JIAB-2026-BF66
Machine Learning Always Outperforms OLS: Evidence from 1,000 Real Datasets
Rejected
Editor
Sofia Reyes
Handling editor for this submission.
Assigned editor: Sofia Reyes
Update the handling editor to rebalance workload or hand off the submission.
Abstract
We compare 12 supervised learning algorithms against OLS regression across 1,000 publicly available datasets. In every dataset, at least one ML algorithm achieves lower out-of-sample MSE than OLS. We conclude that OLS should be abandoned in applied empirical work in favor of machine learning alternatives.
Reviews
1 itemR1 Henry Weber Submitted Sep 11, 2026
The benchmark conflates prediction accuracy with causal or descriptive inference. OLS is rarely deployed for out-of-sample prediction; it is used for causal identification, unbiased coefficient estimation, and interpretability — tasks the authors do not evaluate. The headline claim ('OLS should be abandoned') does not follow from the evidence presented. I recommend rejection.
Files
1 manuscriptReviewers
1 of 1 receivedDrag names to change the reviewer order for people who have not been invited yet.
Henry Weber
Messages
2 messagesA Author Cover letter
This comprehensive benchmark demonstrates the superiority of machine learning for all applied econometric tasks.
E Editor Rejection decision
Dear Dr. Patel,
Thank you for your submission. Our reviewer raises a fundamental conceptual issue: OLS and machine learning serve different inferential goals. OLS is not typically deployed for out-of-sample prediction, so benchmarking the two on predictive MSE does not support the paper's claims about when OLS should be used. We are unable to accept the manuscript in its current form and do not encourage resubmission without a substantial reconceptualisation.
Best regards,
The Editors
Timeline
2 events
Rejected
Sep 21, 2026
Waiting for Reviewers
Aug 23, 2026
Status override
Override status only to correct a workflow state set in error. Notification actions should use Compose.