BigQuery · Status at source date: Preview
BigQuery previews TabFM predictions without custom model training
BigQuery now offers a preview of TabFM, Google’s pretrained model for structured tables, supporting regression and classification without training a separate model for each task.
What changed
The 31 August BigQuery release entry introduces TabFM in preview for regression and classification on structured tables. It names AI.PREDICT for predictions and AI.EVALUATE for assessment. Google Research’s linked explanation, authored by Weihao Kong and Abhimanyu Das, describes learning from examples supplied in context without updating a separate model’s parameters for each task.
This is a way to explore predictive work in the warehouse, not evidence that an advertiser’s particular dataset will produce reliable scores. The historical release remains a preview in this article. The linked research post discusses benchmarks, but those results are not a substitute for evaluation on the business question the advertiser intends to solve.
Why it matters for advertisers
PMC analysis: define the prediction before choosing the technology. Predicting whether a lead becomes qualified is different from estimating the eventual value of a customer, and each needs a clear observation window. A label that changes halfway through the evaluation can make an apparently strong result impossible to use consistently.
An illustrative test could ask whether information available when a lead arrives helps prioritise later follow-up. Reserve records with known outcomes for evaluation, and exclude information added only after the outcome occurred. Otherwise, the model may appear to predict the future while merely reading evidence from it. Compare the result with a simple existing rule, such as the team’s present qualification process. A more sophisticated score is useful only if it improves an actual decision enough to justify its cost and maintenance. Keep the initial exercise separate from live bidding until that case has been made.
What to check next
- Choose one clearly defined numeric or classification outcome and document which fields would be available at prediction time.
- Use held-out business records to assess errors and compare TabFM with a simple baseline; inspect the cases where the methods disagree.
- Record the preview’s applicable limits and costs before operational use, and validate the effect of the proposed score on a real business decision before connecting it to advertising.
Sources & contributor credit
- Newsletter coverage · Paid Media Collective newsletter
Original newsletter text, contributor labels and media for this update.
- Source referenced in newsletterBigQuery release notes
Linked from the original newsletter. The source publication date has not been independently confirmed.
- Underlying research explanationGoogle Research
Read during historical rewrite; research explanation and original authorship, distinct from BigQuery preview release date.
Source attribution checked
Google publishes the release notes. Weihao Kong and Abhimanyu Das wrote the linked research explanation; its acknowledgements name Erez Louidor Ilan, Taman Narayan, Shuxin Nie, Rajat Sen, Yichen Zhou, Joe Toth, Deqing Fu and Samet Oymak as joint contributors and Kimberly Schwede as graphics designer. No profile or release-note authorship inferred.
Attribution evidence and limitations
Read the exact 31 August release-note section rather than unrelated accumulated releases. Followed its TabFM link and read Google Research article including author line and acknowledgements. Research post date differs from BigQuery release date; they are not conflated.
Retained original release-note and newsletter URLs. Followed linked Google Research article; no guessed profiles. Research authors are distinct from the unnamed release-note author.
Viewed web/public/news/newsletter/bigquery-tabfm-structured-table-predictions.png. Screenshot shows 31 August release section and preview label, no personal watermark. Google Research credits Kimberly Schwede for its graphics; those graphics are not this screenshot and were not visually inspected or reproduced.
- Read exact date section, TabFM preview, AI.PREDICT and AI.EVALUATE.
- Read full substantive article, Weihao Kong and Abhimanyu Das author line, team acknowledgement and separate Kimberly Schwede graphics credit.
Attribution checked . This is a review date, not the original publication date.
- Published on this site
- Article updated
This update reflects the dated source reporting. Availability may have changed. Further coverage of this same development will be added to this page.
Original newsletter text and archive evidence
BigQuery previews TabFM predictions without custom model training.g
BigQuery now offers a preview of TabFM, Google’s pretrained model for structured tables, supporting regression and classification without training a separate model for each task. Teams can request predictions through AI.PREDICT and assess performance with AI.EVALUATE, while the release remains a preview. For you, this means a new option for exploring customer or commercial predictions in the warehouse, with validation against held-out business data before those scores influence advertising decisions.
Source captured . No explicit first-contributor label was provided for this update.
