FinTFM · own record

Where this model actually sits

FinTFM is a tabular foundation model for corporate credit risk, pretrained only on synthetic data. This page is its standing under TabArena's protocol, published in full — including the part where it loses.

763 Elo · 94th of 95 · the published model, own classification head (re-checked 2026-10-06)

Every method's Elo on one axis, from the 2026-09-27 snapshot. The marked rule is the unpublished, exploratory linear-probe variant (910 Elo); it clears a tuned k-nearest-neighbours and sits just under an untuned logistic regression. The published model's own head is not in this snapshot. The leaders are a long way right.

This is not tabarena.ai. It is FinTFM's own comparison: 27 TabArena-Lite datasets — the subset this model's published checkpoint can run on — at one split each, against a cached snapshot of 95 other methods.

The official board covers all 51 datasets across every split, so its Elo values are not comparable to these, even for a method with the same name. Elo is only meaningful relative to the population it was computed over. The real board is at tabarena.ai.

Two variants of the same weights

The first is the entry submitted to TabArena and withdrawn on competitiveness; it uses the model's own classification head. The second swaps that head for a plain logistic regression fit on the model's internal representation — identical weights, different final layer. Why that helps is written up in full.

763

Own classification head

rank 94 of 95 · submitted, withdrawn

910

Linear probe on the representation

rank 93 of 95 · unpublished, exploratory, +145 Elo

The full comparison

Sort by any column. Bars are Elo across the observed range; the faint rule behind each bar is its confidence interval.

#	Method	Elo & interval	Elo	Win rate