Public scorecard · accuracy & provenance

Audit us first.

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 written to a provenance ledger whose retention status this page reads from the served scorecard, and outcomes are not yet recorded, so the accuracy page is a validation record, not a track record.

Read the retired claims Request access

signals on the record, each badged with its level · 76 graded above L0

accuracy claims withdrawn, each with its cause

outcomes graded — a validation record, not a track record

2026-11-09

date of the published validation record

The ladder, and the caps it enforces.

Confidence scores in the served envelope are capped by validation grade. A figure is never served with more confidence than its level allows.

L0 · Illustrative

0.30confidence cap

Sample, synthetic or scaffold data with no fitted model behind it. Shown only to describe a structure, never as a basis for a decision.

L1 · Real · Unvalidated

0.50confidence cap

Real underlying data and a fitted model, but no validation gate has been passed yet.

L2 · In-Sample

0.70confidence cap

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.

L3 · OOS Validated

0.95confidence cap

Passed a held-out, point-in-time out-of-sample test on its recorded date. The level describes that record.

The scorecard.

Each signal with the grade it earned rather than the best number available.

Validation record

loading the validation record…

Served predictions are written to a provenance ledger whose retention status this page reads from the served scorecard, and outcomes are not yet recorded, so the accuracy page is a validation record, not a track record.

Validation record

Signal

What it measures

Headline grade

Level

Private Credit · 5 rows on the record · open the Private Credit page

PC · Disclosed-impairment score

L2 — a measured grade at L2, not a pre-registered forward claim. Grade as of 2026-11-09. Private credit · disclosed-impairment score (archived: the borrower score served now is the rebuilt one). Out-of-filer grade: leave-one-BDC-out AUC 0.5892 [95% CI 0.53, 0.65] on the filer's own disclosed non-accrual label (97 events) and 0.6916 [95% CI 0.67, 0.72] on a wider label that also counts a mark below 0.85× amortized cost (545 events). It ranks disclosed impairment; it is not a validated probability of default. It scored 2,735 of 3,065 borrowers and refused 330 (10.8%) by name. Panel as of 2026-09-30 filings (GBDC 2026-06-30) (filed 2026-08-03 to 2026-11-06), re-measured 2026-11-09 — the latest quarter on file for every filer: period ended 2026-09-30 for 13 of 14, and 2026-06-30 for GBDC, whose fiscal year ends 2026-09-30 and whose annual report for it is not yet filed. Scores describe that panel, not today.

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

LOFO AUC 0.59 [0.53, 0.65] strict, 97 events / 0.69 [0.67, 0.72] union, 545 events (leave-one-filer-out; strict on 97 non-accrual events, union on 545 events)

In-Sample

PCLEAD · Co-lender mark lead (next quarter)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-22: Forecast error vs no change when a BDC mark is moved toward its co-lenders' marks — 1.025x (worse), 90% CI of the improvement -0.043 to -0.025 pp; no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-22, never as current

as of 2026-10-22

1.025x

Real · Unvalidated

PCMD · First mark below 90% of cost -> below 80% within two quarters

L3 — passing out of sample; live grading begins once the forecast is issued and its outcome settles. Result as of 2026-10-22: PASS: 43 of 148 first crossings (29.0%) went below 80% within two quarters vs 2.4% of other at-risk borrowers; Brier skill +0.038 vs the base rate (90% CI +0.024 to +0.051), n 9,694, origins 2024-06-30..2025-09-30

verdict as of 2026-10-22 · current until 2027-01-22 (settles every quarter)

Calibration on 9,694 settled outcomes: Brier 0.0262 vs benchmark 0.0272 (skill score 0.038: better than the benchmark) · mean calibration gap 0.096

reliability: forecast vs what happened

bin	n	forecast	happened
2024-06-30	15	17.6%	20.0%
2024-09-30	20	21.6%	55.0%
2024-12-31	46	21.4%	19.6%
2025-03-31	22	26.0%	31.8%
2025-06-30	23	24.5%	17.4%
2025-09-30	22	25.2%	40.9%

as of 2026-10-22

+0.038

OOS Validated

PCEW · Co-lenders' mark -> will a BDC first file the borrower on non-accrual within two quarters?

L3 — passing out of sample; live grading begins once the forecast is issued and its outcome settles. Result as of 2026-11-03: PASS: AUC 0.892 (90% CI 0.860-0.919) on 13,021 borrower-quarters (1,658 borrowers, 155 first non-accruals, origins 2023Q1..2025Q4) vs 'already below 90%' 0.747 (difference 90% CI +0.114..+0.179) and spread 0.559 (difference +0.263..+0.401); 2025 holdout AUC 0.8906 on 76 events; base rate 1.19%

verdict as of 2026-11-03 · current until 2027-02-03 (settles every quarter)

as of 2026-11-03

0.892

OOS Validated

PCSC · Rebuilt impairment score -> will this lender first file its position on non-accrual within two quarters?

L1 — honest negative — no accuracy claim. Result as of 2026-11-04: FAIL: out-of-time AUC 0.863 (90% CI 0.816-0.902) on 29,688 2025 lender positions (290 first non-accruals at the same lender, 2,081 borrowers, 148 lenders) - LOSES to the position's own mark alone 0.926 (difference 90% CI -0.112..-0.021); spread alone 0.664; trained on 30,987 rows 2023Q1-2024Q3

verdict as of 2026-11-04 · current until 2027-02-04 (settles every quarter)

as of 2026-11-04

0.863

Real · Unvalidated

Private Credit: each claim's horizon, baseline and out-of-sample result

Private Credit · the block on its page

In-SamplePCDisclosed-impairment scoreas of 2026-11-09

L2 — a measured grade at L2, not a pre-registered forward claim

Result as of 2026-11-09: ARCHIVED GRADE: since VERT-PC-3 this 14-BDC in-sample classifier scores no borrower on any route; the workspace card, /predict and the Excel impairment-score cell serve PCSC, the score rebuilt out of time (row PCSC, L1, beside the lender's own-mark baseline). Out-of-filer grade: leave-one-BDC-out AUC 0.5892 on the filer's own disclosed non-accrual label and 0.6916 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. The model scores 2,735 of the 3,065 borrowers in the panel and refuses 330 — 10.8% — by name: a refused borrower is a holding a lending BDC reports in its own 1940 Act control or affiliate section (its own fund, joint venture or controlled company), or one whose 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 2026-09-30 filings (GBDC as of 2026-06-30; filed 2026-08-03 to 2026-11-06), 7,639 tranches over 3,065 borrowers from 14 BDC filers. Panel breadth: 16 filers attempted, 14 in the pinned canon, 14 contributing a position. The parser fixes ALSO recover TCPC and ORCC (302 and 352 rows staged 2026-08-28); admitting them is a queued membership decision, not a silent join — ORCC is OBDC's pre-rename ticker and may be the same filer twice. Concurrent impairment classification — no forward window.

L2 on the validation record (as of 2026-11-09) · LOFO AUC 0.59 [0.53, 0.65] strict, 97 events / 0.69 [0.67, 0.72] union, 545 events (leave-one-filer-out; strict on 97 non-accrual events, union on 545 events) · no accuracy claim

Real · UnvalidatedPCLEADCo-lender mark lead (next quarter)as of 2026-10-22

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-22: Forecast error vs no change when a BDC mark is moved toward its co-lenders' marks — 1.025x (worse), 90% CI of the improvement -0.043 to -0.025 pp; no accuracy claim

L1 on the validation record (as of 2026-10-22) · 1.025x · no accuracy claim

OOS ValidatedPCMDFirst mark below 90% of cost -> below 80% within two quartersas of 2026-10-22

L3 — passing out of sample; live grading begins once the forecast is issued and its outcome settles

When the lenders that held a borrower last quarter and this quarter move from all marking it at or above 90% of cost to at least one marking it below 90%, the borrower is more likely than the base rate to be marked below 80% of cost by one of those lenders within the next two quarterly schedules.

Horizon: the next two quarter-end schedules of investments. Baseline: the base rate over all at-risk borrowers (training origins only). Settled by: SEC BDC Data Sets: any continuing lender's fair value / cost below 0.80 on the borrower at t+1 or t+2 = 1; reported and none below = 0; reported by none = void.

Result as of 2026-10-22: PASS: 43 of 148 first crossings (29.0%) went below 80% within two quarters vs 2.4% of other at-risk borrowers; Brier skill +0.038 vs the base rate (90% CI +0.024 to +0.051), n 9,694, origins 2024-06-30..2025-09-30

L3 on the validation record (as of 2026-10-22) · +0.038

Current forecast: served on each entity's own page (as-of and forecast status with it)

OOS ValidatedPCEWCo-lenders' mark -> will a BDC first file the borrower on non-accrual within two quarters?as of 2026-11-03

L3 — passing out of sample; live grading begins once the forecast is issued and its outcome settles

For a borrower that at least two BDCs lend to, one minus the median of the co-lenders' own marks (fair value / cost) this quarter ranks which borrowers a BDC first files on non-accrual within the next two quarters, better than 'already below 90% of cost', better than the borrower's spread and better than the base rate (AUC, borrower block bootstrap).

Horizon: the next two quarter-ends' 10-Q / 10-K schedules. Baseline: 'already below 90% of cost' (median co-lender mark < 0.90), the weighted spread, the base rate. Settled by: each BDC's own 10-K/10-Q non-accrual markers (SEC EDGAR, pc_nonaccrual): the borrower's first filed non-accrual quarter at any lender.

Result as of 2026-11-03: PASS: AUC 0.892 (90% CI 0.860-0.919) on 13,021 borrower-quarters (1,658 borrowers, 155 first non-accruals, origins 2023Q1..2025Q4) vs 'already below 90%' 0.747 (difference 90% CI +0.114..+0.179) and spread 0.559 (difference +0.263..+0.401); 2025 holdout AUC 0.8906 on 76 events; base rate 1.19%

L3 on the validation record (as of 2026-11-03) · 0.892

Current forecast: served on each entity's own page (as-of and forecast status with it)

Real · UnvalidatedPCSCRebuilt impairment score -> will this lender first file its position on non-accrual within two quarters?as of 2026-11-04

L1 — honest negative — no accuracy claim

A logistic score of a lender's own debt position on the loan terms the SEC BDC Data Sets carry (own mark, the other lenders' median mark, co-lender count, spread, PIK, lien, maturity), trained 2023Q1-2024Q3 and tested 2025, ranks which positions that same lender first files on non-accrual within two quarters better than the position's own mark alone (AUC).

Horizon: the next two quarter-ends' 10-Q / 10-K schedules. Baseline: the position's own mark alone (1 - fair value / cost); spread alone reported. Settled by: the same lender's own 10-K/10-Q non-accrual marker on the borrower (pc_nonaccrual).

Result as of 2026-11-04: FAIL: out-of-time AUC 0.863 (90% CI 0.816-0.902) on 29,688 2025 lender positions (290 first non-accruals at the same lender, 2,081 borrowers, 148 lenders) - LOSES to the position's own mark alone 0.926 (difference 90% CI -0.112..-0.021); spread alone 0.664; trained on 30,987 rows 2023Q1-2024Q3

L1 on the validation record (as of 2026-11-04) · 0.863 · no accuracy claim

Macro & ESG · 6 rows on the record · open the Macro & ESG page

MA6 · Recession probit

L1 — a measured grade at L1, not a pre-registered forward claim. US recession probability · RETIRED 2026-11-01 and not shown: the model locked on 2026-10-15 is worse than the long-run rate on its settled outcomes and reads licence-held inputs; its licence-clean replacement (yield curve, the Fed Board's excess bond premium, unemployment change) failed its pre-registered test. Retired model's AUC across label gaps, n=165, out-of-sample 2011–2024 — one recession in the out-of-sample window; not a validated discrimination claim. The earlier 0.862 was retired on 2026-07-20. Research record — not offered in the product today.

a measured grade with no declared settlement cadence: it reads as of 2026-09-01, never as current

Calibration on 161 settled outcomes: Brier 0.1763 vs benchmark 0.0844 (skill score -1.089: WORSE than the benchmark) · mean calibration gap 0.260

Point-in-time check: on the data as first published (ALFRED vintages, 173 months 2011-05–2025-09) the AUC is 0.836, not 0.863: revisions add 0.027. The Federal Reserve's published out-of-sample benchmark is 0.88; with its excess bond premium in place of the high-yield spread the same test covers 4 recessions (1986-09–2025-09) at 0.855.

reliability: forecast vs what happened

bin	n	forecast	happened
0.0–1.6%	16	0.3%	0.0%
1.6–5.7%	16	3.8%	0.0%
5.7–11.4%	17	8.4%	0.0%
11.4–15.5%	15	13.8%	0.0%
15.5–21.4%	16	17.9%	0.0%
21.4–29.5%	17	24.5%	0.0%
29.5–46.5%	16	38.1%	0.0%
46.5–66.4%	15	58.1%	33.3%
66.4–89.3%	16	80.3%	37.5%
89.3–100.0%	17	94.0%	11.8%

as of 2026-09-01

0.79 (gap-18) to 0.81 (gap-12)

Real · Unvalidated

MA2 · Taylor-rule policy direction (6m)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-20: Hit rate of the 6-month cut/hold/hike call — does NOT beat persistence (0.624) or always-hold (0.462); no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

as of 2026-10-20

0.310

Real · Unvalidated

MA3 · Fiscal gap → debt direction (1y)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-09: Hit rate of the debt-ratio direction call — beats the base rate (0.525) but NOT persistence (0.668); no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

0.635

Real · Unvalidated

MAFOMC · Fed statement tone after a hold -> next FOMC decision

L1 — honest negative — no accuracy claim. Result as of 2026-10-22: FAIL: right on 68.8% of 173 statements vs 79.2% for the last-move rule (2006-01-31..2026-07-29); where they differ the tone was right 7 times and the rule 25 (McNemar p 1.00)

verdict as of 2026-10-22 · current until 2026-12-17 (settles every scheduled FOMC meeting (about eight weeks))

as of 2026-10-22

68.8%

Real · Unvalidated

MAECB · ECB introductory-statement tone after a hold -> next ECB decision

L1 — honest negative — no accuracy claim. Result as of 2026-10-30: FAIL: right on 65.4% of 318 ECB meetings vs 75.8% for the last-move rule (1999-03-04..2026-07-23); where they differ the tone was right 10 times and the rule 43 (McNemar p 1.00)

verdict as of 2026-10-30 · current until 2026-12-18 (settles every scheduled ECB meeting (about seven weeks))

as of 2026-10-30

65.4%

Real · Unvalidated

MACLM · Initial jobless claims -> will the next unemployment print rise?

L1 — honest negative — no accuracy claim. Result as of 2026-11-02: FAIL: Brier 0.2102 vs 0.2219 for the base rate (Holm p 0.02) and 0.2233 for persistence (Holm p 0.02); 434 months 1990-01..2026-05; family Holm p 0.10 across the 7 INTEL-4 claims (lower Brier is better)

verdict as of 2026-11-02 · current until 2026-12-03 (settles every month)

as of 2026-11-02

0.2102

Real · Unvalidated

Macro & ESG: each claim's horizon, baseline and out-of-sample result

Macro & ESG · the block on its page

Real · UnvalidatedMA6Recession probitas of 2026-09-01

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-09-01: RETIRED 2026-11-01, not shown on any surface: the 4-feature model locked 2026-10-15 (recession-probit-4f-L20261015) is worse than climatology on its settled outcomes and reads licence-held inputs. Its licence-clean replacement (T10Y2Y + Fed Board excess bond premium + 12m change in UNRATE) FAILED its pre-registered gate on 2026-11-09 (published at L1): 469 point-in-time months 1986-09..2025-09, 4 recessions, AUC 0.8243, Brier 0.138 vs climatology 0.158 (one-sided DM p 0.24, not significant), recession-state persistence 0.107, since 2007 Brier 0.209 vs climatology 0.155. Recursive AUC of the retired model — 0.862 retired (gate failed 2026-07-20); 2026-09-01 removed the look-ahead from the serving harness (the fit at issue month t sees only labels resolved at t)

L1 on the validation record (as of 2026-09-01) · retired, no current figure · no accuracy claim

Real · UnvalidatedMA2Taylor-rule policy direction (6m)as of 2026-10-20

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-20: Hit rate of the 6-month cut/hold/hike call — does NOT beat persistence (0.624) or always-hold (0.462); no accuracy claim

L1 on the validation record (as of 2026-10-20) · 0.310 · no accuracy claim

Real · UnvalidatedMA3Fiscal gap → debt direction (1y)as of 2026-11-09

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-09: Hit rate of the debt-ratio direction call — beats the base rate (0.525) but NOT persistence (0.668); no accuracy claim

L1 on the validation record (as of 2026-11-09) · 0.635 · no accuracy claim

Real · UnvalidatedMAFOMCFed statement tone after a hold -> the next FOMC decisionas of 2026-10-22

L1 — honest negative — no accuracy claim

The next FOMC decision repeats this one, except that after a hold a unanimous statement tone (every counted clause hawkish, or every one dovish) calls a move in that direction.

Horizon: the next scheduled FOMC meeting. Baseline: the last-move rule (the next decision repeats this one). Settled by: FRED DFEDTARU the day after the next meeting vs the day after this one (up / equal / down).

Result as of 2026-10-22: FAIL: right on 68.8% of 173 statements vs 79.2% for the last-move rule (2006-01-31..2026-07-29); where they differ the tone was right 7 times and the rule 25 (McNemar p 1.00)

L1 on the validation record (as of 2026-10-22) · 68.8% · no accuracy claim

Real · UnvalidatedMAECBECB introductory-statement tone after a hold -> the next ECB decisionas of 2026-10-30

L1 — honest negative — no accuracy claim

The next ECB decision repeats this one, except that after a hold an introductory-statement tone in the top (bottom) decile of all earlier statements calls a raise (cut).

Horizon: the next scheduled ECB monetary-policy meeting. Baseline: the last-move rule (the next decision repeats this one). Settled by: the ECB's policy-rate table at the next meeting (up / equal / down).

Result as of 2026-10-30: FAIL: right on 65.4% of 318 ECB meetings vs 75.8% for the last-move rule (1999-03-04..2026-07-23); where they differ the tone was right 10 times and the rule 43 (McNemar p 1.00)

L1 on the validation record (as of 2026-10-30) · 65.4% · no accuracy claim

Real · UnvalidatedMACLMInitial jobless claims -> will the next unemployment print rise?as of 2026-11-02

L1 — honest negative — no accuracy claim

The probability that next month's first-print unemployment rate is above this month's is forecast from the change in initial jobless claims (4 weeks vs 13-16 weeks earlier) and the latest unemployment change, better than the base rate or persistence (Brier).

Horizon: the next monthly BLS Employment Situation release. Baseline: the rolling base rate of rises, and persistence (rise frequency given the last change). Settled by: ALFRED / FRED UNRATE first print for the next month vs this month at that vintage.

Result as of 2026-11-02: FAIL: Brier 0.2102 vs 0.2219 for the base rate (Holm p 0.02) and 0.2233 for persistence (Holm p 0.02); 434 months 1990-01..2026-05; family Holm p 0.10 across the 7 INTEL-4 claims (lower Brier is better)

L1 on the validation record (as of 2026-11-02) · 0.2102 · no accuracy claim

Manufacturing · 3 rows on the record · open the Manufacturing page

MF7 · Industrial nowcast

L1 — a measured grade at L1, not a pre-registered forward claim. Industrial-production nowcast: +3.6% RMSE against the prevailing mean, not significant (DM p 0.45) — an honest negative, no accuracy claim.

a measured grade with no declared settlement cadence: it reads as of 2026-10-22, never as current

as of 2026-10-22

+3.6%

Real · Unvalidated

MFREV · Revision of the advance industrial-production print

L1 — honest negative — no accuracy claim. Result as of 2026-10-30: FAIL: mean absolute error 0.1767 pp vs 0.1769 for no revision (Holm p 0.16) and 0.1790 for the rolling mean revision (Holm p 0.0052), 318 months 1999-12..2026-05: the edge over no revision is not significant

verdict as of 2026-10-30 · current until 2026-11-30 (settles every month)

as of 2026-10-30

0.1767 pp

Real · Unvalidated

MFEMP · Empire State + factory orders -> will the next industrial-production print be up?

L1 — honest negative — no accuracy claim. Result as of 2026-11-02: FAIL: Brier 0.2368 vs 0.2358 for the base rate (Holm p 0.77) and 0.2406 for persistence (Holm p 0.77); 235 months 2006-08..2026-06; family Holm p 1.00 across the 7 INTEL-4 claims (lower Brier is better)

verdict as of 2026-11-02 · current until 2026-12-03 (settles every month)

as of 2026-11-02

0.2368

Real · Unvalidated

Manufacturing: each claim's horizon, baseline and out-of-sample result

Manufacturing · the block on its page

Real · UnvalidatedMF7Industrial nowcastas of 2026-10-22

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-22: RMSE skill vs the prevailing mean (the stronger rolling baseline) — not significant (DM p 0.45); no accuracy claim

L1 on the validation record (as of 2026-10-22) · +3.6% · no accuracy claim

Real · UnvalidatedMFREVRevision of the advance industrial-production printas of 2026-10-30

L1 — honest negative — no accuracy claim

The second print of INDPRO growth = the advance print + beta x the revision the advance release made to the previous month (beta fitted on earlier settled months).

Horizon: the next G.17 release (about one month). Baseline: no revision, and the rolling mean revision. Settled by: ALFRED / FRED INDPRO at the next vintage (growth within the vintage).

Result as of 2026-10-30: FAIL: mean absolute error 0.1767 pp vs 0.1769 for no revision (Holm p 0.16) and 0.1790 for the rolling mean revision (Holm p 0.0052), 318 months 1999-12..2026-05: the edge over no revision is not significant

L1 on the validation record (as of 2026-10-30) · 0.1767 pp · no accuracy claim

Real · UnvalidatedMFEMPEmpire State survey + factory orders -> will the next industrial-production print be up?as of 2026-11-02

L1 — honest negative — no accuracy claim

On the 20th of the month, the probability that the month's first-print industrial production growth is positive is forecast from the Empire State survey (level and change), M3 new orders growth and the last IP print, better than the base rate or persistence (Brier).

Horizon: the next G.17 release (about one month). Baseline: the rolling base rate of positive prints, and persistence (given the last print's sign). Settled by: Federal Reserve G.17 first print (ALFRED INDPRO) for the month.

Result as of 2026-11-02: FAIL: Brier 0.2368 vs 0.2358 for the base rate (Holm p 0.77) and 0.2406 for persistence (Holm p 0.77); 235 months 2006-08..2026-06; family Holm p 1.00 across the 7 INTEL-4 claims (lower Brier is better)

L1 on the validation record (as of 2026-11-02) · 0.2368 · no accuracy claim

Commodities · 5 rows on the record · open the Commodities page

COT · Commodities positioning

L1 — a measured grade at L1, not a pre-registered forward claim. Commodity futures positioning (CFTC COT). Headline market WTI crude (CL); 5 of 22 markets tested passed the point-in-time gauntlet (CL, CT, HG, ZM, ZS; record 2026-07-23), but re-graded on 2026-10-22 against rolling baselines with the test counted over all 22, none beats a coin: no accuracy claim, positioning served as data. L1 on the validation record (as of 2026-10-22).

a measured grade with no declared settlement cadence: it reads as of 2026-10-22, never as current

Calibration: — not in the calibration record: the COT gauntlet artifact carries per-market Brier, not the (forecast, outcome) pairs a reliability curve needs

as of 2026-10-22

0.6137 (CI90 0.5341–0.6933, n=140)

Real · Unvalidated

CM4 · Commodities seasonal base rate

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-22: Walk-forward AUC — SRW wheat (ZW) headline; 4 of 22 markets passed the original gate (KE, PL, ZC, ZW) but none beats a coin or the running base rate on Brier once all 22 are counted (ZW skill +3.5%, DM p 0.07, Holm-adjusted 1.0). Claim: P(next month's average price > this month's), 1 month, settled on the same series — no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-22, never as current

Calibration: — not in the calibration record: the seasonal gauntlet carries AUC per market, not the (forecast, outcome) pairs a reliability curve needs

as of 2026-10-22

0.615

Real · Unvalidated

CMLIVE · COT probabilities of the five AUC-gate markets, settled forward from deploy

L1 — settles forward from deploy — no accuracy claim yet. Result as of 2026-10-22: PENDING: 0 months settled since registration; decision at 60 settled market-months

verdict as of 2026-10-22 · current until 2026-11-22 (settles every month)

Calibration: — no settled live outcome yet (live grading begins once the forecast is issued and its outcome settles), so there is nothing to calibrate on

as of 2026-10-22

0 of 60

Real · Unvalidated

CMVOL · Next month's move size per commodity (EWMA volatility)

L1 — honest negative — no accuracy claim. Result as of 2026-11-03: FAIL: mean QLIKE -4.332 vs -4.183 (last12, Holm p 0.0025) and -4.283 (expanding, Holm p 0.09); 20 markets, 321 months 2000-01..2026-09 (lower is better)

verdict as of 2026-11-03 · current until 2026-12-04 (settles every month)

as of 2026-11-03

-4.332 vs -4.283

Real · Unvalidated

NGVOL · Henry Hub next month's realized variance from daily spot (HAR)

L1 — honest negative — no accuracy claim. Result as of 2026-11-09: FAIL: mean QLIKE -1.271 vs -1.302 (last 12 months) / -1.234 (all history), lower is better; one-sided Newey-West DM Holm p 0.644 / 0.644 on 296 test months 2002-02..2026-09; it is worse than the last-12-month mean; no accuracy claim

verdict as of 2026-11-09 · current until 2026-12-10 (settles every month)

as of 2026-11-09

-1.271 vs -1.302

Real · Unvalidated

Commodities: each claim's horizon, baseline and out-of-sample result

Commodities · the block on its page

Real · UnvalidatedCOTCommodities positioningas of 2026-10-22

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-22: OOS AUC — CL (WTI) headline; 5 of 22 markets passed the AUC gate (CL, CT, HG, ZM, ZS) but none beats a coin or the running up-share on Brier once all 22 are counted (CL skill +3.9%, DM p 0.19; best ZM p 0.07; Holm-adjusted 1.0) — no accuracy claim

L1 on the validation record (as of 2026-10-22) · 0.6137 (CI90 0.5341–0.6933, n=140) · no accuracy claim

Real · UnvalidatedCM4Commodities seasonal base rateas of 2026-10-22

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-22: Walk-forward AUC — SRW wheat (ZW) headline; 4 of 22 markets passed the original gate (KE, PL, ZC, ZW) but none beats a coin or the running base rate on Brier once all 22 are counted (ZW skill +3.5%, DM p 0.07, Holm-adjusted 1.0). Claim: P(next month's average price > this month's), 1 month, settled on the same series — no accuracy claim

L1 on the validation record (as of 2026-10-22) · 0.615 · no accuracy claim

Real · UnvalidatedCMLIVECOT positioning probabilities of the five AUC-gate markets, settled forward from deployas of 2026-10-22

L1 — settles forward from deploy — no accuracy claim yet

For CL, CT, HG, ZM and ZS, the COT positioning probability that next month's price label rises beats the up-month base rate known at issue (Brier), on months settling after 2026-10-22.

Horizon: the next calendar month's price label. Baseline: the up-month base rate of the label history known at issue, and a coin. Settled by: 1 if next month's label is above this month's (FRED MCOILWTICO for CL; World Bank Pink Sheet).

Result as of 2026-10-22: PENDING: 0 months settled since registration; decision at 60 settled market-months

L1 on the validation record (as of 2026-10-22) · 0 of 60 · no accuracy claim

Real · UnvalidatedCMVOLNext month's move size per commodity (volatility clustering)as of 2026-11-03

L1 — honest negative — no accuracy claim

Next month's squared log price change of each commodity market is forecast by the EWMA (lambda 0.94) of its earlier monthly squared changes, better than its last-12-month or all-history mean (QLIKE).

Horizon: the next calendar month's print of the market's monthly series. Baseline: the rolling 12-month mean and the expanding mean of squared monthly changes. Settled by: the same monthly series (World Bank Pink Sheet / FRED): squared log change into the month.

Result as of 2026-11-03: FAIL: mean QLIKE -4.332 vs -4.183 (last12, Holm p 0.0025) and -4.283 (expanding, Holm p 0.09); 20 markets, 321 months 2000-01..2026-09 (lower is better)

L1 on the validation record (as of 2026-11-03) · -4.332 vs -4.283 · no accuracy claim

Real · UnvalidatedNGVOLHenry Hub daily volatility (HAR) -> next month's realized varianceas of 2026-11-09

L1 — honest negative — no accuracy claim

At the end of each month, a HAR model of the daily Henry Hub spot's realized variance (last day, last week, last month; refit on data through the month) forecasts next month's realized variance better than the last 12 months' mean and the all-history mean (QLIKE, one-sided Newey-West DM, Holm over the two).

Horizon: the next calendar month's daily returns. Baseline: the mean realized variance of the last 12 months; the all-history mean. Settled by: the sum of squared daily log returns of the Henry Hub spot (EIA via FRED DHHNGSP) in that month.

Result as of 2026-11-09: FAIL: mean QLIKE -1.271 vs -1.302 (last 12 months) / -1.234 (all history), lower is better; one-sided Newey-West DM Holm p 0.644 / 0.644 on 296 test months 2002-02..2026-09; it is worse than the last-12-month mean; no accuracy claim

L1 on the validation record (as of 2026-11-09) · -1.271 vs -1.302 · no accuracy claim

Fixed Income · 8 rows on the record · open the Fixed Income page

FI · Treasury curve repricing (held-out tenors)

L1 — a measured grade at L1, not a pre-registered forward claim. Fixed income · NSS curve repricing, 5-year held-out-tenor error. A repricing measurement, not a forecast.

a measured grade with no declared settlement cadence: it reads as of 2026-11-03, never as current

as of 2026-11-03

6.1 bps

Real · Unvalidated

FI10 · Treasury 10Y VaR coverage (1-day 99%)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-22: Share of days the next-day loss exceeded the served VaR — target 1%; coverage is calibrated (Kupiec, Christoffersen pass) but on the quantile loss it does not beat a rolling 250-day normal VaR (skill −0.9%, DM p 0.70); no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-22, never as current

as of 2026-10-22

1.09%

Real · Unvalidated

FI30 · Treasury 30Y VaR coverage (1-day 99%)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-22: Share of days the next-day loss exceeded the served VaR — target 1%; coverage is calibrated but on the quantile loss it does not beat a rolling 250-day normal VaR (skill −2.6%, DM p 0.88); no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-22, never as current

as of 2026-10-22

1.03%

Real · Unvalidated

FIVAR · Treasury 1Y/2Y/5Y VaR coverage

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-20: Exceedance rate vs the 1% target — FAILS (1Y and 2Y too many, 5Y clustered); no accuracy claim, served at L1

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

as of 2026-10-20

1.27% / 1.46% / 1.22%

Real · Unvalidated

FISPRD · Corporate credit spread change forecast (1m, GZ spread)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-03: MAE skill vs a rolling-mean baseline — HONEST NEGATIVE: not significant (RMSE skill +0.009, Diebold-Mariano p 0.21, direction AUC 0.55); credit spreads are served as observations only

a measured grade with no declared settlement cadence: it reads as of 2026-11-03, never as current

as of 2026-11-03

+0.016

Real · Unvalidated

FIDAILY · One-day 99% Treasury VaR exceptions, settled daily

L1 — honest negative — no accuracy claim. Result as of 2026-10-22: FAIL: 10Y 1.09% and 30Y 1.03% exceptions are calibrated, but no tenor beats a rolling 250-day normal VaR on the quantile loss (Holm-adjusted p 1.0); live daily record pending

verdict as of 2026-10-22 · not current: tested 2026-10-22; the claim settles every day, so outcomes settled since 2026-10-23 are not in this verdict - it is the verdict as of 2026-10-22, re-graded when live grading begins once the forecast is issued and its outcome settles

as of 2026-10-22

1.09% / 1.03%

Real · Unvalidated

FITP · Term-premium change -> 10Y direction over 3 months

L1 — honest negative — no accuracy claim. Result as of 2026-10-30: FAIL: the 10Y moved with the term-premium change in 46.2% of 143 quarters (1990-03..2026-06); Holm p 1.00 vs a coin, 1.00 vs yield momentum (45.5%), 1.00 vs the rolling majority (49.0%)

verdict as of 2026-10-30 · current until 2026-11-30 (settles every month)

as of 2026-10-30

46.2%

Real · Unvalidated

FITVOL · Next month's realised volatility of the 10-year Treasury yield (HAR)

L3 — passing out of sample; live grading begins once the forecast is issued and its outcome settles. Result as of 2026-11-02: PASS: mean QLIKE 7.465 vs 7.536 (last month's RV, Holm p < 0.0001) and 7.497 (last 12 months' mean RV, Holm p 0.0019); 441 months 1990-01..2026-09; family Holm p 0.01 across the 7 INTEL-4 claims (lower is better)

verdict as of 2026-11-02 · current until 2026-12-03 (settles every month)

as of 2026-11-02

7.465 vs 7.536

OOS Validated

Fixed Income: each claim's horizon, baseline and out-of-sample result

Fixed Income · the block on its page

Real · UnvalidatedFITreasury curve repricing (held-out tenors)as of 2026-11-03

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-03: Held-out-tenor MAE of the served curve vs linear interpolation, six tenors of the Treasury par file pooled — repricing measurement, not a forecast

L1 on the validation record (as of 2026-11-03) · 6.1 bps · no accuracy claim

Real · UnvalidatedFI10Treasury 10Y VaR coverage (1-day 99%)as of 2026-10-22

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-22: Share of days the next-day loss exceeded the served VaR — target 1%; coverage is calibrated (Kupiec, Christoffersen pass) but on the quantile loss it does not beat a rolling 250-day normal VaR (skill −0.9%, DM p 0.70); no accuracy claim

L1 on the validation record (as of 2026-10-22) · 1.09% · no accuracy claim

Real · UnvalidatedFI30Treasury 30Y VaR coverage (1-day 99%)as of 2026-10-22

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-22: Share of days the next-day loss exceeded the served VaR — target 1%; coverage is calibrated but on the quantile loss it does not beat a rolling 250-day normal VaR (skill −2.6%, DM p 0.88); no accuracy claim

L1 on the validation record (as of 2026-10-22) · 1.03% · no accuracy claim

Real · UnvalidatedFIVARTreasury 1Y/2Y/5Y VaR coverageas of 2026-10-20

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-20: Exceedance rate vs the 1% target — FAILS (1Y and 2Y too many, 5Y clustered); no accuracy claim, served at L1

L1 on the validation record (as of 2026-10-20) · 1.27% / 1.46% / 1.22% · no accuracy claim

Real · UnvalidatedFISPRDCorporate credit spread change forecast (1m, GZ spread)as of 2026-11-03

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-03: MAE skill vs a rolling-mean baseline — HONEST NEGATIVE: not significant (RMSE skill +0.009, Diebold-Mariano p 0.21, direction AUC 0.55); credit spreads are served as observations only

L1 on the validation record (as of 2026-11-03) · +0.016 · no accuracy claim

Real · UnvalidatedFIDAILYOne-day 99% Treasury VaR exceptions per tenor, settled dailyas of 2026-10-22

L1 — honest negative — no accuracy claim

The next day's loss on a par Treasury exceeds the served 99% VaR on 1% of days, independently, and the VaR is a better 1% quantile forecast than a rolling normal VaR.

Horizon: the next business day. Baseline: rolling normal VaR, 500-day and 250-day. Settled by: the loss implied by the next FRED DGS<T> print (duration + convexity at today's yield).

Result as of 2026-10-22: FAIL: 10Y 1.09% and 30Y 1.03% exceptions are calibrated, but no tenor beats a rolling 250-day normal VaR on the quantile loss (Holm-adjusted p 1.0); live daily record pending

L1 on the validation record (as of 2026-10-22) · 1.09% / 1.03% · no accuracy claim

Real · UnvalidatedFITPTerm-premium change -> 10Y yield direction over 3 monthsas of 2026-10-30

L1 — honest negative — no accuracy claim

Over the next 3 months the 10Y yield moves in the direction of last month's Kim-Wright term-premium change.

Horizon: three calendar months. Baseline: a coin, yield momentum, and the rolling majority direction. Settled by: FRED DGS10 month-end three months on.

Result as of 2026-10-30: FAIL: the 10Y moved with the term-premium change in 46.2% of 143 quarters (1990-03..2026-06); Holm p 1.00 vs a coin, 1.00 vs yield momentum (45.5%), 1.00 vs the rolling majority (49.0%)

L1 on the validation record (as of 2026-10-30) · 46.2% · no accuracy claim

OOS ValidatedFITVOLNext month's realised volatility of the 10-year Treasury yield (HAR)as of 2026-11-02

L3 — passing out of sample; live grading begins once the forecast is issued and its outcome settles

Next calendar month's realised variance of daily 10-year Treasury yield changes is forecast by a recursive HAR regression on last month's, last quarter's and last year's realised variance, better than last month's or the last 12 months' realised variance (QLIKE).

Horizon: the next calendar month. Baseline: last month's realised variance and the last 12 months' mean realised variance. Settled by: FRED DGS10 daily values of the month: sum of squared daily changes in basis points.

Result as of 2026-11-02: PASS: mean QLIKE 7.465 vs 7.536 (last month's RV, Holm p < 0.0001) and 7.497 (last 12 months' mean RV, Holm p 0.0019); 441 months 1990-01..2026-09; family Holm p 0.01 across the 7 INTEL-4 claims (lower is better)

L3 on the validation record (as of 2026-11-02) · 7.465 vs 7.536

Current forecast: served on each entity's own page (as-of and forecast status with it)

Equities · 5 rows on the record · open the Equities page

EQR · S&P 500 return forecast (1m)

L1 — a measured grade at L1, not a pre-registered forward claim. Equities · next-month S&P 500 return forecast. Loses to a rolling-mean forecast (MAE skill −0.012) and is not significantly better than AR(1) (DM p = 0.645, 75 months). No accuracy claim; not served as a forecast.

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

as of 2026-10-20

-0.012

Real · Unvalidated

EQO · 13F ownership breadth (1q)

L1 — a measured grade at L1, not a pre-registered forward claim. Equities · 13F institutional-breadth change vs the next quarter's return. Inconclusive: tradeable rank IC −0.008 over 20 quarters (t −1.22, not significant), no better than a random placebo, on a small historical filer panel. No accuracy claim; the breadth figure is served as data.

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

as of 2026-10-20

-0.008

Real · Unvalidated

EQHF · 13F holders fell + fails rose -> next-quarter underperformance

L1 — honest negative — no accuracy claim. Result as of 2026-10-22: FAIL: 50.5% of 30,431 flagged security-quarters finished below the median (90% CI 49.3% to 51.7%); mean relative return +1.29% (90% CI +0.46% to +2.26%); 27 origins 2019-06-30..2025-12-31

verdict as of 2026-10-22 · current until 2027-01-22 (settles every quarter)

as of 2026-10-22

50.5%

Real · Unvalidated

EQCRWD · 13F crowding -> next-quarter fails-to-deliver rise

L1 — honest negative — no accuracy claim. Result as of 2026-10-30: FAIL: crowded securities' fails rose in 43.9% of cases vs 47.8% for the rest - the opposite of the claim (lift -3.9 pp, 90% CI -5.0 to -2.9; 2 of 27 origins positive), 18,802 flagged security-quarters 2019-06-30..2025-12-31

verdict as of 2026-10-30 · current until 2027-01-30 (settles every quarter)

as of 2026-10-30

-3.9 pp

Real · Unvalidated

EQINS · Insider open-market buying -> will a stock gain 13F holders next quarter?

L1 — honest negative — no accuracy claim. Result as of 2026-11-02: FAIL: right on 57.9% of 107,774 issuer-quarters (44 origins 2015Q2..2026Q1) - the same as the rolling base rate (57.9%) and the no-insider logistic (57.9%): the model called a rise for every stock at every origin; persistence 53.3% (Holm p < 0.0001), rolling majority 57.1% (Holm p 0.18); claim p 1.00, family Holm p 1.00 across the 5 equities trials

verdict as of 2026-11-02 · current until 2027-02-02 (settles every quarter)

as of 2026-11-02

57.9%

Real · Unvalidated

Equities: each claim's horizon, baseline and out-of-sample result

Equities · the block on its page

Real · UnvalidatedEQRS&P 500 return forecast (1m)as of 2026-10-20

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-20: MAE skill vs the rolling-mean forecast — loses to it; +0.006 vs AR(1), not significant (DM p=0.645); no accuracy claim

L1 on the validation record (as of 2026-10-20) · -0.012 · no accuracy claim

Real · UnvalidatedEQO13F ownership breadth (1q)as of 2026-10-20

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-20: INCONCLUSIVE: tradeable (lagged) rank IC of institutional-breadth change vs the next-quarter 13F-implied return — no edge over a random placebo (-0.004) and no significant negative (t -1.22); no accuracy claim

L1 on the validation record (as of 2026-10-20) · -0.008 · no accuracy claim

Real · UnvalidatedEQHF13F holders fell and fails rose -> next-quarter underperformanceas of 2026-10-22

L1 — honest negative — no accuracy claim

A security whose number of 13F filers fell this quarter while its SEC fails-to-deliver rose finishes the first full quarter after the 13F filings are public below the cross-sectional median return (13F-implied quarter-end price).

Horizon: quarter-end Q+1 to quarter-end Q+2. Baseline: no information (half of securities finish below the median). Settled by: 13F-implied price return vs the universe median over Q+1..Q+2 (SEC 13F data sets).

Result as of 2026-10-22: FAIL: 50.5% of 30,431 flagged security-quarters finished below the median (90% CI 49.3% to 51.7%); mean relative return +1.29% (90% CI +0.46% to +2.26%); 27 origins 2019-06-30..2025-12-31

L1 on the validation record (as of 2026-10-22) · 50.5% · no accuracy claim

Real · UnvalidatedEQCRWD13F crowding (top-decile filer growth) -> next-quarter fails-to-deliver riseas of 2026-10-30

L1 — honest negative — no accuracy claim

Securities in the top decile of 13F filer-count growth see their mean daily fails rise over the first full quarter after the 13F deadline more often than the rest of the universe.

Horizon: quarter Q+2 vs Q+1. Baseline: the unflagged universe's rise rate at the same origin. Settled by: SEC fails-to-deliver files for Q+1 and Q+2.

Result as of 2026-10-30: FAIL: crowded securities' fails rose in 43.9% of cases vs 47.8% for the rest - the opposite of the claim (lift -3.9 pp, 90% CI -5.0 to -2.9; 2 of 27 origins positive), 18,802 flagged security-quarters 2019-06-30..2025-12-31

L1 on the validation record (as of 2026-10-30) · -3.9 pp · no accuracy claim

Real · UnvalidatedEQINSInsider open-market buying -> will a stock gain 13F holders next quarter?as of 2026-11-02

L1 — honest negative — no accuracy claim

For a US issuer whose insiders traded in the open market in a quarter, whether the number of 13F filers holding its stock rises or falls by the next quarter-end is forecast by a recursive logistic on the quarter's insider net-buy breadth (Form 4 purchases minus sales, per insider) and the stock's last holder-count change, better than the rolling base rate, persistence, the rolling majority and the same logistic without the insider signal.

Horizon: the next quarter-end's 13F filings (45 days after it). Baseline: the rolling base rate, persistence of the last holder-count change, the stock's rolling majority direction, and the no-insider logistic. Settled by: SEC Form 13F data sets: distinct filers whose original 13F-HR, filed by the deadline, lists the CUSIP, at quarter-end t+1 vs t.

Result as of 2026-11-02: FAIL: right on 57.9% of 107,774 issuer-quarters (44 origins 2015Q2..2026Q1) - the same as the rolling base rate (57.9%) and the no-insider logistic (57.9%): the model called a rise for every stock at every origin; persistence 53.3% (Holm p < 0.0001), rolling majority 57.1% (Holm p 0.18); claim p 1.00, family Holm p 1.00 across the 5 equities trials

L1 on the validation record (as of 2026-11-02) · 57.9% · no accuracy claim

FX · 10 rows on the record · open the FX page

FX1 · REER misalignment reversion (12m)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-20: RMSE skill vs the random walk — does NOT beat it (24m: −1.56%); valuation context, no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

as of 2026-10-20

−1.06%

Real · Unvalidated

FX2 · Crowded-carry crash flag (4w)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-20: Lift of P(adverse 2σ move | flag) — 90% CI 0.71–1.68 includes 1; the flag is not served, no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

as of 2026-10-20

1.18

Real · Unvalidated

FX3 · Positioning x carry x value composite (3m)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-21: Direction hit rate vs USD — 51.8%, binomial p 0.28, not above the majority-direction baseline; the composite is served as context, no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-21, never as current

as of 2026-10-21

51.8%

Real · Unvalidated

FXPOS · CFTC positioning extremes -> 4-week reversal vs USD

L1 — honest negative — no accuracy claim. Result as of 2026-10-22: FAIL: the currency reversed after 52.6% of 152 extremes (2009-12-29..2026-06-09), one-sided p 0.29 vs a coin; the rolling majority direction was right 52.0% (McNemar p 0.50)

verdict as of 2026-10-22 · not current: tested 2026-10-22; the claim settles every week (the weekly COT report), so outcomes settled since 2026-10-29 are not in this verdict - it is the verdict as of 2026-10-22, re-graded when live grading begins once the forecast is issued and its outcome settles

as of 2026-10-22

52.6%

Real · Unvalidated

FXVOL · Next month's realised volatility per G10 currency vs USD

L3 — passing out of sample; live grading begins once the forecast is issued and its outcome settles. Result as of 2026-10-30: PASS: mean QLIKE -6.084 vs -6.051 (last1, Holm p < 0.0001) and -6.035 (last12, Holm p 0.0053); 9 currencies, 306 months 2001-01..2026-06 (lower is better)

verdict as of 2026-10-30 · current until 2026-11-30 (settles every month)

as of 2026-10-30

-6.084 vs -6.051

OOS Validated

FXCARRY · G10 carry: long the 3 highest policy rates, short the 3 lowest, vs USD (1m)

L3 — a measured grade at L3, not a pre-registered forward claim. Measured as of 2026-11-09: Mean monthly excess return > 0 — PASSES: +0.264% a month (Sharpe 0.43 a year, max drawdown 0.35 log), one-sided Newey-West p 0.018, Holm p 0.035 over the two carry strategies; 317 months; dollar carry FAILS (p 0.147, Sharpe 0.22). Tested on history, not yet graded live

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

+0.264%/month

OOS Validated

FXVALUE · Five-year REER change sort: long the 3 cheapest, short the 3 dearest, vs USD (1m)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-09: Mean monthly excess return > 0 — FAILS: -0.009% a month (Sharpe -0.02), one-sided Newey-West p 0.54; 317 months; no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

-0.009%/month

Real · Unvalidated

FXTAYLOR · Taylor-rule differential -> next month's exchange-rate change vs a random walk

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-09: Clark-West vs the random walk — FAILS: one-sided p 0.25, MSPE ratio 1.040 (above 1 = worse than the random walk); 192 months 2010-03..2026-03, 6 currencies (CHF no gap history; AUD/NZD no monthly CPI); no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

MSPE ratio 1.040

Real · Unvalidated

FXDENSITY · Next month's exchange-rate density (EWMA-vol Normal) vs a random-walk Normal (CRPS)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-09: CRPS vs the random walk — FAILS: 0.01631 vs 0.01637 (lower is better), one-sided Newey-West DM p 0.25; 268 months 2004-02..2026-05; no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

0.01631 vs 0.01637

Real · Unvalidated

FXCRASH · VIX above its 5-year 80th percentile -> a bottom-5% month for broad carry (G10 + EM)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-09: One-sided Fisher exact — FAILS: crash rate 8.3% after a flag vs 4.1% without (lift 2.04, p 0.13; 16 events in 317 months); the flat-when-flagged overlay cut the worst drawdown 0.23 -> 0.19 log but lowered the Sharpe 0.73 -> 0.66; no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

lift 2.04

Real · Unvalidated

FX: each claim's horizon, baseline and out-of-sample result

FX · the block on its page

Real · UnvalidatedFX1REER misalignment reversion (12m)as of 2026-10-20

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-20: RMSE skill vs the random walk — does NOT beat it (24m: −1.56%); valuation context, no accuracy claim

L1 on the validation record (as of 2026-10-20) · −1.06% · no accuracy claim

Real · UnvalidatedFX2Crowded-carry crash flag (4w)as of 2026-10-20

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-20: Lift of P(adverse 2σ move | flag) — 90% CI 0.71–1.68 includes 1; the flag is not served, no accuracy claim

L1 on the validation record (as of 2026-10-20) · 1.18 · no accuracy claim

Real · UnvalidatedFX3Positioning x carry x value composite (3m)as of 2026-10-21

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-21: Direction hit rate vs USD — 51.8%, binomial p 0.28, not above the majority-direction baseline; the composite is served as context, no accuracy claim

L1 on the validation record (as of 2026-10-21) · 51.8% · no accuracy claim

Real · UnvalidatedFXPOSCFTC leveraged-fund positioning extremes -> 4-week reversal vs USDas of 2026-10-22

L1 — honest negative — no accuracy claim

When leveraged funds' net position in a currency future is beyond +/-2 standard deviations of its last 156 weekly reports, the currency moves against the crowd versus the dollar over the four weeks starting one report later.

Horizon: four weeks from the report after the extreme. Baseline: a coin (random walk) and the currency's rolling majority 4-week direction. Settled by: sign of the ECB-reference USD value of the currency, report i+5 vs report i+1.

Result as of 2026-10-22: FAIL: the currency reversed after 52.6% of 152 extremes (2009-12-29..2026-06-09), one-sided p 0.29 vs a coin; the rolling majority direction was right 52.0% (McNemar p 0.50)

L1 on the validation record (as of 2026-10-22) · 52.6% · no accuracy claim

OOS ValidatedFXVOLNext month's realised volatility per G10 currency vs USDas of 2026-10-30

L3 — passing out of sample; live grading begins once the forecast is issued and its outcome settles

Next month's realised variance of each G10 currency vs USD (ECB reference rates) is forecast by the EWMA (lambda 0.94) of daily squared returns, better than last month's or the last 12 months' realised variance (QLIKE).

Horizon: the next calendar month. Baseline: last month's realised variance and the last 12 months', per day. Settled by: ECB euro reference rates of that month: sum of squared daily log returns vs USD.

Result as of 2026-10-30: PASS: mean QLIKE -6.084 vs -6.051 (last1, Holm p < 0.0001) and -6.035 (last12, Holm p 0.0053); 9 currencies, 306 months 2001-01..2026-06 (lower is better)

L3 on the validation record (as of 2026-10-30) · -6.084 vs -6.051

Current forecast: not in this payload (as-of and forecast status with it)

OOS ValidatedFXCARRYG10 carry: long the 3 highest policy rates, short the 3 lowest, vs USD (1m)as of 2026-11-09

L3 — a measured grade at L3, not a pre-registered forward claim

Result as of 2026-11-09: Mean monthly excess return > 0 — PASSES: +0.264% a month (Sharpe 0.43 a year, max drawdown 0.35 log), one-sided Newey-West p 0.018, Holm p 0.035 over the two carry strategies; 317 months; dollar carry FAILS (p 0.147, Sharpe 0.22). Tested on history, not yet graded live

L3 on the validation record (as of 2026-11-09) · +0.264%/month · no accuracy claim

Real · UnvalidatedFXVALUEFive-year REER change sort: long the 3 cheapest, short the 3 dearest, vs USD (1m)as of 2026-11-09

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-09: Mean monthly excess return > 0 — FAILS: -0.009% a month (Sharpe -0.02), one-sided Newey-West p 0.54; 317 months; no accuracy claim

L1 on the validation record (as of 2026-11-09) · -0.009%/month · no accuracy claim

Real · UnvalidatedFXTAYLORTaylor-rule differential -> next month's exchange-rate change vs a random walkas of 2026-11-09

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-09: Clark-West vs the random walk — FAILS: one-sided p 0.25, MSPE ratio 1.040 (above 1 = worse than the random walk); 192 months 2010-03..2026-03, 6 currencies (CHF no gap history; AUD/NZD no monthly CPI); no accuracy claim

L1 on the validation record (as of 2026-11-09) · MSPE ratio 1.040 · no accuracy claim

Real · UnvalidatedFXDENSITYNext month's exchange-rate density (EWMA-vol Normal) vs a random-walk Normal (CRPS)as of 2026-11-09

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-09: CRPS vs the random walk — FAILS: 0.01631 vs 0.01637 (lower is better), one-sided Newey-West DM p 0.25; 268 months 2004-02..2026-05; no accuracy claim

L1 on the validation record (as of 2026-11-09) · 0.01631 vs 0.01637 · no accuracy claim

Real · UnvalidatedFXCRASHVIX above its 5-year 80th percentile -> a bottom-5% month for broad carry (G10 + EM)as of 2026-11-09

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-09: One-sided Fisher exact — FAILS: crash rate 8.3% after a flag vs 4.1% without (lift 2.04, p 0.13; 16 events in 317 months); the flat-when-flagged overlay cut the worst drawdown 0.23 -> 0.19 log but lowered the Sharpe 0.73 -> 0.66; no accuracy claim

L1 on the validation record (as of 2026-11-09) · lift 2.04 · no accuracy claim

Energy · 11 rows on the record · open the Energy page

ENRP · Tomorrow's net-load ramp above its 30-day median (per balancing authority)

L2 — passing out of sample; live grading begins once the forecast is issued and its outcome settles. Result as of 2026-10-22: PASS: Brier skill +0.097 vs seasonal naive (90% CI +0.089 to +0.106) and +0.042 vs persistence; AUC 0.691 vs 0.547; n 34,625 BA-days, 55 BAs, 2025-01-01..2026-10-19

Calibration on 34,625 settled outcomes: Brier 0.2240 vs benchmark 0.2337 (skill score 0.042: better than the benchmark) · no reliability curve: the walk-forward's (forecast, outcome) pairs come from the six EIA-930 six-month files named (with sha256) in the record, which are not on this host; Brier and n are the record's

as of 2026-10-22

+0.097

In-Sample

EN4 · Crack-spread reversion (1m)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-03: RMSE skill vs the random walk — does NOT beat it (DM p 0.89; direction 50.3%); observed margin, no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-11-03, never as current

as of 2026-11-03

−1.29%

Real · Unvalidated

EN1 · Gas-storage signal (1m)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-03: Direction hit rate for next month's Henry Hub — 90% CI 39%–53% includes 50%, below the majority-sign baseline (47.1%); context, no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-11-03, never as current

as of 2026-11-03

45.7%

Real · Unvalidated

ENFE · Grid operator's day-ahead demand forecast runs high tomorrow (per balancing authority)

L1 — passing out of sample; live grading begins once the forecast is issued and its outcome settles. Result as of 2026-11-03: PASS: Brier skill +0.163 vs each BA's base rate (90% CI +0.149 to +0.177) and +0.109 vs sign persistence; AUC 0.857 vs 0.773; n 34,757 BA-days, 55 BAs, 2025-01-01..2026-10-30; back-tested on latest-vintage history, not live-graded

verdict as of 2026-11-03 · not current: tested 2026-11-03; the claim settles every day, so outcomes settled since 2026-11-04 are not in this verdict - it is the verdict as of 2026-11-03, re-graded when live grading begins once the forecast is issued and its outcome settles

as of 2026-11-03

+0.163

Real · Unvalidated

ENS1 · Weekly gas-storage surprise -> next-week Henry Hub change

L1 — honest negative — no accuracy claim. Result as of 2026-11-03: FAIL: RMSE skill vs no change -0.02% (DM p 0.53); direction 48.2% (90% CI 44.7% to 51.7%) vs the usual direction 49.8%; n 560; no accuracy claim

verdict as of 2026-11-03 · not current: tested 2026-11-03; the claim settles every week, so outcomes settled since 2026-11-10 are not in this verdict - it is the verdict as of 2026-11-03, re-graded when live grading begins once the forecast is issued and its outcome settles

as of 2026-11-03

−0.02%

Real · Unvalidated

ENS3 · NY Harbor 3-2-1 crack reversion, weekly

L1 — honest negative — no accuracy claim. Result as of 2026-11-03: FAIL: RMSE skill vs no change -0.70% (DM p 0.76); direction 52.3% (90% CI 49.5% to 55.0%) vs the usual direction 52.3%; n 903; no accuracy claim

as of 2026-11-03

−0.70%

Real · Unvalidated

ENR1 · ENRP ramp claim vs a weekday-calendar + persistence table (stronger-baseline check)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-03: Brier skill of the ENRP model vs the 4-cell table — +0.033 (90% CI +0.029 to +0.038): the model adds skill beyond calendar + persistence, but the table alone recovers about a fifth of its gain over persistence; a check on ENRP, not a claim of its own

a measured grade with no declared settlement cadence: it reads as of 2026-11-03, never as current

as of 2026-11-03

+0.033

Real · Unvalidated

ENT1 · Weekly Lower-48 gas-storage change nowcast (Bcf), the day before EIA's report

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-09: RMSE 19.2 Bcf vs 45.2 for the 5-year same-week average, 52.4 for last week's change and 45.4 for a seasonal-only model (skill +0.57, 90% CI +0.53 to +0.62 vs the 5-year average; every Holm-adjusted DM p < 0.000001); back-tested on latest-vintage inputs, not live-graded

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

+0.57

Real · Unvalidated

ENT2 · ENRP v2: tomorrow's net-load ramp above its 30-day median, per grid operator, with weather

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-09: Brier skill +0.055 vs the ENRP v1 model (90% CI +0.047 to +0.063), +0.094 vs persistence, +0.086 vs the weekday + persistence table; 27 grid operators individually above v1; back-tested on latest-vintage EIA-930, not live-graded

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

+0.055

Real · Unvalidated

ENT4 · Tomorrow's 24 hourly day-ahead hub prices (LEAR) at CAISO SP15, NYISO Zone J, MISO Indiana

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-09: MAE skill vs the same hour a week ago +0.61 / +0.49 / +0.58 and vs yesterday +0.20 / +0.24 / +0.25 (all Holm-adjusted DM p <= 0.0011; 75 / 89 / 98 test days, short of the 90-day intent at two hubs); back-tested, not live-graded; hub prices served under the pre-public licence hold

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

+0.20

Real · Unvalidated

ENT5 · Tomorrow's net-load ramp range (10th / 50th / 90th percentile, MW) per grid operator

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-09: pinball-loss skill +0.054 vs the last 30 days' percentiles (90% CI +0.043 to +0.067) and +0.159 vs the same weekday over 8 weeks; the 10-90 band held 78.8% of outcomes (target 80%); back-tested on latest-vintage EIA-930, not live-graded

a measured grade with no declared settlement cadence: it reads as of 2026-11-09, never as current

as of 2026-11-09

+0.054

Real · Unvalidated

Energy: each claim's horizon, baseline and out-of-sample result

Energy · the block on its page

In-SampleENRPTomorrow's max 3-hour net-load ramp above its 30-day medianas of 2026-10-22

L2 — passing out of sample; live grading begins once the forecast is issued and its outcome settles

For a balancing authority, the probability that tomorrow's largest 3-hour rise in net load (demand minus solar and wind) is above the median of its last 30 daily ramps.

Horizon: the balancing authority's next local day. Baseline: seasonal naive (same weekday last week above the median) and persistence (today above it). Settled by: EIA-930 hourly demand, solar and wind for the next day, the same ramp arithmetic: above the frozen 30-day median = 1, else 0; an incomplete or unpublished day = void.

Result as of 2026-10-22: PASS: Brier skill +0.097 vs seasonal naive (90% CI +0.089 to +0.106) and +0.042 vs persistence; AUC 0.691 vs 0.547; n 34,625 BA-days, 55 BAs, 2025-01-01..2026-10-19

L2 on the validation record (as of 2026-10-22) · +0.097

Current forecast: served on each entity's own page (as-of and forecast status with it)

Real · UnvalidatedEN4Crack-spread reversion (1m)as of 2026-11-03

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-03: RMSE skill vs the random walk — does NOT beat it (DM p 0.89; direction 50.3%); observed margin, no accuracy claim

L1 on the validation record (as of 2026-11-03) · −1.29% · no accuracy claim

Real · UnvalidatedEN1Gas-storage signal (1m)as of 2026-11-03

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-03: Direction hit rate for next month's Henry Hub — 90% CI 39%–53% includes 50%, below the majority-sign baseline (47.1%); context, no accuracy claim

L1 on the validation record (as of 2026-11-03) · 45.7% · no accuracy claim

Real · UnvalidatedENFETomorrow's day-ahead demand forecast runs high (per balancing authority)as of 2026-11-03

L1 — passing out of sample; live grading begins once the forecast is issued and its outcome settles

For a balancing authority, the probability that its own published day-ahead demand forecast for tomorrow, summed over the day, comes in above tomorrow's actual demand, from today's forecast error and the last 7 days' mean error.

Horizon: the balancing authority's next local day. Baseline: each balancing authority's own base rate and sign persistence (today's error sign). Settled by: EIA-930 hourly demand forecast and demand for the next local day: sum(forecast - actual) over the hours with both > 0 = 1, else 0; a day with no such hour = void.

Result as of 2026-11-03: PASS: Brier skill +0.163 vs each BA's base rate (90% CI +0.149 to +0.177) and +0.109 vs sign persistence; AUC 0.857 vs 0.773; n 34,757 BA-days, 55 BAs, 2025-01-01..2026-10-30; back-tested on latest-vintage history, not live-graded

L1 on the validation record (as of 2026-11-03) · +0.163

Current forecast: served on each entity's own page (as-of and forecast status with it)

Real · UnvalidatedENS1Weekly gas-storage surprise -> next-week Henry Hub changeas of 2026-11-03

L1 — honest negative — no accuracy claim

Lower-48 working gas against the five prior years' same week (level and weekly change), known on the Thursday release, predicts the next week's log change in Henry Hub spot.

Horizon: one week (release Thursday to the next). Baseline: no change, the expanding mean and AR(1) of past weekly changes; the usual direction. Settled by: EIA RNGWHHD: last price on or before Thursday + 7 days vs on or before the release Thursday.

Result as of 2026-11-03: FAIL: RMSE skill vs no change -0.02% (DM p 0.53); direction 48.2% (90% CI 44.7% to 51.7%) vs the usual direction 49.8%; n 560; no accuracy claim

L1 on the validation record (as of 2026-11-03) · −0.02% · no accuracy claim

Real · UnvalidatedENS3NY Harbor 3-2-1 crack reversion, weeklyas of 2026-11-03

L1 — honest negative — no accuracy claim

Next week's change in the NY Harbor 3-2-1 crack moves toward its trailing 52-week mean.

Horizon: one week (Friday to Friday). Baseline: no change, AR(1) of weekly changes, the expanding mean change; the usual direction. Settled by: EIA daily WTI, NYH gasoline and NYH ULSD, last day of the next ISO week, 3-2-1 formula.

Result as of 2026-11-03: FAIL: RMSE skill vs no change -0.70% (DM p 0.76); direction 52.3% (90% CI 49.5% to 55.0%) vs the usual direction 52.3%; n 903; no accuracy claim

L1 on the validation record (as of 2026-11-03) · −0.70% · no accuracy claim

Real · UnvalidatedENR1ENRP ramp claim vs a weekday-calendar + persistence table (stronger-baseline check)as of 2026-11-03

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-03: Brier skill of the ENRP model vs the 4-cell table — +0.033 (90% CI +0.029 to +0.038): the model adds skill beyond calendar + persistence, but the table alone recovers about a fifth of its gain over persistence; a check on ENRP, not a claim of its own

L1 on the validation record (as of 2026-11-03) · +0.033 · no accuracy claim

Real · UnvalidatedENT1Weekly Lower-48 gas-storage change nowcast (Bcf), the day before EIA's reportas of 2026-11-09

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-09: RMSE 19.2 Bcf vs 45.2 for the 5-year same-week average, 52.4 for last week's change and 45.4 for a seasonal-only model (skill +0.57, 90% CI +0.53 to +0.62 vs the 5-year average; every Holm-adjusted DM p < 0.000001); back-tested on latest-vintage inputs, not live-graded

L1 on the validation record (as of 2026-11-09) · +0.57 · no accuracy claim

Real · UnvalidatedENT2ENRP v2: tomorrow's net-load ramp above its 30-day median, per grid operator, with weatheras of 2026-11-09

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-09: Brier skill +0.055 vs the ENRP v1 model (90% CI +0.047 to +0.063), +0.094 vs persistence, +0.086 vs the weekday + persistence table; 27 grid operators individually above v1; back-tested on latest-vintage EIA-930, not live-graded

L1 on the validation record (as of 2026-11-09) · +0.055 · no accuracy claim

Real · UnvalidatedENT4Tomorrow's 24 hourly day-ahead hub prices (LEAR) at CAISO SP15, NYISO Zone J, MISO Indianaas of 2026-11-09

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-09: MAE skill vs the same hour a week ago +0.61 / +0.49 / +0.58 and vs yesterday +0.20 / +0.24 / +0.25 (all Holm-adjusted DM p <= 0.0011; 75 / 89 / 98 test days, short of the 90-day intent at two hubs); back-tested, not live-graded; hub prices served under the pre-public licence hold

L1 on the validation record (as of 2026-11-09) · +0.20 · no accuracy claim

Real · UnvalidatedENT5Tomorrow's net-load ramp range (10th / 50th / 90th percentile, MW) per grid operatoras of 2026-11-09

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-09: pinball-loss skill +0.054 vs the last 30 days' percentiles (90% CI +0.043 to +0.067) and +0.159 vs the same weekday over 8 weeks; the 10-90 band held 78.8% of outcomes (target 80%); back-tested on latest-vintage EIA-930, not live-graded

L1 on the validation record (as of 2026-11-09) · +0.054 · no accuracy claim

Private Equity · 4 rows on the record · open the Private Equity page

PEIG · Industry-group mark stress -> BDC next-quarter share below 90% of cost

L1 — honest negative — no accuracy claim. Result as of 2026-10-22: FAIL: the fitted lead was zero at every origin, so the forecast equals no change (MAE ratio 1.0, n 723); no accuracy claim

verdict as of 2026-10-22 · current until 2027-01-22 (settles every quarter)

as of 2026-10-22

1.000x

Real · Unvalidated

PEMARK · BDC mark-down stress (next-quarter non-accrual)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-21: Does a BDC mark below 90% of cost flag the borrowers its filer puts on non-accrual next quarter? Graded 2026-10-21 on two quarters parsed by one parser under a rule written before the run: it FAILS the rule on sample size (20 new non-accruals, 30 required) although the ranking was strong (AUC 0.950). No accuracy claim, served at L1

a measured grade with no declared settlement cadence: it reads as of 2026-10-21, never as current

as of 2026-10-21

no accuracy claim

Real · Unvalidated

PESTRS · Two quarters of rising markdowns -> the industry group's marks fall

L1 — not enough data yet — no accuracy claim. Result as of 2026-10-30: INSUFFICIENT: 28 flagged group-quarters (30 required); marks fell after 50.0% vs a rolling base rate of 35.2% and 42.9% for persistence (Holm p 0.15 / 0.39); scored again as quarters arrive

verdict as of 2026-10-30 · current until 2027-01-30 (settles every quarter)

as of 2026-10-30

28 of 30

Real · Unvalidated

PESLO · Bond-market stress -> next SLOOS C&I tightening direction

L1 — honest negative — no accuracy claim. Result as of 2026-11-02: FAIL: right on 59.2% of 142 surveys 1990-07..2026-07 (3 ties void); persistence 59.2%, rolling majority 52.8%; Holm p 0.05 vs a coin, 0.55 vs persistence, 0.33 vs rolling majority; family Holm p 1.00 across the 7 INTEL-4 claims

verdict as of 2026-11-02 · current until 2027-02-02 (settles every quarter)

as of 2026-11-02

59.2%

Real · Unvalidated

Private Equity: each claim's horizon, baseline and out-of-sample result

Private Equity · the block on its page

Real · UnvalidatedPEIGIndustry-group mark stress -> a BDC's next-quarter share below 90% of costas of 2026-10-22

L1 — honest negative — no accuracy claim

A BDC's share of borrowers marked below 90% of cost next quarter moves with this quarter's change in the other BDCs' below-90% share in the industry groups it lends to.

Horizon: the BDC's next quarterly schedule of investments. Baseline: no change (the BDC's share now). Settled by: the BDC's share of grouped positions below 0.90 of cost at t+1 (SEC BDC Data Sets).

Result as of 2026-10-22: FAIL: the fitted lead was zero at every origin, so the forecast equals no change (MAE ratio 1.0, n 723); no accuracy claim

L1 on the validation record (as of 2026-10-22) · 1.000x · no accuracy claim

Real · UnvalidatedPEMARKBDC mark-down stress (next-quarter non-accrual)as of 2026-10-21

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-21: Does a BDC mark below 90% of cost flag the borrowers its filer puts on non-accrual next quarter? Graded 2026-10-21 on two quarters parsed by one parser under a rule written before the run: it FAILS the rule on sample size (20 new non-accruals, 30 required) although the ranking was strong (AUC 0.950). No accuracy claim, served at L1

L1 on the validation record (as of 2026-10-21) · no accuracy claim · no accuracy claim

Real · UnvalidatedPESTRSTwo quarters of rising markdowns in an industry group -> its marks fall next quarteras of 2026-10-30

L1 — not enough data yet — no accuracy claim

When an industry group's share of BDC positions below 90% of cost rose two quarters running, its cost-weighted mark falls next quarter.

Horizon: the next quarter's BDC filings. Baseline: the rolling base rate of falls, and persistence. Settled by: SEC BDC Data Sets: fair value / cost by group.

Result as of 2026-10-30: INSUFFICIENT: 28 flagged group-quarters (30 required); marks fell after 50.0% vs a rolling base rate of 35.2% and 42.9% for persistence (Holm p 0.15 / 0.39); scored again as quarters arrive

L1 on the validation record (as of 2026-10-30) · 28 of 30 · no accuracy claim

Real · UnvalidatedPESLOBond-market stress -> will banks tighten buyout-style loan standards next survey?as of 2026-11-02

L1 — honest negative — no accuracy claim

The next Senior Loan Officer Survey's net tightening of C&I standards to large and middle-market firms moves in the direction of the Federal Reserve Board excess bond premium's change over the preceding quarter.

Horizon: the next quarterly SLOOS release. Baseline: a coin, persistence of direction, and the rolling majority direction. Settled by: FRED DRTSCILM at the next survey vs the one before.

Result as of 2026-11-02: FAIL: right on 59.2% of 142 surveys 1990-07..2026-07 (3 ties void); persistence 59.2%, rolling majority 52.8%; Holm p 0.05 vs a coin, 0.55 vs persistence, 0.33 vs rolling majority; family Holm p 1.00 across the 7 INTEL-4 claims

L1 on the validation record (as of 2026-11-02) · 59.2% · no accuracy claim

Crypto · 2 rows on the record · open the Crypto page

Crypto · Next BTC difficulty (80% band)

L3 — a measured grade at L3, not a pre-registered forward claim. Crypto · the next Bitcoin difficulty adjustment, median and 80% band from mid-epoch block times, tested on 330 epochs. Re-graded 2026-10-22 against four rolling baselines with a significance test: it still beats the best of them (no change) by 53% at mid-epoch.

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

as of 2026-10-20

79.7%

OOS Validated

CRHRI · Next difficulty median vs Hashrate Index's rule

L1 — honest negative — no accuracy claim. Result as of 2026-10-22: FAIL: our median's error 2.45 points vs 2.36 for Hashrate Index's rule over 330 epochs (-3.7% skill, DM p 0.81); our 80% band still holds 79.7%

verdict as of 2026-10-22 · not current: tested 2026-10-22; the claim settles every difficulty epoch (about two weeks), so outcomes settled since 2026-11-05 are not in this verdict - it is the verdict as of 2026-10-22, re-graded when live grading begins once the forecast is issued and its outcome settles

as of 2026-10-22

2.45 vs 2.36

Real · Unvalidated

Crypto: each claim's horizon, baseline and out-of-sample result

Crypto · the block on its page

OOS ValidatedCryptoNext BTC difficulty (80% band)as of 2026-10-20

L3 — a measured grade at L3, not a pre-registered forward claim

Result as of 2026-10-20: Share of difficulty epochs inside the 80% band, forecast at mid-epoch (block 1,008 of 2,016); the median's error is 2.45 points against 5.18 for "no change" (53% lower)

L3 on the validation record (as of 2026-10-20) · 79.7% · no accuracy claim

Real · UnvalidatedCRHRINext difficulty median vs Hashrate Index's ruleas of 2026-10-22

L1 — honest negative — no accuracy claim

At block 1,008, our median of the next difficulty adjustment is closer to the realised change than Hashrate Index's rule (the uncorrected retarget estimate).

Horizon: the first block of the next difficulty epoch. Baseline: Hashrate Index's rule: the retarget-rule estimate at block 1,008 (+/- 5 points, ~90%). Settled by: that block's difficulty / this epoch's - 1, from the chain, 6 blocks deep.

Result as of 2026-10-22: FAIL: our median's error 2.45 points vs 2.36 for Hashrate Index's rule over 330 epochs (-3.7% skill, DM p 0.81); our 80% band still holds 79.7%

L1 on the validation record (as of 2026-10-22) · 2.45 vs 2.36 · no accuracy claim

Logistics · 4 rows on the record · open the Logistics page

LG · TSI freight band (1m, 80%)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-20: Share of months inside the 80% band — HONEST NEGATIVE: the right width on average, but misses cluster at turning points and the median does no better than "no change"; served with no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

as of 2026-10-20

80.7%

Real · Unvalidated

LGPORT · Port tonnage the year after a hurricane passage

L1 — honest negative — no accuracy claim. Result as of 2026-10-22: FAIL: tonnage rose in 54.8% of 31 port-years after a passage season (19 ports, 2020-2023) vs 48.4% for persistence (McNemar p 0.41): not significant

verdict as of 2026-10-22 · current until 2027-10-23 (settles every year)

as of 2026-10-22

54.8%

Real · Unvalidated

LGMOM · TSI freight 3-month momentum -> next month's direction

L1 — honest negative — no accuracy claim. Result as of 2026-10-30: FAIL: right on 51.5% of 266 months (2003-01..2026-03); persistence 44.4%, rolling majority 56.0%; Holm p 0.67 / 0.09 / 0.90

verdict as of 2026-10-30 · current until 2026-11-30 (settles every month)

as of 2026-10-30

51.5%

Real · Unvalidated

LGBRD · Truck border crossings -> next month's freight index direction

L1 — honest negative — no accuracy claim. Result as of 2026-11-02: FAIL: right on 51.5% of 297 months 2001-01..2026-04 (8 ties void); persistence 43.4%, rolling majority 54.5%; Holm p 0.64 vs a coin, 0.09 vs persistence, 0.79 vs rolling majority; family Holm p 1.00 across the 7 INTEL-4 claims

verdict as of 2026-11-02 · current until 2026-12-03 (settles every month)

as of 2026-11-02

51.5%

Real · Unvalidated

Logistics: each claim's horizon, baseline and out-of-sample result

Logistics · the block on its page

Real · UnvalidatedLGTSI freight band (1m, 80%)as of 2026-10-20

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-20: Share of months inside the 80% band — HONEST NEGATIVE: the right width on average, but misses cluster at turning points and the median does no better than "no change"; served with no accuracy claim

L1 on the validation record (as of 2026-10-20) · 80.7% · no accuracy claim

Real · UnvalidatedLGPORTPort tonnage the year after a hurricane passageas of 2026-10-22

L1 — honest negative — no accuracy claim

At a US port, total tonnage rises in the year after a season with a hurricane track within 150 km.

Horizon: the next calendar year. Baseline: persistence of the port's direction, and its rolling majority direction. Settled by: BTS Port Performance total tonnage, year t vs t-1.

Result as of 2026-10-22: FAIL: tonnage rose in 54.8% of 31 port-years after a passage season (19 ports, 2020-2023) vs 48.4% for persistence (McNemar p 0.41): not significant

L1 on the validation record (as of 2026-10-22) · 54.8% · no accuracy claim

Real · UnvalidatedLGMOMTSI freight 3-month momentum -> next month's directionas of 2026-10-30

L1 — honest negative — no accuracy claim

Next month's TSI freight change has the sign of its last 3-month change.

Horizon: the next monthly TSI print. Baseline: a coin, persistence of direction, and the rolling majority direction. Settled by: BTS TSI freight (bw6n-ddqk).

Result as of 2026-10-30: FAIL: right on 51.5% of 266 months (2003-01..2026-03); persistence 44.4%, rolling majority 56.0%; Holm p 0.67 / 0.09 / 0.90

L1 on the validation record (as of 2026-10-30) · 51.5% · no accuracy claim

Real · UnvalidatedLGBRDTruck border crossings -> next month's freight index directionas of 2026-11-02

L1 — honest negative — no accuracy claim

Next month's BTS TSI freight change has the sign of the acceleration of truck crossings at the US land borders (3-month mean YoY growth vs the 3 months before).

Horizon: the next monthly TSI print. Baseline: a coin, persistence of direction, and the rolling majority direction. Settled by: BTS TSI freight.

Result as of 2026-11-02: FAIL: right on 51.5% of 297 months 2001-01..2026-04 (8 ties void); persistence 43.4%, rolling majority 54.5%; Holm p 0.64 vs a coin, 0.09 vs persistence, 0.79 vs rolling majority; family Holm p 1.00 across the 7 INTEL-4 claims

L1 on the validation record (as of 2026-11-02) · 51.5% · no accuracy claim

Infrastructure · 5 rows on the record · open the Infrastructure page

IF2 · Highway traffic nowcast (1m)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-22: MAE skill vs the strongest rolling baseline (AR(1) on year-on-year growth) −5.1% (one-sided DM p 0.73) — it beats only the weaker seasonal-naive baseline (+31.4%); no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-22, never as current

as of 2026-10-22

−5.1%

Real · Unvalidated

IF6 · NRI physical-risk ranking

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-20: Rank AUC for FEMA major-disaster declarations after the NRI release — 90% CI 0.499–0.546 includes chance; past declarations rank better (0.606); context, no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

as of 2026-10-20

0.522

Real · Unvalidated

IFBTS · VMT nowcast vs BTS preliminary estimate (monthly)

L1 — not enough data yet — no accuracy claim. Result as of 2026-10-22: INSUFFICIENT: 1 month settled (BTS closer: 1, ours: 0); 2026-09 pending; no claim before 12 settled months

verdict as of 2026-10-22 · current until 2026-11-22 (settles every month)

as of 2026-10-22

0 of 1

Real · Unvalidated

IFBAND · VMT nowcast 90% band vs the rolling AR(1) band

L1 — honest negative — no accuracy claim. Result as of 2026-10-30: FAIL: the band covered 90.3% of 320 months (Kupiec p 0.85) but its misses cluster (Christoffersen p < 0.0001) and its interval score 48,751 loses to the rolling AR(1) band's 38,239 (DM p 0.93)

verdict as of 2026-10-30 · current until 2026-11-30 (settles every month)

as of 2026-10-30

90.3%

Real · Unvalidated

IFARX · Highway travel next month from its own trend plus factory output and jobs

L1 — honest negative — no accuracy claim. Result as of 2026-11-02: FAIL: mean squared error 15.227 vs 15.230 pp^2 for the rolling AR(1) (DM p 0.49); 317 months 2000-01..2026-07; family Holm p 1.00 across the 7 INTEL-4 claims

verdict as of 2026-11-02 · current until 2026-12-03 (settles every month)

as of 2026-11-02

15.227

Real · Unvalidated

Infrastructure: each claim's horizon, baseline and out-of-sample result

Infrastructure · the block on its page

Real · UnvalidatedIF2Highway traffic nowcast (1m)as of 2026-10-22

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-22: MAE skill vs the strongest rolling baseline (AR(1) on year-on-year growth) −5.1% (one-sided DM p 0.73) — it beats only the weaker seasonal-naive baseline (+31.4%); no accuracy claim

L1 on the validation record (as of 2026-10-22) · −5.1% · no accuracy claim

Real · UnvalidatedIF6NRI physical-risk rankingas of 2026-10-20

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-20: Rank AUC for FEMA major-disaster declarations after the NRI release — 90% CI 0.499–0.546 includes chance; past declarations rank better (0.606); context, no accuracy claim

L1 on the validation record (as of 2026-10-20) · 0.522 · no accuracy claim

Real · UnvalidatedIFBTSVMT nowcast vs BTS's preliminary estimate, month by monthas of 2026-10-22

L1 — not enough data yet — no accuracy claim

For each month both were made before FHWA printed it, the IF2 nowcast of US vehicle-miles is closer to FHWA's first print than BTS's preliminary estimate.

Horizon: the month's FHWA Traffic Volume Trends print (about seven weeks after the month). Baseline: BTS's own preliminary estimate (earliest archived vintage). Settled by: FRED TRFVOLUSM227NFWA first print; absolute % error of each.

Result as of 2026-10-22: INSUFFICIENT: 1 month settled (BTS closer: 1, ours: 0); 2026-09 pending; no claim before 12 settled months

L1 on the validation record (as of 2026-10-22) · 0 of 1 · no accuracy claim

Real · UnvalidatedIFBANDVMT nowcast 90% band: calibrated and sharper than the rolling AR(1) bandas of 2026-10-30

L1 — honest negative — no accuracy claim

The IF2 nowcast's 90% band from its own earlier errors covers FHWA's print 90% of months, independently, with a better interval score than the rolling AR(1)-on-YoY band.

Horizon: the month's FHWA Traffic Volume Trends print. Baseline: the rolling AR(1) on YoY growth with its own band. Settled by: FRED TRFVOLUSM227NFWA.

Result as of 2026-10-30: FAIL: the band covered 90.3% of 320 months (Kupiec p 0.85) but its misses cluster (Christoffersen p < 0.0001) and its interval score 48,751 loses to the rolling AR(1) band's 38,239 (DM p 0.93)

L1 on the validation record (as of 2026-10-30) · 90.3% · no accuracy claim

Real · UnvalidatedIFARXHighway travel next month from its own trend plus factory output and jobsas of 2026-11-02

L1 — honest negative — no accuracy claim

Next month's year-over-year growth of US vehicle-miles travelled is forecast by a recursive ARX with industrial production and payroll growth (point-in-time vintages) better than the rolling AR(1) on VMT growth alone (squared error).

Horizon: the month's FHWA Traffic Volume Trends print. Baseline: the rolling AR(1) on VMT year-over-year growth (the IF2 baseline). Settled by: FRED TRFVOLUSM227NFWA.

Result as of 2026-11-02: FAIL: mean squared error 15.227 vs 15.230 pp^2 for the rolling AR(1) (DM p 0.49); 317 months 2000-01..2026-07; family Holm p 1.00 across the 7 INTEL-4 claims

L1 on the validation record (as of 2026-11-02) · 15.227 · no accuracy claim

Insurance & ILS · 5 rows on the record · open the Insurance & ILS page

IL3 · State 1-in-10 storm loss (1y)

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-11-01: Exceedance rate of the served 1-in-10 level vs the stated 10% — 90% CI 6.7%–9.5% excludes 10% (conservative); historical statistic, no accuracy claim

a measured grade with no declared settlement cadence: it reads as of 2026-11-01, never as current

as of 2026-11-01

7.98%

Real · Unvalidated

IL2 · Chain-ladder reserve back-test

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-20: Median absolute reserve error vs development-year-10 incurred — chain-ladder 4.3% vs undeveloped 6.0% (beats it in 63% of triangles); method test, no company triangle served

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

as of 2026-10-20

4.3%

Real · Unvalidated

ILNRI · FEMA NRI loss ratio ranks next year's state tail years

L1 — not enough data yet — no accuracy claim. Result as of 2026-11-01: INSUFFICIENT: only 7 tail years in 104 state-years (2024, 2025); 15 are needed. Rank AUC 0.387 (90% CI 0.285-0.492), below chance on this sample

verdict as of 2026-11-01 · current until 2027-11-02 (settles every year)

as of 2026-11-01

0.387

Real · Unvalidated

ILRANK · FEMA NRI expected annual loss ranks next year's state storm damage

L1 — honest negative — no accuracy claim. Result as of 2026-11-01: FAIL: Spearman 0.460 for the NRI vs 0.527 for the last-10-year mean, 52 states x 2024, 2025; difference -0.067 (90% CI -0.214 to +0.070)

verdict as of 2026-11-01 · current until 2027-11-02 (settles every year)

as of 2026-11-01

0.460

Real · Unvalidated

ILEXC · Will a state's storm damage next year top its 5-year median?

L1 — honest negative — no accuracy claim. Result as of 2026-11-02: FAIL: Brier 0.2469 vs 0.2511 for the base rate (Holm p 0.02) and 0.2501 for persistence (Holm p 0.11); 17 years (884 state-years) 2009..2025; family Holm p 0.56 across the 7 INTEL-4 claims (lower Brier is better)

verdict as of 2026-11-02 · current until 2027-11-03 (settles every year)

as of 2026-11-02

0.2469

Real · Unvalidated

Insurance & ILS: each claim's horizon, baseline and out-of-sample result

Insurance & ILS · the block on its page

Real · UnvalidatedIL3State 1-in-10 storm loss (1y)as of 2026-11-01

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-11-01: Exceedance rate of the served 1-in-10 level vs the stated 10% — 90% CI 6.7%–9.5% excludes 10% (conservative); historical statistic, no accuracy claim

L1 on the validation record (as of 2026-11-01) · 7.98% · no accuracy claim

Real · UnvalidatedIL2Chain-ladder reserve back-testas of 2026-10-20

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-20: Median absolute reserve error vs development-year-10 incurred — chain-ladder 4.3% vs undeveloped 6.0% (beats it in 63% of triangles); method test, no company triangle served

L1 on the validation record (as of 2026-10-20) · 4.3% · no accuracy claim

Real · UnvalidatedILNRIFEMA NRI loss ratio ranks next year's tail years by stateas of 2026-11-01

L1 — not enough data yet — no accuracy claim

A state's NRI expected annual loss relative to its recent recorded storm damage ranks which states' damage next year exceeds their own 1-in-10 level.

Horizon: the next calendar year. Baseline: climatology (every state 10%, no ranking). Settled by: NOAA NCEI Storm Events damage (CPI-adjusted) above the state's 1-in-10 level from earlier years.

Result as of 2026-11-01: INSUFFICIENT: only 7 tail years in 104 state-years (2024, 2025); 15 are needed. Rank AUC 0.387 (90% CI 0.285-0.492), below chance on this sample

L1 on the validation record (as of 2026-11-01) · 0.387 · no accuracy claim

Real · UnvalidatedILRANKFEMA NRI expected annual loss ranks next year's state storm damageas of 2026-11-01

L1 — honest negative — no accuracy claim

Across states, the NRI expected annual loss ranks next year's recorded storm damage better than recent damage does.

Horizon: the next calendar year. Baseline: climatology: the state's last-10-year mean damage. Settled by: NOAA NCEI Storm Events damage (CPI-adjusted).

Result as of 2026-11-01: FAIL: Spearman 0.460 for the NRI vs 0.527 for the last-10-year mean, 52 states x 2024, 2025; difference -0.067 (90% CI -0.214 to +0.070)

L1 on the validation record (as of 2026-11-01) · 0.460 · no accuracy claim

Real · UnvalidatedILEXCWill a state's storm damage next year top its 5-year median?as of 2026-11-02

L1 — honest negative — no accuracy claim

The probability that a state's CPI-adjusted storm damage next year exceeds the median of its previous five years is forecast from its own recent damage profile better than the base rate or persistence (Brier, per year).

Horizon: the next calendar year. Baseline: the rolling pooled base rate, and persistence (given last year's exceedance). Settled by: NOAA NCEI Storm Events damage (CPI-adjusted).

Result as of 2026-11-02: FAIL: Brier 0.2469 vs 0.2511 for the base rate (Holm p 0.02) and 0.2501 for persistence (Holm p 0.11); 17 years (884 state-years) 2009..2025; family Holm p 0.56 across the 7 INTEL-4 claims (lower Brier is better)

L1 on the validation record (as of 2026-11-02) · 0.2469 · no accuracy claim

Banking · 3 rows on the record · open the Banking page

B1 · Bank stress

L2 — a measured grade at L2, not a pre-registered forward claim. Bank stress model · out-of-time AUC on one held-out window; beats the level-persistence baseline (0.5838). Research record — not offered in the product today.

a measured grade with no declared settlement cadence: it reads as of 2026-11-01, never as current

as of 2026-11-01

0.6522

In-Sample

BKFAIL · Bank failure-risk screen

L1 — a measured grade at L1, not a pre-registered forward claim. Measured as of 2026-10-20: Does the screen rank banks that fail within 12 months better than the best single ratio? NO: AUC 0.909 vs 0.942 for the nonperforming-assets ratio (FDIC NPERFV) alone (pre-registered walk-forward 2007Q4-2024Q4); calibration held (observed/expected 1.09). No accuracy claim, served at L1

a measured grade with no declared settlement cadence: it reads as of 2026-10-20, never as current

Calibration on 434,649 settled outcomes: Brier 0.0043 vs benchmark 0.0049 (skill score 0.121: better than the benchmark) · mean calibration gap 0.001

reliability: forecast vs what happened

bin	n	forecast	happened
decile 1	43,465	0.0%	0.0%
decile 2	43,465	0.0%	0.1%
decile 3	43,465	0.0%	0.1%
decile 4	43,465	0.0%	0.1%
decile 5	43,465	0.0%	0.1%
decile 6	43,465	0.0%	0.1%
decile 7	43,465	0.1%	0.1%
decile 8	43,465	0.1%	0.1%
decile 9	43,465	0.2%	0.2%
decile 10	43,464	4.0%	4.0%

as of 2026-10-20

no accuracy claim

Real · Unvalidated

BKNP · Bank nonperforming-assets rise next quarter (>= 0.25 pp)

L2 — passing out of sample; live grading begins once the forecast is issued and its outcome settles. Result as of 2026-10-22: PASS: Brier skill +0.028 vs persistence (90% CI +0.023 to +0.033), AUC 0.653 vs 0.560, observed / expected 1.03; n 733,645 bank-quarters 2000-2025

verdict as of 2026-10-22 · current until 2027-01-22 (settles every quarter)

Calibration on 733,645 settled outcomes: Brier 0.1175 vs benchmark 0.1209 (skill score 0.028: better than the benchmark) · mean calibration gap 0.013

reliability: forecast vs what happened

bin	n	forecast	happened
0.6–8.6%	73,365	6.6%	5.3%
8.6–9.9%	73,364	9.3%	8.0%
9.9–10.8%	73,365	10.4%	9.3%
10.8–11.6%	73,364	11.2%	10.6%
11.6–12.4%	73,364	12.0%	11.9%
12.4–13.2%	73,365	12.8%	13.5%
13.2–14.3%	73,364	13.7%	15.1%
14.3–15.9%	73,365	15.0%	17.5%
15.9–19.3%	73,364	17.3%	21.0%
19.3–83.9%	73,365	29.0%	29.7%

as of 2026-10-22

+0.028

In-Sample

Banking: each claim's horizon, baseline and out-of-sample result

Banking · the block on its page

In-SampleB1Bank stressas of 2026-11-01

L2 — a measured grade at L2, not a pre-registered forward claim

Result as of 2026-11-01: Held-out AUC (two quarters, pre-registered) — beats ranking by today's ratio (0.5991)

L2 on the validation record (as of 2026-11-01) · 0.6522 · no accuracy claim

Real · UnvalidatedBKFAILBank failure-risk screenas of 2026-10-20

L1 — a measured grade at L1, not a pre-registered forward claim

Result as of 2026-10-20: Does the screen rank banks that fail within 12 months better than the best single ratio? NO: AUC 0.909 vs 0.942 for the nonperforming-assets ratio (FDIC NPERFV) alone (pre-registered walk-forward 2007Q4-2024Q4); calibration held (observed/expected 1.09). No accuracy claim, served at L1

L1 on the validation record (as of 2026-10-20) · no accuracy claim · no accuracy claim

In-SampleBKNPNext-quarter rise of at least 0.25 pp in nonperforming assets / total assetsas of 2026-10-22

L2 — passing out of sample; live grading begins once the forecast is issued and its outcome settles

The probability that a bank's nonperforming assets ratio (FDIC NPERFV: noncurrent assets plus other real estate owned, % of total assets) rises by at least 0.25 percentage points from this call-report quarter to the next.

Horizon: the next call-report quarter (known about 60 days after it ends). Baseline: persistence (the event rate given last quarter's rise) and the base rate. Settled by: FDIC call report NPERFV(t+1) - NPERFV(t) >= 0.25 -> 1, else 0; no report for t+1 -> void.

Result as of 2026-10-22: PASS: Brier skill +0.028 vs persistence (90% CI +0.023 to +0.033), AUC 0.653 vs 0.560, observed / expected 1.03; n 733,645 bank-quarters 2000-2025

L2 on the validation record (as of 2026-10-22) · +0.028

Current forecast: served on each entity's own page (as-of and forecast status with it)

loading the validation record…

Which private-credit grade to trust · grade as of 2026-11-09

Out-of-filer grade: leave-one-BDC-out AUC 0.5892 on the filer's own disclosed non-accrual label and 0.6916 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 0.59 [95% CI 0.53, 0.65] on 97 disclosed non-accrual events, and 0.69 [95% CI 0.67, 0.72] on 545 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 2,735 of the 3,065 borrowers in the panel and refuses 330 — 10.8% — by name: a refused borrower is a holding a lending BDC reports in its own 1940 Act control or affiliate section (its own fund, joint venture or controlled company), or one whose 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 2026-09-30 filings (GBDC as of 2026-06-30; filed 2026-08-03 to 2026-11-06), 7,639 tranches over 3,065 borrowers from 14 BDC filers.

strict (leave-one-filer-out, disclosed non-accrual): 0.5892 on 97 events · union (leave-one-filer-out, non-accrual or mark below 0.85×): 0.6916 on 545 events · as of 2026-11-09

Read this before you read the number

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 the score itself is not a forecast. The one forward claim we register for it is narrower: that its ranking persists into each borrower's next filed quarter. That claim is graded against the filer's own non-accrual flag once the next filing is in.

Retired claims log

Four numbers we published, then withdrew.

Each with the specific failure that killed it and the number, if any, that replaced it.

RETIRED 2026-07-20

US recession probit

AUC 0.862

Replaced by

0.79–0.81as of 2026-09-01AUC across label gaps · one recession in the out-of-sample window · that model was itself retired on 2026-11-01 and no recession probability is shown today

Real · Unvalidated

What killed it

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.

RETIRED 2026-07-18

Private-credit PD

AUC 0.852

also 0.879 · 0.903 — retired

Replaced by

—no default-forecasting skill is claimed; replaced by the disclosed-impairment score

What killed it

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.

RETIRED 2026-07-19

Market-cycle national accuracy

90.7%

Replaced by

—no replacement figure; no forecasting-skill claim is made

What killed it

A concurrent-nowcast artifact. A corrected version does not beat persistence, so we make no forecasting-skill claim for market cycles at all.

RETIRED 2026-07-23

COT commodities positioning

0.661 / 10 markets

Replaced by

0.6137as of 2026-10-22WTI crude (CL); 0 of 22 markets pass the rolling-baseline re-grade (2026-10-22)

Real · Unvalidated

What killed it

Retired when the point-in-time re-run of 2026-07-23 across 22 markets passed only 5. The surviving headline is WTI crude at 0.6137 (CI90 0.5341–0.6933, n=140), as of 2026-10-22 on the validation record.

A validation record, not a track record.

Every grade on this page comes from held-out, out-of-time or out-of-sample evaluation. Served predictions are written to a provenance ledger whose retention status this page reads from the served scorecard, 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.

Predictions

Not retained

served predictions are written only to the answering instance's log, which is not retained

Outcomes

Not yet recorded

so no signal is graded on results in use

This record

2026-11-09

published validation record

Scope

What this page is not.

Not a compliance claim

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.

Not a track record

Held-out and out-of-time evaluation is not the same thing as having been right in use over time. No forecast has been graded on its outcome yet, so this page is a validation record, not a track record.

Not a forecast, in two places

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.

Then buy.

You have now seen the grades, the verification levels and the four withdrawals. The price is published on the next page.

See pricing Request access

Sealed forecast commitment · as of 2026-10-14

Q3 2026 BDC filing wave

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.

Audit us first.

bin

forecast

happened

2024-06-30

17.6%

20.0%

2024-09-30

21.6%

55.0%

2024-12-31

21.4%

19.6%

2025-03-31

26.0%

31.8%

2025-06-30

24.5%

17.4%

2025-09-30

25.2%

40.9%

bin

forecast

happened

0.0–1.6%

0.3%

0.0%

1.6–5.7%

3.8%

0.0%

5.7–11.4%

8.4%

0.0%

11.4–15.5%

13.8%

0.0%

15.5–21.4%

17.9%

0.0%

21.4–29.5%

24.5%

0.0%

29.5–46.5%

38.1%

0.0%

46.5–66.4%

58.1%

33.3%

66.4–89.3%

80.3%

37.5%

89.3–100.0%

94.0%

11.8%

bin

forecast

happened

decile 1

43,465

0.0%

decile 2

43,465

0.0%

0.1%

decile 3

43,465

0.0%

0.1%

decile 4

43,465

0.0%

0.1%

decile 5

43,465

0.0%

0.1%

decile 6

43,465

0.0%

0.1%

decile 7

43,465

0.1%

decile 8

43,465

0.1%

decile 9

43,465

0.2%

decile 10

43,464

4.0%

bin

forecast

happened

0.6–8.6%

73,365

6.6%

5.3%

8.6–9.9%

73,364

9.3%

8.0%

9.9–10.8%

73,365

10.4%

9.3%

10.8–11.6%

73,364

11.2%

10.6%

11.6–12.4%

73,364

12.0%

11.9%

12.4–13.2%

73,365

12.8%

13.5%

13.2–14.3%

73,364

13.7%

15.1%

14.3–15.9%

73,365

15.0%

17.5%

15.9–19.3%

73,364

17.3%

21.0%

19.3–83.9%

73,365

29.0%

29.7%

A validation record, not a track record.

Q3 2026 BDC filing wave

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 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.