← Freddie Mac Interview Insights

Freddie Mac·Data Scientist·Technical Phone Screen·Senior

Senior
Jul 2026

Summary

Freddie Mac data scientist interview, heavy quant focus on panel econometrics for real estate data. The whole thing felt like a graduate seminar more than a job interview, which I was not fully prepared for.

Questions Asked (5)

Q1

You have monthly panel data on rental properties across MSAs from 2010 to 2025. Write a regression to estimate the elasticity of rent with respect to vacancy, including property and month fixed effects. Justify your choice of log transforms and any seasonal terms.

Data ModelingProduct Analytics & Metrics
Author's notes

I knew the log-log setup cold, elasticity interpretation is straightforward.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by specifying a two-way fixed effects model with log rent as the dependent variable and log vacancy rate as the key regressor, including property and month fixed effects. Justify log transforms for interpretability as elasticities and to handle skewness, and include seasonal terms (e.g., month dummies or Fourier terms) to capture within-year patterns. Discuss identification concerns and potential endogeneity, suggesting robustness checks like lagged vacancy or instrumental variables.

Pro tip: Emphasize that property fixed effects absorb time-invariant property characteristics, while month fixed effects control for common shocks; but be cautious about the reflection problem if vacancy and rent are jointly determined in the same market. Mention clustering standard errors at the MSA level to account for spatial correlation.

1. Specify the baseline model

Write the regression equation: log(Rent_it) = β log(Vacancy_it) + α_i + γ_t + ε_it, where α_i are property fixed effects and γ_t are month fixed effects. Explain that β is the elasticity of rent with respect to vacancy.

2. Justify log transforms

Log transforms allow interpreting β as an elasticity, reduce skewness in rent and vacancy, and linearize multiplicative relationships. Also, log(vacancy) handles the bounded nature of vacancy rates (0-1) better than levels.

3. Address seasonality

Include month fixed effects (or month-of-year dummies) to capture seasonal patterns in rental markets. Alternatively, use Fourier terms (sin/cos) if you want a more parsimonious specification, but month dummies are more flexible.

4. Discuss identification and endogeneity

Acknowledge that vacancy may be endogenous due to simultaneity (rent and vacancy determined together) or omitted time-varying confounders. Suggest using lagged vacancy, instrumental variables (e.g., local construction permits), or dynamic panel models as robustness checks.

5. Inference and robustness

Cluster standard errors at the MSA level to account for within-MSA correlation over time. Perform robustness checks: alternative functional forms (e.g., Poisson for counts), different fixed effects (MSA-by-month), and placebo tests.

Key Points to Mention

  • Elasticity interpretation: β is the percentage change in rent for a 1% change in vacancy.
  • Property fixed effects control for time-invariant property and location characteristics.
  • Month fixed effects control for common seasonal and macroeconomic shocks.
  • Log transforms mitigate skewness and heteroskedasticity, and handle multiplicative effects.
  • Endogeneity concerns: simultaneity between rent and vacancy, omitted variables like local economic conditions.
  • Clustered standard errors at the MSA level to address spatial correlation.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

Given serial correlation and heteroskedasticity in your panel data, which standard error estimator would you use and why? Compare two-way clustering by property and month, Driscoll-Kraay, and Newey-West. What are the finite-sample tradeoffs?

Data ModelingTechnical Trade-offs
Author's notes

This is where I started sweating.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem: serial correlation and heteroskedasticity in panel data violate OLS assumptions, so you need a robust covariance estimator. Then compare the three estimators on their assumptions, asymptotic properties, and finite-sample behavior, and conclude with a recommendation tailored to the data's dimensions and structure.

Pro tip: Mention that in Freddie Mac's context, the choice often hinges on the relative size of N and T; for short panels with large N, two-way clustering is common, but for long panels with small N, Driscoll-Kraay may be preferable. Also, note that Newey-West is a special case of Driscoll-Kraay when there is no cross-sectional dependence.

1. Diagnose the error structure

Identify the presence of serial correlation, heteroskedasticity, and cross-sectional dependence. Discuss tests like Wooldridge test for serial correlation and Breusch-Pagan for heteroskedasticity.

2. Compare estimator assumptions

Explain that two-way clustering assumes independence across clusters but allows arbitrary correlation within clusters; Driscoll-Kraay allows for cross-sectional dependence and serial correlation; Newey-West handles serial correlation and heteroskedasticity but assumes cross-sectional independence.

3. Evaluate finite-sample tradeoffs

Discuss how two-way clustering can be unreliable when the number of clusters is small; Driscoll-Kraay may over-reject in small samples; Newey-West can be biased with many time periods. Mention the need for bias adjustments or bootstrap methods.

4. Match estimator to data dimensions

Recommend based on N and T: if N is large and T small, two-way clustering is feasible; if T is large and N small, Driscoll-Kraay is better; if no cross-sectional dependence, Newey-West suffices.

5. Conclude with a recommendation

State which estimator you would use and why, acknowledging that the choice depends on the specific dataset and the presence of cross-sectional dependence.

Key Points to Mention

  • Serial correlation and heteroskedasticity violate OLS assumptions, leading to biased standard errors.
  • Two-way clustering by property and month accounts for correlation within properties and within months, but requires a sufficient number of clusters.
  • Driscoll-Kraay standard errors are robust to cross-sectional dependence, serial correlation, and heteroskedasticity, and are consistent as T grows.
  • Newey-West standard errors handle serial correlation and heteroskedasticity but assume no cross-sectional dependence.
  • Finite-sample performance: two-way clustering can be downward biased with few clusters; Driscoll-Kraay may over-reject in small samples; Newey-West may be inefficient with many time periods.
  • In practice, use Driscoll-Kraay when cross-sectional dependence is suspected and T is large; use two-way clustering when N is large and clusters are well-defined; use Newey-West when cross-sectional independence is plausible.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

Test for unit roots and cointegration in log rent and log vacancy. If both series are I(1), how would you set up an error correction model using either the Engle-Granger or Johansen approach to avoid spurious regression?

Data ModelingTechnical Trade-offs
Author's notes

Honestly the unit root part I handled fine, ADF test, mention the lag selection criteria, done.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by explaining the importance of testing for unit roots and cointegration to avoid spurious regression. Then outline the steps: test each series for unit roots (e.g., ADF, PP, KPSS), test for cointegration (Engle-Granger or Johansen), and finally specify an error correction model (ECM) that incorporates the long-run relationship and short-run dynamics. Emphasize how the ECM captures both levels and differences.

Pro tip: Mention that in practice, you should also consider structural breaks and lag selection criteria (AIC/BIC) when testing for unit roots and cointegration, as ignoring them can lead to incorrect conclusions.

1. Test for Unit Roots

Apply unit root tests (ADF, PP, KPSS) to log rent and log vacancy to determine if they are I(1). Check both levels and first differences.

2. Test for Cointegration

If both are I(1), test for cointegration using Engle-Granger (two-step) or Johansen (system) approach. Engle-Granger: estimate long-run regression, test residuals for stationarity. Johansen: use trace and max eigenvalue tests.

3. Estimate Long-Run Relationship

If cointegrated, estimate the long-run equilibrium relationship (e.g., log rent = α + β log vacancy + ε) to obtain the error correction term (residuals).

4. Specify Error Correction Model

Construct ECM: Δlog rent_t = γ1 + δ1 * ECT_{t-1} + lagged Δlog rent + lagged Δlog vacancy + ε_t, and similarly for Δlog vacancy. Include lagged differences to capture short-run dynamics.

5. Validate and Interpret

Check ECM residuals for autocorrelation and normality. Interpret the speed of adjustment coefficient (δ) and ensure it is negative and significant for convergence.

Key Points to Mention

  • Spurious regression occurs when regressing non-stationary series without cointegration, leading to misleading high R² and t-statistics.
  • Engle-Granger is simpler but assumes a single cointegrating vector and is sensitive to the choice of dependent variable; Johansen can handle multiple cointegrating vectors and is more robust in systems.
  • The error correction term (ECT) represents the deviation from long-run equilibrium and its coefficient measures the speed of adjustment back to equilibrium.
  • Lag selection in ECM should be based on information criteria (AIC, BIC) and residual diagnostics.
  • In practice, also test for structural breaks (e.g., Chow test, Zivot-Andrews) as they can affect unit root and cointegration results.
  • For Freddie Mac, consider economic interpretation: log rent and log vacancy likely have a long-run relationship due to supply and demand in housing markets.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q4

If your model returns a coefficient of -0.35 on log vacancy, what does that mean? Specifically, compute the predicted percent change in rent when vacancy increases from 8% to 10% at the MSA level.

Product Analytics & MetricsData Modeling
Author's notes

The interpretation part is easy, a 1% increase in vacancy is associated with a 0.35% decrease in rent.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, interpret the coefficient as the effect of a one-unit increase in log vacancy on log rent, then compute the log difference for the given change and convert to a percentage. Emphasize that this is an elasticity and that the calculation assumes a log-log model specification.

Pro tip: Always clarify that the coefficient is an elasticity and that the percentage change is approximate for large changes; for precise results, exponentiate the log difference and subtract 1. Also, mention that this is a ceteris paribus effect and that real-world predictions may differ due to model limitations.

1. Interpret the coefficient

Explain that -0.35 is the estimated elasticity of rent with respect to vacancy: a 1% increase in vacancy is associated with a 0.35% decrease in rent, holding other factors constant.

2. Compute the change in log vacancy

Calculate the natural log of the new vacancy rate minus the natural log of the old rate: ln(0.10) - ln(0.08) = ln(1.25) ≈ 0.2231.

3. Apply the coefficient to get log rent change

Multiply the coefficient by the change in log vacancy: -0.35 * 0.2231 ≈ -0.0781. This is the change in log rent.

4. Convert log change to percentage change

Exponentiate the log change and subtract 1: exp(-0.0781) - 1 ≈ -0.0751, or about a 7.5% decrease in rent. Alternatively, for small changes, the log change itself approximates the percentage change.

5. State the final answer with context

Conclude that increasing vacancy from 8% to 10% is associated with an approximate 7.5% decrease in rent at the MSA level, assuming the model is correctly specified and other factors are held constant.

Key Points to Mention

  • The coefficient is an elasticity in a log-log model.
  • The calculation uses the difference in logs: ln(10/8) = ln(1.25).
  • The exact percentage change is obtained via exponentiation: exp(β*Δln(x)) - 1.
  • For small changes, the log change approximates the percentage change.
  • The interpretation is ceteris paribus and assumes the model is correctly specified.
  • The result is an association, not necessarily causation, and may be subject to omitted variable bias.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q5

How would you diagnose multicollinearity among macro regressors like unemployment, CPI, and interest rates? What remedies would you consider, such as orthogonalization, ridge regression, or Bayesian priors, and how does each affect your inference?

Data ModelingTechnical Trade-offsRoot Cause Analysis
Author's notes

VIF, condition number, I rattled those off.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by explaining how you would detect multicollinearity using VIF, condition number, and correlation matrices, then discuss the trade-offs of remedies like orthogonalization, ridge regression, and Bayesian priors, emphasizing their impact on inference. Conclude with a practical recommendation tailored to Freddie Mac's macro forecasting context, balancing interpretability and predictive stability.

Pro tip: Mention that multicollinearity doesn't bias coefficients but inflates standard errors, so if prediction is the goal, you might tolerate it; if inference on individual effects is needed, consider remedies. Also, highlight that orthogonalization changes the interpretation of coefficients, which can be problematic for policy analysis.

1. Detect multicollinearity

Use correlation matrices, variance inflation factors (VIF), and condition number to quantify the severity. Check if VIF > 5 or 10 and if condition number > 30 indicates potential issues.

2. Assess impact on inference

Determine whether multicollinearity is harmful for your goal: it inflates standard errors, making coefficients unstable and insignificant, but doesn't bias estimates. Consider if the model is for prediction or explanation.

3. Evaluate remedies

Discuss orthogonalization (e.g., PCA or residualization), ridge regression, and Bayesian priors. For each, explain how it addresses multicollinearity and its effect on coefficient interpretation and inference.

4. Choose and justify

Select a remedy based on the context: if interpretability is key, ridge may not be ideal; if forecasting, orthogonalization or ridge can help. Bayesian priors can regularize and provide probabilistic inference.

5. Validate and communicate

Validate the chosen approach with cross-validation or out-of-sample testing, and clearly communicate the trade-offs and limitations to stakeholders.

Key Points to Mention

  • VIF and condition number as diagnostic tools
  • Multicollinearity does not bias coefficients but inflates standard errors
  • Orthogonalization (e.g., PCA) creates uncorrelated components but complicates interpretation
  • Ridge regression introduces bias to reduce variance, improving prediction but complicating inference
  • Bayesian priors (e.g., regularizing priors) can stabilize estimates and provide full posterior inference
  • Trade-offs between prediction accuracy and interpretability in macro forecasting

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.