Unverified50% confidenceBenchmarkExact time
TabPFN-2's median prediction had an average error as a share of sales of 0.502, versus 0.486 for the 4-week mean × k baseline; TabPFN-2 also outperformed a gradient boosting model trained on the same features (0.527), the last-week rule (0.631), and the same-week-last-year rule (0.884).
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOf the 586 invoices for which TabPFN gave an overdue probability below 20%, 8.7% were actually overdue (the model predicted 6.2% on average); of the 116 for which it gave above 80%, 90.5% were actually overdue (the model predicted 87.4%).71% similarUnverifiedIn a benchmark of 30 held-out test windows, TabPFN achieved 100.0% overall accuracy, 100.0% shock recall, 0.33 ms CPU latency, 0.0% step-frequency false alarms, and required no training or tuning.71% similarUnverifiedIn the specific experimental case, the overlapping-anchor scheme (n=912) showed the model at 55.2% versus the majority class at 51.1%, a net advantage of +4.1pp, about 2.4 standard errors; the non-overlapping-anchor scheme (n=163) showed 59.5% versus 54.0%, a net advantage of +5.5pp, about 1.4 standard errors.69% similarUnverifiedIn public health benchmarks, an ordinary AI model scored higher than Tabibu in the developer's own re-run, 0.522 versus 0.426 (non-official leaderboard data), and the developer believes the reason is that Tabibu's safety layer lowered its score in the benchmark.69% similarUnverifiedUnder the condition of a single table and zero feature engineering, TabFM outperformed the author's previously carefully tuned tree model in a customer churn prediction demo, but the score was only 0.60.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1051648API
curl https://kongchang.com/api/v1/knowledge/claims/1051648MCP
get_claim(id=1051648)