A pricing analyst fits a model to predict monthly revenue per store. The target ranges from $8,000 to $2.1 million and is heavily right-skewed. Residuals fan out: small stores are predicted within a few thousand dollars, large stores miss by hundreds of thousands.
What is the most appropriate fix?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Fanning residuals on a skewed, strictly positive target are the classic signature for a log transform. On the log scale the model predicts percentage error, so a 5% miss is a 5% miss for both an $8K and a $2M store. Predictions are transformed back with exp(); report errors on the original scale so the business can read them.
The large stores are the revenue. Deleting them makes the model useless for the decisions that matter most.
A different scoreboard does not make the predictions better. R² still rewards fitting the large stores.
Polynomials change the shape of the mean, not the spread of the errors. They will overfit long before they fix heteroscedasticity.