A Synthetic Estimate of September Consumer Sentiment, Frozen Before Michigan's Release
Protocol dated September 9, 2026 · n=1,000 · 48 states + D.C. · lewis-v17 only
A public protocol commitment is timestamped on GitHub in swarmgrampublic, commit 1c0b054, dated September 9. That commit froze the method via cryptographic commitment before inference. It did not publish the forecast of 55.4. The number is published here, before Friday's Michigan release.
ICS (primary)
55.4
Frozen September forecast
ICC
61.8
Secondary
ICE
51.4
Secondary
What this run is, and is not
This run used lewis-v17 only. The synthetic respondents had no live web access, news summaries, Brave results, Google Knowledge Graph results, or economist consensus in their prompts. Under Method B, each respondent received only a sentence stating the date.
In the August reconstruction, Lewis estimated Michigan's published ICS at 55.0 versus an actual 51.7, an error of 3.3 points. That is within the registered five-point band. ICC was much farther off at 61.0 versus 51.9, an error of plus 9.1. We are keeping that miss in full view.
The frozen September estimate is 55.4, which is 3.7 points above Michigan's published August result of 51.7. But most of that difference already existed in the August reconstruction. Lewis itself moved only about 0.4 points from August to September.
The useful finding is narrower: a census-grounded synthetic panel, run without external research tools, came reasonably close to the headline index in a retrospective month. That does not prove forecasting skill, and it does not show that Lewis beat economists. The 3.7-point gap should not be read as evidence that Lewis detected a September swing.
August diagnostic
We used the same roster with an August 24, 2026 dateline and compared the result with Michigan's published August final. The registered qualification test applied only to ICS. An absolute error greater than 15 would have changed the title to a failed-qualification prospective test. Lewis did not hit that threshold.
Passing the qualification test does not prove the model had not previously encountered Michigan's August result.
| Index | Lewis Aug 24 | Michigan Aug final | Signed error |
|---|---|---|---|
| ICS (primary) | 55.0 | 51.7 | +3.3 |
| ICC | 61.0 | 51.9 | +9.1 |
| ICE | 51.1 | 51.5 | -0.4 |
ICC is the largest miss, and we are leaving it visible. The registered primary measure is ICS.
September freeze
The September run used a September 7, 2026 dateline. The frozen primary estimate is ICS 55.4. The component relatives are PAGO 78, PEXP 82, BUS12 63, BUS5 58, and DUR 80. Full-precision ICS is 55.4356. The no-change comparator is Michigan's published August final of 51.7.
The September estimate of 55.4 is 3.7 points above the human August final of 51.7. Lewis itself moved only about 0.4 points from its own August estimate of 55.0. That distinction matters. We are not claiming Lewis detected a 3.7-point September improvement.
How Friday is scored
- Under the registered five-point rule, an absolute ICS error of 5 or less is a win. An error greater than 5 is a loss.
- The win interval around 55.4 is 50.4 to 60.4.
- If Friday prints 51.7, Lewis is still inside the win band, but the no-change comparator is closer.
- If Friday prints 60.0, Lewis is also inside the win band.
- Lewis is closer than no-change if Friday prints above 53.55.
- For larger errors, 5 to 10 is a miss, 10 to 20 is bad, and greater than 20 is embarrassing.
- The Street comparison uses the Reuters median, or Bloomberg if Reuters is unavailable, only if it is posted by Thursday 23:59 ET. We will not claim that Lewis beat the Street.
Friday official print (empty until 10:00 a.m. ET, September 11, 2026)
| Index | Official | Lewis | Signed error |
|---|---|---|---|
| ICS (primary) | [FRIDAY ICS] | 55.4 | [FRIDAY ICS ERROR] |
| ICC | [FRIDAY ICC] | 61.8 | [FRIDAY ICC ERROR] |
| ICE | [FRIDAY ICE] | 51.4 | [FRIDAY ICE ERROR] |
No-change comparator: 51.7. No-change ICS error: [FRIDAY NO-CHANGE ICS ERROR]. Recorder: [FRIDAY RECORDER].
Companion, not the forecast
After the August and September synthetic outputs were complete, but before Friday's human result, we calculated an August-anchored nowcast of 52.1. Using full-precision ICS, the arithmetic is 55.4356 plus (51.7 minus 54.9915).
The original registration did not allow this adjusted number to become the lead forecast. It remains a secondary comparison. The primary forecast is still 55.4, and we will not retroactively call whichever number is closer on Friday our prediction.
Five items · relatives and option shares
For each item, A is favorable and C is unfavorable. B and D are neutral in the index and are shown separately.
The panel is pessimistic about business conditions and only mildly pessimistic about personal finances. A large share of the probability sits in neutral or mixed responses. That helps explain why ICS lands near 55 rather than 51.7 or 76.
PAGO · x1
Would you say that you (and your family living there) are better off or worse off financially than you were a year ago?
August x=77 · September x=78
August 24
- A Better off17.3%
- B About the same39.8%
- C Worse off40.0%
- D Uncertain / don't know2.9%
September 7
- A Better off17.4%
- B About the same40.2%
- C Worse off39.4%
- D Uncertain / don't know2.9%
PEXP · x2
A year from now, will you be better off financially, worse off, or just about the same as now?
August x=81 · September x=82
August 24
- A Better off11.6%
- B Just about the same49.6%
- C Worse off30.4%
- D Uncertain / don't know8.4%
September 7
- A Better off12.1%
- B Just about the same49.7%
- C Worse off30.2%
- D Uncertain / don't know8.0%
BUS12 · x3
During the next twelve months, will we have good times financially, or bad times, or what?
August x=62 · September x=63
August 24
- A Good times8.0%
- B Uncertain / mixed43.3%
- C Bad times45.7%
- D Don't know3.0%
September 7
- A Good times8.3%
- B Uncertain / mixed43.7%
- C Bad times45.1%
- D Don't know3.0%
BUS5 · x4
Looking ahead five years, continuous good times, or periods of widespread unemployment or depression, or what?
August x=59 · September x=58
August 24
- A Continuous good times3.5%
- B Uncertain / mixed48.0%
- C Unemployment or depression44.6%
- D Don't know3.9%
September 7
- A Continuous good times3.3%
- B Uncertain / mixed48.3%
- C Unemployment or depression44.8%
- D Don't know3.6%
DUR · x5
Is now a good or bad time for people to buy major household items?
August x=79 · September x=80
August 24
- A Good time to buy24.3%
- B Uncertain / depends27.0%
- C Bad time to buy45.8%
- D Don't know2.9%
September 7
- A Good time to buy24.8%
- B Uncertain / depends27.5%
- C Bad time to buy44.7%
- D Don't know2.9%
Methods
This experiment uses Michigan's published index questions and scoring formula in a synthetic adaptation. It is not a reproduction of the Michigan survey. Michigan has used web interviewing since July 2024. A preliminary release is based on about 420 interviews, and a final is based on about 1,000. Michigan's human sample covers the 48 coterminous states and D.C.
Our panel contains 1,000 census-grounded adult synthetic respondents across the 48 coterminous states and D.C., excluding Alaska and Hawaii. Each respondent carries equal weight. The same roster answered all five items on both dates.
Outputs were processed using Lewsearch's frozen production calibration pipeline. The pipeline and configuration were cryptographically committed before inference and were not modified after observing results. The primary forecast was not adjusted toward the known August value of 51.7.
Method B is a dateline conditioning, not a true knowledge cutoff. Each wave used a stated interview date: August 24, 2026 for the diagnostic and September 7, 2026 for the forecast. The underlying model may still contain knowledge from later in 2026.
The relative score is x = floor(100 + F - U + 0.5). ICS = (sum x) / 6.7558 + 2, published to one decimal using half-up rounding. Each wave contains 5,000 cells, with zero missing in Issue 05.
We also ran an independent second implementation on CPU. It matched the valid-cell counts, roster membership, index arithmetic, five integer relatives, and one-decimal indexes. The primary forecast remains 55.4.
All 5,000 of 5,000 cells passed validation. August required a retry on 70 cells. September required a retry on 58.
Hashes on this page
- Registration: 1c0b054cbfa6506aea6eee9c96a72ea41df87b93
- Roster: e16e31518c1546d4b4b4ca8618aed217aa0232b0f14deeb071c3dd00f793b524
- August cells: b500dbd826718d1c10ece28091c337286840120318ee1184a5f946ba5da7779b
- September cells: 523c04b069e81f66b0d97e9fc3cc0e54dbca60c1ff2273a0b5c05c00ccb7b8e5
- Runner: f06ca50ec6866d98d5e77baa1cb8e97e7327b687a733f6b157aee2dad3a5ceac
Read before citing
This is a synthetic panel, not a probability sample. No humans were interviewed. This is a custom Issue for The Lewsearch Report, not a result from the standard production pipeline. The protocol was frozen publicly before the result, but the numerical forecast itself was not published in that protocol freeze.