Evidence
The 500-question benchmark showcase records mean error of 3.72 points per answer in the September 1, 2026 historical evaluation.
Memory
Setup. 300 simulated respondents over eight simulated weeks. Topic-matched memory compared with the same questions before that memory accumulated.
Result. High-signal questions showed up to 73% lower error than the same questions before that memory. Mid-signal questions fell about 46%. Stable questions with no relevant signal did not move. That figure is for this experiment only. The product score is the 50-state book above.
Matching
Setup. A Vermont question on whether people are likely to move. One run used the census panel draw. The other rebuilt the sample to that study's age, sex, party, and education.
Result. 20.4 points on the census draw. 9.5 points when the sample was rebuilt to that study's age, sex, party, and education. One question, not the 50-state book.
The rows below summarize previously published results. Types are not pooled into one accuracy score. Research files and full instruments are available by request.
Request research materials · Methodology
| Study | Type and freeze | Metric and result | Materials |
|---|---|---|---|
| 500-question benchmark showcase lewsearch-as-shipped-50state | Retrospective benchmark Freeze unknown | MAE · percentage points The 500-question benchmark showcase records mean error of 3.72 points per answer in the September 1, 2026 historical evaluation. | Request materials |
| April 2026 five-place calibrated benchmark lewsearch-april-calibrated | Retrospective benchmark Freeze unknown | MAE · percentage points 7.47 pp on 404 non-electoral questions of a 460-question bank. | Request materials |
| Held-out set sourced 2026-04-18 lewsearch-heldout-2026-04-18 | Held-out test Freeze 2026-04-18 | MAE · percentage points 9.97 pp non-electoral; 10.68 pp overall. | Request materials |
| Public per-question CSV (April bank subset) lewsearch-public-csv-443 | Retrospective benchmark Freeze unknown | MAE · percentage points About 7.44 pp on 389 non-electoral rows. Do not replace with the 50-state book. | Request materials |
| Issue 01: NRF back-to-school lewsearch-report-issue1-nrf | Prospective forecast Freeze 2026-07-07 | mean K-12 spend miss · percent of the NRF figure, as published on the issue Mean K-12 spend within 1%. | Request materials |
| Issue 04: 225 pre-registered predictions lewsearch-report-issue4-lock | Prospective forecast Freeze unknown | mixed · see issue See the issue. This record does not invent a pooled accuracy score. | Request materials |
| Issue 07: Deloitte 2026 holiday budget, Partial (original MISS retained) lewsearch-report-issue7-deloitte-holiday | Baseline-anchored forecast Freeze 2026-10-01 | Deloitte headline average expected holiday spend · U.S. dollars PARTIAL, category added October 18 after the original October 17 MISS. Forecast $1,589; official $1,654. Absolute error $65 (3.9% of the official amount), inside the $80 tolerance; the $1,595 baseline's $59 error was $6 closer. The same recipe run one year back gives $1,772 for 2025, against Deloitte's $1,595. | Request materials |
| Issue 06: August household spending, Partial (original MISS retained) lewsearch-report-issue6-sce-f1 | Prospective forecast Freeze 2026-09-15 | F1 interpolated median, reported year-over-year spending change · percentage points PARTIAL added October 18 after grading; original September 23 MISS retained. Official 4.7 (raw 4.7047). Forecast absolute error 0.1953 (inside 0.4 band). April 4.8 absolute error 0.0953 (closer). Failed closer-than-last-print; frozen rule unchanged. | Request materials |
| Issue 05: September Consumer Sentiment, primary MISS lewsearch-report-issue5-ics | Prospective forecast Freeze 2026-09-09 | ICS miss · index points Primary MISS: 55.4 vs 47.8, error 7.6 index points. No-change 51.7 error 3.9. Companion 52.1 error 4.3; not the primary grade. | Request materials |
| Public Pew forecasts lewsearch-predictions-pew | Prospective forecast Freeze unknown | varies by item · see page See /predictions. Do not fold these into the 10 pp MAE. | Request materials |
| Polymarket scoreboard lewsearch-proof-polymarket | Ungraded editorial pulse Freeze unknown | market comparison · see page Credibility check only. Historical disclosure, not a product claim. | Request materials |