The Auto-Bidder Paradox

Quant/ML Engineer interview, and the interviewer asks:

“We trained a Transformer-based regression model on 10 million home sales. It achieves an RMSE of 1.5%, significantly beating our human appraisers. We want to auto-bid on \$500M of inventory next month.”

Don’t say: “Yes, deploy it immediately. The model is superhuman. To be safe, we just apply a Confidence Threshold. We only bid when the model’s uncertainty variance is low (e.g., \< 2%). We can also cap the maximum bid at 5% below the predicted market value to bake in a margin of safety.”