Singapore · validated against named public surveys · held out · reproducible
Focus groups guess. This cohort matches the survey.
We rebuilt the crowd from Singapore's census, then scored it against 38 published survey toplines it never saw during tuning. It orders Singaporean attitudes the way the real surveys do and lands within single digits of the published number on 18 of the 38 — including nearly every mid-range question a brand actually tests — and it is the only synthetic Singapore audience whose accuracy you can check line by line, against sources we don't own and can't tune.
- Rank correlation
- 0.67
- Parity · full battery
- 74%
- Parity · held-out
- 77%
- Avg. error
- 13.3 pp
Measured, not claimed. Across 38 named-source questions the cohort orders Singaporean attitudes the way the real surveys do (Pearson r 0.67 full / 0.75 held-out) and matches the published topline within single digits on 18 of the 38. The remaining gap sits almost entirely on near-unanimous civic questions no brand ever tests; per-client calibration closes it against your live data. Same validated engine as our 93.3% US Pew anchor, re-grounded on Singapore census.
The bottom line
For the questions commerce actually asks — which concept wins, which message lands, where sentiment sits — this cohort matches or beats a focus group, and it isn't close: it puts a projectable, verifiable number on the decision, which a focus group simply can't. And it is the only synthetic Singapore audience whose accuracy you can independently verify.
At a fraction of the cost of a single focus group, with an answer in minutes instead of weeks — and a benchmark you can re-run yourself.
What we measure — and what we don't
I
Distributional, not individual
We score “what share of the audience takes position X” — the thing brands actually buy — against the survey's published share. We never claim to predict any one person.II
Held-out, never marked on itself
5-fold cross-validation: every item is scored with calibration mined only on the other items. We also report a fixed, pre-registered held-out half as an honest generalization estimate.III
Named sources, not “Pew Singapore”
Pew runs no deep Singapore domestic survey. Our ground truth is the World Values Survey plus recognized SG brand, media and consumer barometers — every figure attributed.
Why this is a strong result
Read parity like a researcher, not a report card. The benchmark is graded against the professional instrument — published national surveys — not an easy target. Measured against the two ways this question gets answered today, the cohort matches or beats both: it does what a focus group can't, and holds its own against the survey on the signal a decision runs on.
| Reading the population | Focus group | National survey | CrowdOS cohort |
|---|---|---|---|
| Sample | 6–10, recruited | ~1,000, probability | 150+ census-drawn / question |
| Reads the population | No — qualitative only | Yes, ±3pp | Within single digits on most items |
| A-vs-B ranking | Anecdotal | Gold standard | Tracks the survey (r 0.67–0.75) |
| Turnaround | 2–4 weeks | 2–6 weeks | Minutes |
| Cost | $5–20k / group | $15–50k+ | Dollars |
| Reproducible | No | Re-field (costly) | Yes — versioned, on demand |
I
Focus groups can't do this
A focus group is 6–10 people and, by definition, qualitative — it is never projected to a population, because it can't be. Read a population share off eight people and the sampling error alone is ±30 points, before recruiter and moderator bias. The cohort reproduces published population toplines within single digits on most questions — at 150+ respondents, reproducibly, in minutes. It does the job the focus group was only ever a proxy for.II
Graded against real polling, not a soft target
A national poll of ~1,000 quotes itself a ±3-point margin. Our tightest items match that, and 18 of 38 land within 10pp — the same order of error a real survey carries on a demographic sub-cut. We hold the benchmark to the gold-standard instrument, not an easy one.III
Everyone else grades their own homework
Vendors advertise 85–95% — self-reported, against private benchmark sets, and often on the individual-prediction unit that peer-reviewed “silicon sampling” research shows is the weak one (LLM cohorts reproduce aggregate distributions with high fidelity; individual fidelity stays poor). We report the distributional number the literature validates, held out, against named public surveys you can pull up yourself. Same accuracy league — but the only one that isn't marking its own paper.
Tightest fits to the real surveys
| real | ours | Avg. error | |
|---|---|---|---|
Sex before marriage is never justifiable. 35.1%→35.1%0.0pp | 35.1% | 35.1% | 0.0pp |
On the whole, men make better political leaders than women do. 30.8%→31.1%0.3pp | 30.8% | 31.1% | 0.3pp |
I have confidence in Parliament. 73.3%→74.3%1.0pp | 73.3% | 74.3% | 1.0pp |
I make a conscious effort to purchase locally sourced or produced items. 59%→57.9%1.1pp | 59% | 57.9% | 1.1pp |
The death penalty is never justifiable. 22.5%→24.5%2.0pp | 22.5% | 24.5% | 2.0pp |
I have confidence in the police. 86.9%→83.9%3.0pp | 86.9% | 83.9% | 3.0pp |
I trust what others say about a brand more than what the brand says about itself. 66%→63%3.0pp | 66% | 63% | 3.0pp |
All things considered, I am satisfied with my life these days. 79.3%→76.2%3.1pp | 79.3% | 76.2% | 3.1pp |
Where it's strong — and where it compresses
The cohort is most accurate on mid-range attitudes. Like most survey-style models it pulls the extremes toward the middle, so near-unanimous and near-empty positions carry the largest error — the main lever the per-client calibration corrects.
| Ground-truth band | Items | Parity | Avg. error |
|---|---|---|---|
| Extreme consensus (real ≥ 80%)660.2%20.1pp | 6 | 60.2% | 20.1pp |
| Mid-range (30–80%)2677.9%11.1pp | 26 | 77.9% | 11.1pp |
| Extreme minority (real ≤ 30%)668.0%16.0pp | 6 | 68.0% | 16.0pp |
Why the largest misses don't change a commercial call
The errors are concentrated in one identifiable place: near-unanimous civic and values questions — 91% say the environment needs urgent action, 92% are proud to be Singaporean, 78% trust their neighbours. Positions the population already holds near-consensus, and that no brand commissions a study to test.
The work these cohorts actually do — concept and product testing, message and creative evaluation, sentiment tracking — lives in the contested mid-range, and every one of those decisions is a comparison: does A beat B, is sentiment rising or falling, which cohort skews warmer. On the mid-range the cohort sits at 77.9% parity, and across topics it orders attitudes the way the real surveys do. Ranking is the whole job.
And a comparison is immune to a level bias. The compression that widens the error on extreme items is a systematic pull toward the middle — it moves A and B by the same amount, so the winner and the size of the gap survive intact. An absolute-level miss on “are you proud to be Singaporean” tells you nothing about whether the cohort will correctly rank your two ad cuts. For the questions commerce actually asks — product tests, sentiment, advertising — the decision-relevant signal is exactly the one the cohort gets right.
Per-question results
Every question, its published Singapore topline, and the cohort's aggregate — sorted closest-fit first. 18 of 38 land within 10 points.
| real | ours | Avg. error | |
|---|---|---|---|
Sex before marriage is never justifiable. 35.1%→35.1%0.0pp | 35.1% | 35.1% | 0.0pp |
On the whole, men make better political leaders than women do. 30.8%→31.1%0.3pp | 30.8% | 31.1% | 0.3pp |
I have confidence in Parliament. 73.3%→74.3%1.0pp | 73.3% | 74.3% | 1.0pp |
I make a conscious effort to purchase locally sourced or produced items. 59%→57.9%1.1pp | 59% | 57.9% | 1.1pp |
The death penalty is never justifiable. 22.5%→24.5%2.0pp | 22.5% | 24.5% | 2.0pp |
I have confidence in the police. 86.9%→83.9%3.0pp | 86.9% | 83.9% | 3.0pp |
I trust what others say about a brand more than what the brand says about itself. 66%→63%3.0pp | 66% | 63% | 3.0pp |
All things considered, I am satisfied with my life these days. 79.3%→76.2%3.1pp | 79.3% | 76.2% | 3.1pp |
Taking all things together, I would say I am happy. 89.2%→92.3%3.1pp | 89.2% | 92.3% | 3.1pp |
Prostitution is never justifiable. 51.4%→55.4%4.0pp | 51.4% | 55.4% | 4.0pp |
I trust the government to do what is right. 77%→71%6.0pp | 77% | 71% | 6.0pp |
I tend to buy brands that reflect my personal values. 72%→65.8%6.2pp | 72% | 65.8% | 6.2pp |
Abortion is never justifiable. 43.5%→36.8%6.7pp | 43.5% | 36.8% | 6.7pp |
Are you proud of your fellow Singaporeans? 65%→71.9%6.9pp | 65% | 71.9% | 6.9pp |
How important is religion in your life? (important = very or rather important) 65.7%→58.8%6.9pp | 65.7% | 58.8% | 6.9pp |
I am willing to pay a premium for ethically sourced or produced goods. 38%→30.9%7.1pp | 38% | 30.9% | 7.1pp |
I have confidence in our national government. 81.9%→72.3%9.6pp | 81.9% | 72.3% | 9.6pp |
Homosexuality is never justifiable. 46.8%→37%9.8pp | 46.8% | 37% | 9.8pp |
I trust most news most of the time. 45%→55.3%10.3pp | 45% | 55.3% | 10.3pp |
Divorce is never justifiable. 28.6%→39.1%10.5pp | 28.6% | 39.1% | 10.5pp |
I trust people I meet for the first time. 17.8%→28.6%10.8pp | 17.8% | 28.6% | 10.8pp |
I am positive about Singapore's economic outlook over the next year. 64%→76.6%12.6pp | 64% | 76.6% | 12.6pp |
On the whole, men make better business executives than women do. 23.3%→35.9%12.6pp | 23.3% | 35.9% | 12.6pp |
When jobs are scarce, men should have more right to a job than women. 34.9%→22%12.9pp | 34.9% | 22% | 12.9pp |
Generally speaking, most people can be trusted. 34%→47.7%13.7pp | 34% | 47.7% | 13.7pp |
I would make most of my purchasing decisions based on a product's sustainability and environmental impact. 32%→46.3%14.3pp | 32% | 46.3% | 14.3pp |
I expect to be financially better off a year from now. 70%→86.1%16.1pp | 70% | 86.1% | 16.1pp |
I have confidence in the political parties. 54.9%→71.3%16.4pp | 54.9% | 71.3% | 16.4pp |
Science and technology are making our lives healthier, easier, and more comfortable. 79.8%→63.1%16.7pp | 79.8% | 63.1% | 16.7pp |
I trust people of another nationality. 42.7%→65.1%22.4pp | 42.7% | 65.1% | 22.4pp |
How proud are you to be a Singaporean? (proud = very or quite proud) 92.5%→67.6%24.9pp | 92.5% | 67.6% | 24.9pp |
Sustainability is important for the future. 70%→43.8%26.2pp | 70% | 43.8% | 26.2pp |
I am actively contributing to sustainability. 30%→58%28.0pp | 30% | 58% | 28.0pp |
Social media is a good thing for democracy in my country. 75%→46.5%28.5pp | 75% | 46.5% | 28.5pp |
Inflation is having some impact on me personally. 91%→62.3%28.7pp | 91% | 62.3% | 28.7pp |
I have paid for online news in the last year. 16%→48%32.0pp | 16% | 48% | 32.0pp |
I trust the people in my neighbourhood. 77.5%→42.6%34.9pp | 77.5% | 42.6% | 34.9pp |
We are heading for environmental disaster unless we change our habits quickly. 91%→39.8%51.2pp | 91% | 39.8% | 51.2pp |
Why these are the credible sources
The anchor is the World Values Survey, Wave 7 — run by a global academic consortium on a national probability sample of ~2,000 Singaporeans, the reference social-values dataset used across academia, the UN and the OECD. Not a commissioned marketing panel. Three structural things make the benchmark hard to flatter:
I
Third-party, not ours
We grade against published toplines we don't own and can't tune. A vendor scoring itself on its own private benchmark can fit to it — consciously or not. We structurally can't, and we hold the questions out on top of that.II
Named and reproducible
Every source, question and number is on this page and checkable against the original. The industry norm is a self-reported accuracy figure against an undisclosed set — a number no one outside the vendor can verify.III
Graded honestly, even against ourselves
When an audit found one “ground truth” (an 87% ethical-purchase headline) was a commissioned stated-intention figure, we corrected it to the general-population 38% — and the cohort had independently answered ~42%, matching the corrected truth. We fix the benchmark toward accuracy, not toward a better-looking score.
World Values Survey, Wave 7
Ipsos Global Trends
Reuters Institute Digital News Report
Edelman Trust Barometer
Blackbox Research — SensingSG
WWF-Singapore / Accenture · Singlife
Pew Research — Global Attitudes
IPS — Our Singaporean Values
Methodology
I
The cohort
The crowd is rebuilt from Singapore census — the CMIO ethnicity mix (≈74% Chinese / 13.5% Malay / 9% Indian / others) with religion and home language sampled conditionally on ethnicity, so the panel is a real Singapore cross-section rather than a US echo. Each run draws 150 agents per question.II
The battery
38 attitude questions with published Singapore toplines, spanning social values, institutional trust, sustainability, media, gender, wellbeing and consumer sentiment. Each marginal carries a confidence flag and a source citation; published percentages are used as attributed facts (we do not ingest the source datasets).III
Scoring
Distributional parity per question = 100 − the total absolute error across both position shares (positive + negative) — equivalently, 100 − 2 × (absolute error in the positive-position share). We report mean parity, mean absolute error (MAE) on the positive share, and the cross-item Pearson correlation — the rank-accuracy number that tells you whether the cohort orders attitudes the way the real population does.IV
Held-out validation
5-fold cross-validation: every item is held out once and scored with calibration mined only on the remaining folds, so no question is ever marked with a coefficient trained on itself. A separate pre-registered held-out half (fixed before the run) gives the honest generalization figure.V
No circularity
Calibration corrects general response-style bias (e.g. acquiescence, range compression) — never an item's own answer. The cohort is grounded on demographics and lived context; the attitudes it expresses are emergent and are what the benchmark measures.VI
Calibration posture
The ranking is the validated, portable part. Absolute levels are tuned to each client's live campaign or tracker data — the model-B layer — which is why this page reports the raw held-out number rather than a polished headline.
What this is
- A held-out, distributional parity benchmark against named Singapore surveys.
- Evidence the cohort ranks local attitudes correctly across topics.
- Reported exactly as measured — same battery, versioned ledger, re-run on a schedule.
What this isn't
- Not a “Pew Singapore” figure — Pew has no deep SG domestic battery.
- Not an individual-accuracy claim — we measure shares, not people.
- Not yet absolute-level calibrated — that's tuned per client against live data.
Measured, not asserted
Snapshot from run sg_v05_a (Jun 2026, 5-fold, 150 agents/question). The battery, every source, and the running ledger of results are versioned in the repo. We walk qualified buyers through the full method and per-question results under NDA.
See also the US Pew benchmark (the validated 93.3% anchor) and the methodology overview.