Inverse Regression Prediction Bias

Reported in interviews at DRW

Problem statement

You run a linear regression $y = ax + b$ using OLS. Later, you realize you need the inverse relationship $x = cy + d$ instead. Without re-running the regression, you compute $c = 1/a$ and $d = -b/a$ (algebraic inversion of the fitted line).

Will predictions from this inverted regression overestimate, underestimate, or neither, compared to the correct OLS regression of $x$ on $y$? Explain why.

Show hint

Which loss does each of the two regressions minimize, and are those the same optimization problem?

Relevant roles: Data Scientist, Applied Scientist and ML Engineer.

Related free problems