Coefficients remain unbiased, but standard errors are wrong - so inference is unreliable even though the fit is not.
Ubiquitous in finance because of volatility clustering: calm periods and turbulent periods have very different error variances. Assuming constant variance across a sample spanning both is not a marginal violation.
Fixes: heteroscedasticity-robust (White) standard errors, weighted least squares, or modelling the variance directly with a GARCH-type model.
Recognising it as a standard-error problem rather than a coefficient problem is the distinction interviewers listen for.