For hedge funds & private-markets investors

Synthetic expert cohorts for sharper investment theses.

Built for funds. Thesis Lab pressure-tests a thesis against synthetic expert cohorts before you commit time and budget to expert calls and diligence.

POWERED BY FISHDOG

Vintage radar antenna dish
4 hrs data refresh from a live feed of 359 sources, so an earnings-cycle thesis is tested against today's world, not a stale model.
18 axes of real-money prediction markets, scored every day against events that resolve to fact.
Zero MNPI by construction. Inspectable cohorts, sourced packets, nothing to log or monitor.

The inputs you rely on are degrading

01 / WHY NOW

Most synthetic-research pitches assume human research is ground truth. On the buy side, that baseline is already breaking in specific, familiar ways.

Expert networks Converging and gamed

The same experts rotate across funds, so the edge converges. And a paid call can now be a non-expert with a chatbot running.

Survey panels Self-selected and coached

Panels skew toward people who value their time cheaply, and some agencies coach participants to pass the screener.

Both Slow and expensive

Just as fielding costs rise and the half-life of a tradable signal keeps shrinking. The "real human" baseline is itself noisy.

Recruit, don't create

02 / HOW IT WORKS

The usual way to build synthetic respondents is to write a few personas and prompt a model to act them out. That imports the author's assumptions and the model's tidy average. Thesis Lab inverts the order.

Create a persona
  • Start with an archetype someone imagined
  • The prompt writer decides who exists
  • The model role-plays a tidy average
  • Hard to audit: invented for the prompt
Recruit from a population
  • Start with a population calibrated to Census and ACS
  • You set the question and the eligibility
  • Real occupational context, reasoning you can probe
  • Traceable to a frame, cohort, and source bundle
On-demand cohorts

Recruit a cohort of CIOs at 250-to-500-person firms in a named vertical, without scheduling them one by one.

Interrogate, don't capture

Ask any expert follow-up and "why" questions at volume. A fixed survey can't, without re-fielding.

No fatigue, no attrition

Re-run waves without burning the panel. Consistency is controlled in software, not re-recruited each time.

What it looks like

03 / PRODUCT
Find Experts modal: describing the cohort to recruit in plain English
Describe the cohort in plain English. "Managers and logistics coordinators at restaurant-supply companies that specialise in poultry."
Research group view: cohort participants with occupational backgrounds
Grounded in real data. Drawn from real census and occupational distributions. You see who you're talking to, their background, and the recruitment trail.
Expert Calls dashboard: synthetic cohort interviews running in parallel
Recorded and repeatable. Re-run the same question as new data, earnings and news land. Every interview recorded and cohort-traceable.
Cohorts built for markets like
Ports & logistics Fast-casual dining SaaS Practice-level healthcare Industrials Financial services Foundations

Where it fits the analyst's day

04 / WORKFLOW
Before an expert call

Shape the hypothesis, find the questions worth a paid human hour, and walk in better briefed.

Between prints

Get a fast read on a sector or cohort when fielding a human survey is too slow or too costly to justify.

On any cohort, on demand

Segment into a micro-population no human panel was ever recruited for, and probe it.

After a market-moving event

Re-run the same panel the same day to read the change, rather than waiting a quarter for the next wave.

See it run against two of your live questions.

Book a Demo

How we know it holds

05 / VALIDATION

We don't ask you to trust one number. Thesis Lab is validated in layers, each against an external yardstick.

Construction floorUS Census + ACS / PUMS microdata

~1% average marginal error across selected demographic and labour-force marginals. The population is audited against the real one before a single opinion is elicited.

Prediction markets18 event axes, scored daily

Graded against real-money events that resolve to fact. Strongest on technical and expertise-driven questions, which is exactly where the analyst value is highest.

Gallup / PewSeveral hundred public-survey questions

~9% mean absolute error versus published human results, compared by demographic group where cross-tabs exist.

Michigan sentimentSurveys of Consumers, running since 1946

A standing benchmark against a published, federally-tracked series, continuously monitored for directional accuracy and drift. See the model check →

Read the reliability before the event resolves. Because every benchmark question is scored on the same 18 axes, a new question is matched to prior ones of similar shape, giving an expected reliability range while there is still time to act, not after the event settles. Full methodology available for diligence review.

Design Partner Programme

Bring two live investment questions

Thesis Lab will run synthetic expert cohorts and channel checks, then compare the outputs against your existing expert-network workflow.

The questions funds ask us

06 / FAQ

Then you've hit the data-staleness wall - frontier models train to a fixed cutoff, so an earnings-cycle thesis gets tested against last year's world. Thesis Lab's cohorts are grounded in real census, occupational, and market data, refreshed every four hours. And because answers come from a calibrated population, you get the full distribution - outliers and disagreement included - not the consensus average a chatbot collapses to. Cohort design is inspectable and outputs are linked to sourced research packets - traceability an off-the-shelf chatbot can't give your diligence or compliance teams.

No - it makes them more productive. Expert networks give you depth from real people; Thesis Lab gives you breadth, structure, and repeatability first. You walk into the expert call with a sharper agenda, a better expert match, and the questions already framed - so the call you were going to book anyway returns far more signal.

We don't rest it on one number. Thesis Lab is validated in layers, each against an external yardstick: the population calibrates to Census and ACS microdata at roughly 1% marginal error, cohorts are scored daily against 18 real-money prediction-market axes, and outputs are benchmarked against Gallup, Pew, and the University of Michigan sentiment series. The full methodology is available for diligence review - we would rather you stress-test it than take any figure on faith.

Yes - that's the point. Cohorts are built from real census and occupational data across customer types, roles, regions, income bands, industries, and buyer segments. That includes the procurement managers, sales leaders, dispatchers, practice managers, and supplier reps expert networks tend to gatekeep - so you build the cohort the thesis needs, not just the experts a network can schedule. When a cohort is too narrow for the population to carry, we say so before anything runs.

It is designed to avoid MNPI solicitation by construction - synthetic cohorts have no insider knowledge to manage, log, or monitor. Cohort design is inspectable and every output is linked to a sourced research packet, so research provenance is auditable end to end.

The customer-facing brand is Thesis Lab. FishDog is the platform it is built on - "Powered by FishDog." Every contract, packet, and artifact you receive is Thesis-Lab-branded.

Stay in the loop.

Signal from our synthetic populations, product updates, and the occasional hot take. No spam.