Week 6: Forecasting for Decision-Making

Forecasting for Decision-Making

From patterns in data \(\rightarrow\) real-world decisions

Opening Scenario: Arches National Park

It is October. You must decide for next summer:

  • How many seasonal staff to hire
  • Whether to implement reservations
  • How many shuttle buses to contract
  • How much overtime to budget

What information do you wish you had?

Expected visitation.

Forecasting reduces uncertainty about the future.

What Is Forecasting?

Forecasting is:

A transparent, testable prediction about future values based on past patterns.

Forecasting is not:

  • Guessing
  • A crystal ball
  • A guarantee

Forecasting is:

  • A model of how the world has behaved
  • With quantified uncertainty
  • Evaluated on held-out data

Why Forecasting Matters in Business and Policy

Operations

  • Staffing
  • Inventory
  • Capacity planning

Finance

  • Revenue projections
  • Budgeting
  • Capital planning

Public policy

  • Emergency planning
  • Infrastructure investment
  • Tourism management

Forecasting Is About Decisions

A forecast is useful only relative to the decision it informs.

Two models can have similar accuracy, but:

  • One consistently overpredicts
  • One consistently underpredicts

Those create different costs.

  • Overpredict visitation \(\rightarrow\) overspend on staffing
  • Underpredict visitation \(\rightarrow\) overcrowding and safety risk

Forecast evaluation must consider decision costs.

Example: Arches National Park

The Forecasting Problem

Every forecast must answer:

  1. What patterns exist?
    • Trend
    • Seasonality
    • Cycles
    • Shocks
  2. Will those patterns continue?
  3. How uncertain are we?

Sources of Predictability

Time series are easier to forecast when:

  • Trend is persistent
  • Seasonality is stable
  • Autocorrelation is strong

Forecasting is harder with:

  • Sudden policy changes
  • Structural breaks
  • Technological shifts
  • Extreme shocks

The Role of Benchmarks

Benchmarks are simple forecasts that set a baseline for performance.

Can you beat the guess: “this month-year will look like the same month last year”?

Seasonal naive is strong because:

  • Seasonality is often dominant
  • No parameter estimation required
  • It adapts automatically

Train/Test Split and Honest Evaluation

Never evaluate a forecast on data used to estimate it.

  • Training set \(\rightarrow\) estimate model parameters
  • Test set \(\rightarrow\) simulate future performance

Without out-of-sample evaluation, you reward overfitting.

What Makes a Good Forecast?

Not just low error.

A good forecast is:

  • Transparent
  • Interpretable
  • Stable across samples
  • Appropriate for the decision horizon

Different horizons imply different model choices.

Forecasting Horizon Matters

Short horizon:

  • Autocorrelation dominates
  • Naive methods are often strong

Long horizon:

  • Trend assumptions dominate
  • Structural change risk increases

Would you use the same model for:

  • Next month?
  • Next year?
  • Ten years?

Forecasting Under Policy Change

Arches reservation policy:

  • Introduced in 2020
  • Removed in 2026

Key modeling question:

Should you assume:

  • Past pattern continues?
  • New regime starts?
  • Structural break at policy change?

Forecasting involves causal reasoning, not just extrapolation.

Forecasts and Uncertainty

Point forecast \(\neq\) full forecast.

Decision-makers need:

  • Prediction intervals
  • Range of plausible outcomes
  • Worst-case planning inputs
  • Risk assessment context

Applications:

  • Fire risk planning
  • Tourism congestion management

The Economic Cost of Forecast Errors

Suppose July error is 50,000 visitors.

What does that mean economically?

  • Staff shortfall or overstaffing
  • Parking congestion
  • Lost revenue opportunities
  • Overtime and contracting costs

MAE and RMSE are proxies for decision cost.

When Forecasting Fails

Forecasts fail when:

  • Structural breaks occur
  • Policy changes behavior
  • Climate shocks hit demand
  • Pandemics disrupt patterns
  • Models extrapolate unstable trends

Most models assume the future resembles the past.

That assumption is fragile.

Why Decomposition Makes Sense

Most economic time series combine:

  • Trend
  • Seasonality
  • Irregular shocks

Decomposition helps by:

  • Separating patterns
  • Improving interpretability
  • Supporting better forecasts

\[ y_t = T_t + S_t + R_t \]

Forecasting Is Iterative

Forecasting is not: fit once, done.

Forecasting is:

  1. Fit
  2. Evaluate
  3. Diagnose bias
  4. Update assumptions
  5. Refit

Especially important when policy regimes change.

Linking Back to Business Analytics

In this course pipeline:

  • EDA \(\rightarrow\) understand structure
  • Modeling \(\rightarrow\) estimate relationships
  • Forecasting \(\rightarrow\) project forward
  • Evaluation \(\rightarrow\) test performance
  • Decision \(\rightarrow\) allocate resources

Forecasting is the bridge between data and planning.

Discussion Prompt

If Arches removes reservations in 2026, how would you forecast 2027 visitation?

Possible strategies:

  • Ignore reservation-period years
  • Add a policy indicator
  • Weight pre-policy years more heavily
  • Build scenario forecasts

Optional 10-Minute Case Activity

Scenario: “Visitation is expected to be 15% higher than last year.”

Before hiring 20 more seasonal staff, what else do you need?

  • Forecast uncertainty and interval width
  • Timing within the season
  • Budget constraints
  • Operational bottlenecks

Forecasts support decisions under uncertainty.

Big Takeaway

Forecasting is not about predicting perfectly.

Forecasting is about:

  • Quantifying uncertainty
  • Improving decisions
  • Learning from forecast errors
  • Updating beliefs with new data

Conceptual Flow: From Data to Decision

%%{init: {"themeVariables": {"fontSize": "24px"}}}%%
flowchart LR
  P[Past Data] --> E[Extract Patterns]
  E --> F[Forecast + Uncertainty]
  F --> C[Decision Costs]
  C --> D[Operational or Policy Decision]
  D --> L[Learn and Update]
  L --> E