

Non-negativity:  x₁, x₂, x₃ ≥ 0

Solution  (using linear programming):

• Product A: 150 units
• Product B: 100 units
• Product C: 125 units
• Maximum profit: $22,500

Business value : Instead of guessing or using rules of thumb, you have the mathematically optimal production plan.

Example 2: Marketing Budget Allocation

Scenario : You have $100K marketing budget to allocate across 4 channels. Each channel has different ROI and diminishing returns.

Decision variables:

• Budget allocated to each channel

Objective function  (maximize):

• Total conversions (or revenue, or profit)

Constraints:

• Total budget ≤ $100K
• Minimum spend per channel (to maintain presence)
• Maximum spend per channel (due to capacity or diminishing returns)

Output : Optimal allocation that maximizes conversions given your budget and constraints.

Types of Optimization Problems

1. Linear Programming

• Objective and constraints are linear
• Very efficient to solve, even with thousands of variables
• Applications : Production planning, resource allocation, transportation

2. Integer Programming

• Some variables must be whole numbers
• Example: Number of employees to hire (can't hire 2.7 people)
• Applications : Scheduling, facility location, project selection

3. Nonlinear Programming

• Objective or constraints are nonlinear
• More complex to solve
• Applications : Portfolio optimization, pricing with demand curves

4. Multi-objective Optimization

• Multiple competing objectives (maximize profit AND minimize risk)
• Find trade-offs (Pareto frontier)
• Applications : Portfolio management, product design

When to Use Optimization

Optimization is valuable when:

• You have many possible choices
• The best choice isn't obvious
• The stakes are high (significant money, resources, or risk)
• You need to make the decision repeatedly

Examples:

• ✓ Allocating $10M marketing budget across 20 channels
• ✓ Scheduling 500 employees across 50 shifts
• ✓ Routing 100 delivery trucks to 1,000 locations
• ✗ Deciding whether to launch one new product (use judgment, not optimization)

5.5.2 Simulation and Scenario Planning

Simulation  models complex systems to understand how they behave under different conditions.

Why simulation?

Some problems are too complex for analytical solutions:

• Too many variables
• Too much uncertainty
• Complex interactions and feedback loops

Solution : Build a model of the system, run it thousands of times with different inputs, and see what happens.

Monte Carlo Simulation

Run a model many times with random inputs drawn from probability distributions.

Example: Project Cost Estimation

A project has three phases:

• Phase 1: $50K ± $10K (normal distribution)
• Phase 2: $80K ± $20K (normal distribution)
• Phase 3: $40K ± $15K (normal distribution)

Question : What's the total project cost? What's the probability it exceeds $200K?

Analytical approach :

• Expected cost = $50K + $80K + $40K = $170K
• But this doesn't tell you the range or probability of overruns

Simulation approach :

• Run 10,000 simulations
• Each simulation: randomly draw costs for each phase, sum them up
• Result: Distribution of total project costs

Output:

• Mean: $170K
• 50th percentile (median): $169K
• 90th percentile: $195K
• 95th percentile: $205K
• Probability of exceeding $200K: 18%

Business value : You can now say "We should budget $195K to have 90% confidence we won't overrun" instead of just "$170K."

Scenario Planning

Explore how different future scenarios would affect your business.

Structure:

1. Identify key uncertainties : What factors could significantly impact your business?
2. Define scenarios : Create 3-4 plausible future scenarios
3. Model impacts : How would each scenario affect your metrics?
4. Develop strategies : What actions would be robust across scenarios?

Example: Retail Expansion Decision

Key uncertainties:

• Economic growth (strong, moderate, weak)
• Competitive intensity (high, medium, low)

Scenarios:

For each scenario, model:

• Revenue projections
• Cost structure
• Profitability
• Cash flow

Strategic questions:

• Which expansion plan works best in most scenarios?
• What's our downside risk in the worst scenario?
• What early indicators would tell us which scenario is unfolding?

Discrete Event Simulation

Model systems where events happen at specific points in time.

Example: Call Center Staffing

• Calls arrive randomly (Poisson process)
• Call duration varies (exponential distribution)
• Agents handle calls one at a time
• Customers wait in queue if all agents busy

Questions:

• How many agents do we need to keep average wait time under 2 minutes?
• What's the trade-off between staffing cost and customer satisfaction?

Simulation approach:

• Model the call center as a queuing system
• Run simulations with different staffing levels
• Measure wait times, agent utilization, abandonment rates

Output : Optimal staffing level that balances cost and service quality.

5.6 Choosing the Appropriate Analytics Type for a Problem

How do you decide which type of analytics to use?

Ask these questions:

1. What decision needs to be made?

• No decision, just reporting  → Descriptive
• Need to understand what happened  → Diagnostic
• Need to anticipate what's coming  → Predictive
• Need to choose the best action  → Prescriptive

2. What's the business value?

• Low stakes  (routine reporting) → Descriptive is fine
• Medium stakes  (understanding problems) → Diagnostic or Predictive
• High stakes  (major resource allocation) → Prescriptive

3. What data and capabilities do you have?

• Limited data, basic tools  → Descriptive or Diagnostic
• Good historical data, statistical skills  → Predictive
• Rich data, optimization expertise  → Prescriptive

4. How much time do you have?

• Need answer today  → Descriptive (use existing reports)
• Have a week  → Diagnostic or simple Predictive
• Have a month  → Complex Predictive or Prescriptive

Decision Framework

START: What's the business question?

├─ "What happened?"

│   └─ DESCRIPTIVE ANALYTICS

│       • Dashboards, reports, summaries

│       • Quick, low effort

│

├─ "Why did it happen?"

│   └─ DIAGNOSTIC ANALYTICS

│       • Root cause analysis, segmentation

│       • Medium effort, requires business knowledge

│

├─ "What will happen?"

│   └─ PREDICTIVE ANALYTICS

│       • Forecasting, classification

│       • Higher effort, requires statistical skills

│

└─ "What should we do?"

    └─ PRESCRIPTIVE ANALYTICS

        • Optimization, simulation

        • Highest effort, highest value

Common Mistakes

1. Using descriptive analytics when you need predictive

❌ "Last year we sold 10,000 units, so let's plan for 10,000 this year"

✓ "Based on trend analysis and market conditions, we forecast 11,500 units (±1,000)"

2. Using predictive analytics when you need prescriptive

❌ "We predict 30% of customers will churn"

✓ "We predict 30% will churn. To reduce this to 20%, we should offer retention incentives to the 500 highest-risk customers, which will cost $50K but save $200K in lost revenue"

3. Using prescriptive analytics when you need diagnostic

❌ Building a complex optimization model before understanding the problem

✓ First diagnose why performance is poor, then optimize

4. Over-engineering

❌ Building a machine learning model when a simple report would suffice

✓ Start simple, add complexity only when needed

The Analyst's Judgment

The framework is a guide, not a rule. Sometimes you need multiple types:

Example: Sales Performance Problem

1. Descriptive : "Sales are down 15%" (identify the problem)
2. Diagnostic : "Sales are down because of pricing in Region A" (understand the cause)
3. Predictive : "If we don't act, we'll lose another 10% next quarter" (forecast impact)
4. Prescriptive : "Lower prices 8% in Region A and reallocate marketing budget" (recommend action)

Good analysts move fluidly between these types, using each where appropriate.

5.7 Case Examples Across the Four Pillars

Let's see how all four types of analytics apply to real business problems.

Case 1: E-commerce Customer Retention

Business Context : An e-commerce company notices customer retention is declining.

Descriptive Analytics: What happened?

Analysis:

• Overall retention rate: 65% (down from 72% last year)
• Cohort analysis shows retention declining for customers acquired in last 6 months
• Repeat purchase rate: 35% (down from 42%)

Output : Dashboard showing retention trends by cohort, product category, and acquisition channel.

Value : Confirms there's a problem and quantifies its magnitude.

Diagnostic Analytics: Why did it happen?

Analysis:

• Segment customers by behavior: High-value, Medium-value, Low-value
• High-value retention stable (85%)
• Medium-value retention dropped from 70% to 60%
• Low-value retention dropped from 50% to 35%

Drill-down:

• Medium and low-value customers are churning after first purchase
• Correlation with shipping times: Customers with >5 day shipping have 40% lower retention
• Competitor analysis: Main competitor now offers 2-day shipping

Root cause : Shipping times are too slow compared to competitors, especially affecting price-sensitive customers.

Value : Identifies specific cause and customer segments affected.

Predictive Analytics: What will happen?

Analysis:

• Build churn prediction model using:

• Customer demographics
• Purchase history
• Engagement metrics (email opens, site visits)
• Shipping experience

Output:

• Churn probability for each customer
• Identify 5,000 customers at high risk (>70% churn probability) in next 90 days
• Forecast: If no action taken, retention will drop to 58% next quarter, costing $2M in lost revenue

Value : Quantifies future impact and identifies specific customers to target.

Prescriptive Analytics: What should we do?

Analysis:

• Option 1 : Upgrade all shipping to 2-day ($500K cost)
• Option 2 : Offer 2-day shipping to high-risk customers only ($150K cost)
• Option 3 : Offer discount on next purchase to high-risk customers ($200K cost)

Optimization model:

• Maximize: Retained revenue - Cost of intervention
• Constraints: Budget limit, operational capacity

Simulation:

• Run scenarios with different combinations of interventions
• Model customer response probabilities

Recommendation:

• Offer 2-day shipping to 5,000 high-risk customers (cost: $150K)
• Offer 15% discount on next purchase to 2,000 highest-value at-risk customers (cost: $80K)
• Expected outcome : Retain 60% of at-risk customers, saving $1.2M in revenue
• Net benefit : $1.2M - $230K = $970K

Value : Provides specific, actionable recommendation with quantified ROI.

Case 2: Manufacturing Quality Control

Business Context : A manufacturer is experiencing increased defect rates.

Descriptive : Defect rate increased from 2% to 4.5% over last 3 months.

Diagnostic :

• Defects concentrated in Product Line B
• Correlation with new supplier for Component X
• Root cause: Component X from new supplier has higher failure rate

Predictive :

• If we continue with current supplier, defect rate will likely reach 6% within 2 months
• This will result in $500K in warranty costs and potential loss of major customer

Prescriptive :

• Option 1 : Switch back to original supplier (higher cost, but proven quality)
• Option 2 : Work with new supplier to improve quality (takes 3 months, uncertain outcome)
• Option 3 : Implement additional quality checks (adds cost and time)
• Recommendation : Switch back to original supplier immediately, negotiate long-term contract for better pricing

Case 3: Hospital Emergency Department

Business Context : ER wait times are increasing, patient satisfaction declining.

Descriptive :

• Average wait time: 85 minutes (up from 60 minutes)
• Patient satisfaction score: 3.2/5 (down from 4.1/5)
• Peak times: 6pm-10pm weekdays, all day weekends

Diagnostic :

• Bottleneck analysis: Triage is fast, but physician availability is limited
• Staffing analysis: Physician coverage doesn't match demand patterns
• 40% of patients could be handled by nurse practitioners, but only 2 NPs on staff

Predictive :

• Forecast patient arrivals by day/hour using 2 years of historical data
• Predict wait times under current staffing model
• Identify high-risk periods (>2 hour waits)

Prescriptive :

• Optimization model: Minimize wait times subject to staffing budget
• Simulation: Test different staffing configurations
• Recommendation :

• Add 2 nurse practitioners for evening shifts (cost: $200K/year)
• Adjust physician schedules to match demand patterns (no additional cost)
• Expected outcome : Reduce average wait time to 50 minutes, improve satisfaction to 4.3/5

Key Takeaways

1. Descriptive analytics  (What happened?) is necessary but not sufficient. Don't stop there.
   
2. Diagnostic analytics  (Why?) is where business knowledge matters most. Understanding causation requires domain expertise, not just data skills.
   
3. Predictive analytics  (What will happen?) enables proactive decision-making. The goal is not perfect prediction, but better decisions under uncertainty.
   
4. Prescriptive analytics  (What should we do?) delivers the highest value but requires the most sophistication. Start here for high-stakes decisions with many options.
   
5. Match the analytics to the decision . Don't over-engineer (complex models for simple problems) or under-engineer (simple reports for complex decisions).
   
6. Move up the value chain . Push your organization from reactive (descriptive) to proactive (predictive) to optimized (prescriptive).
   
7. Combine multiple types . Real problems often require descriptive → diagnostic → predictive → prescriptive in sequence.
   
8. Focus on action . The best analysis is worthless if it doesn't change a decision.
   

Exercises

Exercise 1: Classify Analytics Examples

For each scenario, identify whether it's primarily Descriptive, Diagnostic, Predictive, or Prescriptive  analytics:

a) A monthly sales report showing revenue by region and product category.

b) An analysis investigating why customer acquisition cost increased 40% last quarter.

c) A model that forecasts next quarter's demand for each product SKU.

d) A recommendation system that suggests which customers to target with a promotion to maximize ROI.

e) A dashboard showing real-time website traffic and conversion rates.

f) An analysis of which marketing channels contributed to conversions using multi-touch attribution.

g) A simulation that estimates project completion time under different resource allocation scenarios.

h) A cohort analysis showing retention rates for customers acquired in each month.

i) A churn prediction model that assigns a risk score to each customer.

j) An optimization model that determines the best product mix to maximize profit given production constraints.

Exercise 2: Design a Dashboard

Choose one of the following business functions and design a descriptive analytics dashboard:

Options:

• Sales performance
• Marketing campaign effectiveness
• Customer support operations
• Supply chain/inventory management
• Product usage/engagement

Your dashboard should include:

1. Top-level metrics  (3-5 KPIs for executives)
2. Trend visualizations  (how metrics change over time)
3. Segmentation  (break down by relevant dimensions)
4. Alerts or thresholds  (what indicates a problem?)

Deliverable : Sketch or describe the dashboard layout, including:

• What metrics to show
• What visualizations to use
• What interactivity to enable (filters, drill-downs)

Exercise 3: Root Cause Analysis

Scenario : An online subscription service has seen monthly churn rate increase from 5% to 8% over the past quarter.

Your task : Outline a diagnostic analytics approach to identify the root cause.

Include:

1. Decomposition : How would you break down the churn metric?
2. Segmentation : What dimensions would you analyze?
3. Hypotheses : What are 3-5 possible causes?
4. Data needed : What data would you need to test each hypothesis?
5. Analysis plan : What specific analyses would you run?

Deliverable : A structured plan (1-2 pages) for the root cause investigation.

Exercise 4: Predictive and Prescriptive for Capacity Planning

Scenario : A cloud services company needs to plan server capacity for the next 6 months. Under-capacity leads to service outages and lost customers. Over-capacity wastes money on unused servers.

Your task : Propose both predictive and prescriptive approaches.

Predictive Analytics:

1. What would you predict? (Be specific about the target variable)
2. What data/features would you use?
3. What forecasting method would you recommend and why?
4. How would you communicate uncertainty in the forecast?

Prescriptive Analytics:

1. What decision needs to be made?
2. What are the decision variables?
3. What's the objective function (what are you optimizing)?
4. What are the constraints?
5. What additional analysis would help (e.g., scenario planning, simulation)?

Deliverable : A proposal (2-3 pages) outlining your approach for both predictive and prescriptive analytics, including expected business value.