Every signal we serve, the grade it actually earned, and the four accuracy claims we withdrew. Nothing on this page is gated, and no figure appears without the level of verification behind it. Served predictions are logged only on the instance that answered and are not retained, and outcomes are not yet recorded, so the accuracy page is a validation record, not a track record.
signals graded and badged
accuracy claims withdrawn, each with its cause
outcomes recorded — a validation record, not a track record
date of the published validation record
Confidence scores in the served envelope are capped by validation grade. A figure is never served with more confidence than its level allows.
Sample, synthetic or scaffold data with no fitted model behind it. Shown only to describe a structure, never as a basis for a decision.
Real underlying data and a fitted model, but no validation gate has been passed yet.
Validation evidence exists, in sample or out of time on a single window, but no point-in-time out-of-sample test has been passed and no track record is recorded.
Passed a held-out, point-in-time out-of-sample test on its recorded date. The level describes that record.
Each signal with the grade it earned rather than the best number available.
Out-of-filer grade: leave-one-BDC-out AUC 0.5824 on the filer's own disclosed non-accrual label and 0.6790 on the wider label that also counts a mark below 0.85x amortized cost. The disclosed-non-accrual grade is the better-founded of the two, because that label is the filer's own footnote and no duplicated position can inflate it.
Leave-one-BDC-out AUC 0.58 [95% CI 0.53, 0.64] on 126 disclosed non-accrual events, and 0.68 [95% CI 0.65, 0.70] on 649 events of the wider label (Hanley-McNeil; assumes independent borrowers, so the true interval is wider). This is a concurrent impairment classifier, not a probability of default.
The model scores 3,003 of the 3,034 borrowers in the panel and refuses 31 — 1.0% — by name: a refused borrower's filing discloses neither a spread nor an exposure amount the model may use. Share and unit counts are never read as money, a nil-percentage dash is never read as zero, an instrument line that has lost its issuer is not counted as a borrower, and a name that wraps an instrument around a named issuer is resolved back to that issuer.
BDC panel as of 2026-06-30 filings (filed 2026-07-29 to 2026-08-20), 7,828 tranches over 3,034 borrowers from 14 BDC filers.
The private-credit output is a disclosed-impairment score. It is not a validated probability of default. It reads impairment already disclosed in the same filing that supplies its features, so there is no forward observation window and no horizon attaches to it anywhere in the product.
Four numbers we published, then withdrew.
Each with the specific failure that killed it and the number, if any, that replaced it.
0.79–0.81as of 2026-09-01AUC across label gaps · one recession in the out-of-sample window
Label leakage worth 5–7 AUC points, a look-ahead benchmark, and an out-of-sample window containing a single recession. Leak-free walk-forward AUC: 0.79 (gap-18) to 0.81 (gap-12) across label gaps, n=165, out-of-sample 2011–2024 — one recession in the out-of-sample window; not a validated discrimination claim.
—no default-forecasting skill is claimed; replaced by the disclosed-impairment score
Target leakage: a feature echoed the label, so the model was reading the answer out of its own inputs. The leak-free re-baseline of 2026-07-18 reads about 0.57, so we stopped selling a PD and serve a disclosed-impairment score with the out-of-filer grade above.
—no replacement figure; no forecasting-skill claim is made
A concurrent-nowcast artifact. A corrected version does not beat persistence, so we make no forecasting-skill claim for market cycles at all.
0.6137as of 2026-07-23WTI crude (CL), 5 of 22 markets pass
Retired when a point-in-time re-run across 22 markets passed only 5. The surviving headline is WTI crude at 0.6137 (CI90 0.5341–0.6933, n=140).
Every grade on this page comes from held-out, out-of-time or out-of-sample evaluation. Served predictions are logged only on the instance that answered and are not retained, and outcomes are not yet recorded, so the accuracy page is a validation record, not a track record. Nothing here measures how often we have been right in use.
served predictions are written only to the answering instance's log, which is not retained
so no signal is graded on results in use
published validation record
The record is structured against SR 26-2 section VII sound practices. Those practices are non-enforceable by their own terms, and we claim no compliance with them.
Held-out and out-of-time evaluation is not the same thing as having been right in use over time. Outcomes are not yet recorded.
The private-credit output is concurrent classification with no horizon, and the NSS figure is a repricing measurement. Neither tells you what happens next.
Latavis is an analytics and research platform. Nothing here is investment, legal, tax, or accounting advice. Latavis is not registered as an investment adviser, broker-dealer, commodity trading advisor, or credit rating agency. Forecast-skill statistics describe historical model performance against stated benchmarks over stated windows; they are not simulated or actual trading returns, and past accuracy does not guarantee future accuracy.
You have now seen the grades, the verification levels and the four withdrawals. The price is published on the next page.
SHA-256 eb10c5e9f86454a4a3faaf066a1ca7b752297b222751832dd7f5edb4d17ad3c7
Created 2026-10-14 08:39:27 UTC. Timestamped 2026-10-14 08:41:28 UTC (RFC 3161, DigiCert time-stamping authority).
What it is.A hash of one file that fixes, before the Q3 2026 filings arrive, a rule that estimates a late-filing BDC's reported mark on a borrower from the marks early filers have already reported for the same borrower; the baseline forecasts made before any early filer reported, which are not expected to beat simply carrying forward the previous mark; the grading script; and the rule for declaring failure.
What it is not. Not a result: nothing has been graded. Not a prediction of defaults or credit events. Not investment advice.
Graded after late filers report. If the rule does not reduce error against carrying forward the previous mark, that will be published here as a failure.