GLS and Cochrane-Orcutt for AR(1) Errors

Reported in interviews at Bank of America Citadel Two Sigma

Problem statement

You are fitting a linear regression $y_t = x_t' \beta + u_t$ where the errors follow an AR(1) process:

$$u_t = \rho \, u_{t-1} + \varepsilon_t, \quad |\rho| < 1, \quad \varepsilon_t \sim N(0, \sigma^2)$$

OLS ignores the serial correlation in $u_t$, so the standard errors are wrong and the estimator is inefficient.

  1. Assuming $\rho$ is known, derive the GLS estimator $\hat{\beta}_{\text{GLS}}$. Write out the transformed regression explicitly.
  1. In practice $\rho$ is unknown. Describe the Cochrane-Orcutt feasible GLS (FGLS) procedure for jointly estimating $\rho$ and $\beta$. State the asymptotic properties of the resulting estimator.
  1. Suppose you suspect the AR(1) specification for the errors might be wrong. Outline how you would use the Newey-West HAC estimator as a robust alternative, and explain when you would prefer it over FGLS.
Show hint

If you knew the error covariance, what linear transformation of each observation would leave you with uncorrelated errors?

Relevant roles: Data Scientist, Applied Scientist and Research Scientist.

Related free problems