DATA4400: Data-Driven Forecasting
Week 11: Adaptive Forecasting
Matching Methods to Business Problems
By the end of this lesson, you will be able to:
| Week | Topic |
|---|---|
| 1-4 | Foundations: Data preparation, moving averages, stationarity, basic methods |
| 5 | Assessment 1: Group presentation |
| 6-8 | Advanced Methods: Prophet, ARIMA/SARIMA, VAR models |
| 9 | Regression models + Assessment 2 |
| 10 | End-to-End Solutions: Building complete forecasting systems |
| 11 | Adaptive Forecasting: Making intelligent method choices |
| 12 | Assessment 3: Individual forecasting project |
Today's Focus: You've learned the methods. Now learn how to choose and apply them like a professional Business Analyst.
Before we can choose a forecasting method, we need to understand what stakeholders are really asking for.
| What They Say | What They Might Really Mean | Your Follow-Up Questions |
|---|---|---|
| "Forecast customer churn" | • Identify at-risk customers • Budget retention costs • Predict revenue impact • Test intervention strategies |
• What action will you take with the forecast? • What's the cost of losing a customer? • How far ahead do you need to know? |
| "Predict sales for next quarter" | • Set production targets • Plan inventory • Allocate marketing budget • Report to investors |
• Is overproduction or stockout worse? • Can we adjust after seeing trends? • Are there upcoming campaigns/changes? |
| "How accurate can you be?" | • Is this forecast reliable enough? • Should we automate the decision? • Do we need more data? |
• What decision changes at different accuracy levels? • What's your current decision process? • What's the cost of being wrong? |
→ The same words can mean completely different things. Always clarify before building models.
Scenario: A streaming service stakeholder says: "We're losing customers. Can you investigate churn so we can fix it?"
Identify who will churn next month
→ Need: Classification model on customer attributes
→ Data: Customer behavior, demographics, engagement
→ Output: List of at-risk customers for retention calls
Forecast total churn rate over time
→ Need: Time series forecast
→ Data: Historical monthly churn rates
→ Output: Expected churn % for revenue planning
Find what causes churn
→ Need: Causal analysis (regression/survival)
→ Data: Customer journey events, service quality
→ Output: Actionable drivers to fix
Test if marketing reduces churn
→ Need: A/B test or intervention analysis
→ Data: Control vs. treatment groups
→ Output: ROI of retention campaigns
→ Same request. Four completely different projects. Your job: Ask questions before building anything.
Now that you understand the business problem, how do you choose the right forecasting method?
When choosing a forecasting method, consider these factors in this order:
1. BUSINESS CONTEXT
What decision will this forecast drive? What costs are associated with errors?
2. DATA PATTERN
Does your data show trend? Seasonality? Are there external drivers? Multiple variables?
3. EXPLAINABILITY NEEDS
Does the stakeholder need to understand WHY, or just WHAT will happen?
4. PRACTICAL CONSTRAINTS
How much data do you have? How often will you update? Do you need automation?
→ Never start with "Which method is most accurate?" Start with "Which method fits the business need?"
| Data Pattern | Suitable Methods | Business Example | When to Use |
|---|---|---|---|
| Level only (no trend/season) | Moving Average, SES | ATM withdrawals, daily sales in stable business | Quick baseline, short-term tactical decisions |
| Trend, no seasonality | Holt, AR(p), ARIMA | Subscription growth, market expansion | Business is scaling but cycles aren't established |
| Trend + seasonality | Holt-Winters, SARIMA, Prophet | Retail sales, airline passengers, energy demand | Established business with clear cycles |
| Multiple interacting variables | VAR, Multiple Regression | GDP/inflation/interest rates, price-volume dynamics | System-level forecasts, scenario planning |
| External drivers + events | Prophet, Regression with dummies | Promotions, holidays, weather effects | Business has controllable/known external factors |
→ This table is your starting point, not a strict rule. Business context may override pattern matching.
Best for: Stable operations with random fluctuations, no clear trend/seasonality
Business advantage: Fast, transparent, easy to explain to non-technical stakeholders
Use when:
Real example: A convenience store forecasting daily milk purchases (level data, high noise, need quick updates)
BA Decision Rule: If your data looks like random wiggling around a stable average, and you need an answer today, start here. Don't overcomplicate.
Best for: Data with clear trend but no strong seasonality
Business advantage: Captures momentum in growth/decline businesses
Choose Holt when:
Choose ARIMA when:
BA Decision Rule: If the business is scaling but seasonal patterns aren't yet clear or stable, use trend methods. Think: "We're growing, but how fast will growth continue?"
Best for: Established businesses with predictable seasonal patterns
Choose Holt-Winters when:
Choose SARIMA when:
Choose Prophet when:
Best for: When multiple time series influence each other, or external factors drive your target
Choose VAR when:
Choose Regression when:
BA Decision Rule: Use these when the question is "What drives our outcome?" not just "What will happen?"
Two critical questions every BA faces: "How much data do I actually need?" and "What do I do about that disruption?"
Common assumption: "More data is always better." Reality: Sometimes more data makes forecasts worse.
Example: Electricity demand patterns haven't fundamentally changed in 10 years → use all data
Example: Company went digital in 2020 → pre-2020 data may hurt accuracy
Stock price from 1980 vs. recent 5 years:
Pre-COVID (2015-2019) vs. Post-COVID (2021-2024):
→ Just because you have the data doesn't mean it's valid for forecasting.
Disruptions are sudden, significant changes that break historical patterns (COVID-19, regulatory changes, major events).
When: Disruption was temporary and won't repeat
Method: Remove the disruption period from training data
Example: One-time supply chain shock in Q2 2023 → exclude that quarter
When: Disruption changed behavior permanently or semi-permanently
Method: Add intervention variables (dummy/phase variables in Prophet/Regression)
Example: COVID-19 created "pre," "during," "post" phases with different patterns
Create binary (0/1) or phase (1, 2, 3) indicators:
Step 1: Is the disruption one-time or structural?
Step 2: Will similar disruptions happen again?
Step 3: Does your method support intervention variables?
1. UNDERSTAND the business problem
→ Decode stakeholder requests. Ask: What decision? What costs? What constraints?
2. MATCH method to context
→ Consider: Business need → Data pattern → Explainability → Constraints
→ Use the decision framework, not just "most accurate"
3. DECIDE on data scope
→ Ask: Is old data still relevant? Has the business fundamentally changed?
→ More ≠ better if patterns have shifted
4. HANDLE disruptions explicitly
→ Never ignore them. Either exclude or model with intervention variables
→ Choose based on: one-time vs. structural, repeatable vs. unique
→ You're not just running models. You're translating business problems into technical solutions that drive real decisions.
Assessment 3 (40% of grade) requires you to demonstrate everything we've covered:
Key Success Factors:
Next: Discussion time with your facilitator about A3 questions and concerns.
Questions? Concerns? Let's discuss.