The 67× test: our credit model on a year it never saw

Only about 1% of UK companies take on new secured borrowing in any twelve months. The claim behind Scout is that we can tell you which ones. Claims like that deserve a harder test than a train/test split — so we gave it one.

The setup: no peeking

We rebuilt the register exactly as it stood on 30 June 2025 — a vintage our models were never trained on — scored every company blind, and waited. The 12-month window closed on 30 June 2026. Then we opened the envelope.

The results

The lift numbers are loud, but the quieter column matters more: the model predicted 69.1% for its top credit slice, and 72.4% happened. On refinancing it said 30.4% and 30.6% happened. It doesn't just rank companies — it prices the probability, and the price is honest. For anyone underwriting risk, "we say 30% and 30% occurs" is worth more than any multiple.

The other half of the story is the bottom of the book: the lower half of all scores produced near-zero events — 0 to 0.1% outcome rates. Knowing where not to spend origination effort is half the value of a model like this.

The honest footnotes

• The 67× figure is the top 0.1% of scores. On the broader top 1% the deployed setup lifts about 17.6× — still a different sport from SIC-code lists, but we'd rather you hear both numbers from us.
• The sale model's 36.7× is a lower bound: its label window is 24 months and only 13 had elapsed when we measured. It "over-predicts" only because the future hasn't finished arriving.
• Nothing here was retrained, recalibrated or cherry-picked after the fact: one frozen model, one unseen year, one readout.

If you're in commercial lending or broking and your pipeline still starts with a sector list, this is what the alternative looks like: start free or talk to us.