Experimental · Two predictions awaiting a grade · Expected November 16
Credit applications: two experimental calls before the October result
October 21, 2026 · 1,969 synthetic profiles · Reference date October 20, 2026
Our frozen estimates are 12.10% for auto-loan applications and 22.86% for credit-card applications in the preceding 12 months. These forecast application incidence, not approvals, rejection rates or future borrowing. No humans were interviewed. The official October outcomes are pending; these estimates have not been validated against them.
Auto applications · primary
12.10%
First of two separately scored predictions
Card applications · secondary
22.86%
Second prediction; cannot replace auto
Official October outcomes
Pending
Release scheduled November 16, 2026
Large presentation sensitivity limits confidence
Across answer-label and row-order presentations, auto estimates ranged from 5.513029% to 23.997821% (18.484792 percentage points); card estimates ranged from 16.238928% to 35.669538% (19.430610 points). The same profiles appeared in each form. Averaging every form prevents choosing a favorable presentation, but does not establish agreement with human behavior. These ranges are sensitivity diagnostics, not confidence intervals or margins of error.
One issue, two predictions. Both count.
Auto loans remain the primary outcome. Credit cards are separately scored as secondary. Each prediction passes when its absolute error is at most 5 percentage points; otherwise it misses. Both results and both errors will enter the scorecard when comparable official outcomes arrive, including any misses. We will not select a winner, drop a losing item or replace auto with credit cards.
| Prediction | Frozen estimate | Official outcome | Grade |
|---|---|---|---|
| Auto · primary | 12.10% | Pending | Pending |
| Card · secondary | 22.86% | Pending | Pending |
For each item, absolute error = |frozen percentage − official percentage|. The pass threshold is inclusive and uses full precision, not the two-decimal display. It is 5 percentage points of application incidence, not 5% relative error.
The two-item mean absolute error (MAE) will be (auto absolute error + card absolute error) / 2, in percentage points. A low average cannot turn an individual miss into a pass. Any overall scorecard MAE must include both comparable items, one observation per prediction, and state its denominator and units. The Report's existing MAPE is a different measure: it averages absolute errors relative to their official outcomes, expressed as percentages. These two unresolved forecasts add nothing to completed accuracy totals or the existing MAPE denominator today. The front-page tally currently counts issues, so this adds one pending issue containing two pending predictions.
Scoring amendment · October 21, 2026
The owner approved the five-percentage-point rule after reviewing the frozen estimates and before any public release or official October outcome. It replaces an earlier private proposal that required beating both baselines. That earlier strict baseline rule was not publicly preregistered or owner-approved. The original private protocol is preserved; this dated amendment changes scoring and authorizes publication, without changing the estimates, primary/secondary designations or response denominator. Baseline comparisons will still be reported separately, but do not determine pass or miss.
We will use each item's first official October 2026 value from the New York Fed Credit Access workbook, retaining the first-release file, retrieval time and value. Later revisions will be preserved separately. The official calendar, checked October 21, schedules the release for November 16; the exact hour is unverified. A material instrument mismatch will remain unresolved, not be counted as success.
The comparisons we will also report
| Application | Our estimate | June 2026 persistence | October 2025 seasonal |
|---|---|---|---|
| Auto | 12.10% | 12.03% | 12.84% |
| Card | 22.86% | 29.89% | 26.64% |
For each item we will publish our absolute error and each baseline's absolute error against the same first-release outcome. A pass under the five-point rule does not establish an advantage over persistence or the seasonal baseline.
What was measured
The experiment adapted the published SCE questionnaire application activities into individual Yes/No response tasks. The stem asks about activities during the last twelve months; the selected items concern applying for an auto loan and applying for a credit card. Our fixed proxy window was after October 20, 2025, through October 20, 2026. It does not reconstruct the official survey's actual field dates.
The frame comprised 1,969 household reference persons aged 18 or older from 2024 ACS one-year PUMS records across all 50 states and Washington, DC, who owned, were buying or rented their homes. Thirty-one rent-free records were excluded before model outcomes. This household-head proxy is not all adults, not only applicants and not the New York Fed's actual repeated survey panel.
Each profile answered both items across four balanced presentation forms: 1,969 × 2 × 4 = 15,752 assigned, attempted and valid responses, with zero invalid or missing responses, zero repairs and no response retries. The frozen estimates average all four weighted replicate results. We used demographic weights; no outcome-based calibration was fitted to these forecasts. Shared model responses are not independent human interviews, and numerical precision does not imply equivalent accuracy.
Mechanical completion and separate negative evidence
The original mechanical gates passed: token/rendered-answer parity, 16/16 known-answer fixtures and 64/64 complete legal-format diagnostic responses. Candidate completeness also passed. No numerical presentation-stability cutoff was specified. The sensitivity hold was a later editorial recommendation, not a failed original candidate gate. Mechanical completion does not validate forecast accuracy.
A separate historical fitted-model study failed to beat persistence across five evaluation waves. Its auto MAE was 1.233 versus 1.195 percentage points for persistence; card MAE was 5.382 versus 1.832 points. These historical errors are not the current October outcomes and did not alter the frozen forecasts.
A separate later fixture test scored 30/32 and failed its strict correctness requirement. It was not the candidate's original gate. These results and the large presentation ranges limit readiness claims. This issue is experimental and makes no established-accuracy or endorsement claim.
Download and sources
Download the full-precision forecast and scoring record (JSON). This public aggregate record contains the frozen numbers, denominators, sensitivity, baseline values and dated amendment. It excludes proprietary prompts, implementation code and private client information.
- Official SCE Credit Access page
- Official Credit Access chart data
- Published questionnaire and license
- Official release calendar data
Some survey questions were taken or adapted from the Survey of Consumer Expectations, © 2013–2020 Federal Reserve Bank of New York (FRBNY). The SCE questions are available without charge at newyorkfed.org/microeconomics/sce and may be used subject to license terms posted there. FRBNY did not participate in or endorse the Lewsearch October 2026 synthetic credit-application study, and FRBNY disclaims any responsibility or legal liability for the administration of the survey and the analysis and interpretation of data collected.
Source: Survey of Consumer Expectations, © 2013–2026 Federal Reserve Bank of New York (FRBNY). The SCE data are available without charge at www.newyorkfed.org and may be used subject to license terms posted there. FRBNY disclaims any responsibility or legal liability for this analysis and interpretation of Survey of Consumer Expectations data.
Get the next forecast before the answer.
New Issues, Pulses and grades from The Lewsearch Report by email. Free, and you can leave any time.