Building Economic Models: Data Sources & Quantitative Framework

Why Build Economic Models?

Models help you:

• Formalize relationships
• Make systematic forecasts
• Backtest investment ideas
• Quantify uncertainty
• Remove emotional bias

"All models are wrong, but some are useful." — George Box

Model Types for Macro

Nowcasting Models

Real-time GDP estimation.

Examples:

• Atlanta Fed GDPNow
• NY Fed Staff Nowcast
• Private sector trackers

Approach: Mix of partial indicators before official release.

Forecasting Models

Forward-looking predictions.

Types:

• Leading indicator models
• Econometric (regression-based)
• VAR (Vector Autoregression)
• DSGE (Dynamic Stochastic General Equilibrium)
• Machine learning

Factor Models

Reduce many variables to few factors.

Examples:

• Chicago Fed National Activity Index
• Principal component models
• Diffusion indices

Recession Probability Models

Binary outcome prediction.

Types:

• Probit/Logit models
• Yield curve models
• Composite indicator models

Data Requirements

Time Series Properties

Stationarity:

Many models require stationary data.

Solutions: Differencing, detrending, logs.

Frequency:

• Monthly: Most economic data
• Quarterly: GDP, productivity
• Daily: Financial data

Alignment:

Mixed frequency = challenges.

MIDAS models, interpolation needed.

Data Quality Checks

Before modeling:

• Missing values (interpolate? exclude?)

• Outliers (COVID, crises)

• Revisions (vintage data vs real-time)

• Structural breaks (1984, 2008, 2020)

Vintage Data

Real-time data analysis requires:

• ALFRED (FRED real-time database)
• Philadelphia Fed real-time datasets
• Understanding revision patterns

Building a Simple Model: Example

GDP Nowcasting Model

Step 1: Identify inputs

Monthly indicators available before GDP:

• ISM Manufacturing
• Retail sales
• Industrial production
• Payroll employment
• Initial claims

Step 2: Align timing

• GDP: Quarterly, released ~4 weeks after quarter
• Monthly data: Varying release dates

Step 3: Specify relationship

Simple regression:

GDP_growth = α + β₁(ISM) + β₂(Retail) + β₃(IP) + ε

Or bridge equations with monthly aggregation.

Step 4: Estimate parameters

• OLS for simple models
• Kalman filter for state space
• Choose estimation window

Step 5: Validate

• Out-of-sample testing
• Compare to benchmarks
• Real-time vs revised data performance

Regression Models

Simple OLS

Good starting point.

FRED has most data ready.

Example: Inflation model

CPI_change = α + β₁(Unemployment_gap) + β₂(Oil) + β₃(ULC) + ε

Dynamic Models

Include lags.

Capture persistence.

Example: AR(1) with explanatory variables

Y_t = α + ρY_{t-1} + βX_t + ε_t

Error Correction Models

For cointegrated series.

Long-run and short-run dynamics.

VAR Models

What Is VAR?

System of equations where each variable depends on:

• Own lags
• Lags of other variables

Applications

• Impulse response analysis (what happens when Fed raises rates?)
• Variance decomposition (what drives output?)
• Forecasting (all variables simultaneously)

Practical Considerations

• Lag selection (AIC, BIC)
• Variable ordering (for impulse responses)
• Structural identification

Factor Models

Principal Components

Reduce many indicators to few factors.

Process:

• Collect many economic series

• Extract common factors

• Use factors for forecasting

Example: Global activity factor from:

• Multiple countries' GDP
• Trade data
• Industrial production

Diffusion Indices

Chicago Fed NAI:

85 indicators → Single index

Interpretation:

• Above 0: Above-trend growth
• Below 0: Below-trend growth

Machine Learning Approaches

Random Forests

• Handle nonlinearity
• Feature importance ranking
• Less overfitting than single trees

Gradient Boosting (XGBoost)

• Often best predictive performance
• Regularization built in
• Interpretability tools available

Neural Networks

• For complex patterns
• Requires more data
• Less interpretable

ML Cautions

• Economic relationships can change
• Small samples in macro
• Overfitting risk
• Explainability challenges

Backtesting Best Practices

Out-of-Sample Testing

Never test on training data.

• Split sample (70/30)
• Rolling window
• Cross-validation (careful with time series)

Real-Time Evaluation

Use vintage data if possible.

Revised data flatters models.

Benchmark Comparison

Compare to:

• Random walk
• Historical average
• Survey consensus

Performance Metrics

• RMSE (Root Mean Square Error)
• MAE (Mean Absolute Error)
• Direction accuracy
• R-squared (in-sample only)

Practical Implementation

Python Ecosystem

• pandas: Data handling
• statsmodels: Econometrics
• scikit-learn: ML models
• fredapi: FRED data access

R Ecosystem

• quantmod: Financial data
• vars: VAR models
• caret/tidymodels: ML
• fredr: FRED access

Data Pipeline

• Source: FRED, BLS, Census APIs

• Store: Database or files

• Transform: Seasonal adjustment, alignment

• Model: Your analysis

• Output: Forecasts, reports

Model Maintenance

Monitoring

• Track real-time performance
• Compare to consensus
• Flag large errors

Updating

• Re-estimate periodically
• Evaluate structural changes
• Consider new variables

Documentation

• Code version control (Git)
• Model specification record
• Performance history

Pro Tips

• Simple often wins: Complex ≠ better

• Data quality first: Garbage in, garbage out

• Real-time matters: Backtests flatter

• Ensemble approaches: Combine models

• Understand limitations: Models are tools

• Economic intuition: ML needs domain knowledge

Explore model data sources →