What building Marketry taught me about forecasting

Role

Co-founder and CEO

Since

2026

Site

marketry.xyz

Customers

Mid-size companies

Recognition

OakNorth pitch finalist

Funding won

$7K non-dilutive

Marketry is a company I co-founded with Johnny Dawson, who leads product and engineering, and Nicholas Lobkowicz, who leads operations and compliance. I'm the CEO. Mid-size companies, roughly $500 million to $5 billion in revenue, carry risks that are hard to insure: a construction delay, an energy price spike, a shipping disruption, a shortage of data-center parts. Marketry lets them price and move those risks with event-linked contracts that pay out on a measurable outcome, the way parametric insurance does, with forecasts underneath.

It is our third attempt at the same question. Hindswipe was a consumer prediction market. HedgeGuard moved to institutions, routing data and orders. Marketry kept the institutional customer and put the forecast at the center.

The research

• We mapped the parametric insurance market and its main players, including Descartes, Arbol, FloodFlash and Parametrix, along with insurance-linked securities and how prediction markets are regulated in the EU.
• We designed our broker discovery calls around The Mom Test and Disciplined Entrepreneurship, so interviews asked what people had actually done, not what they thought of our idea.
• We built our own forecasting pipeline and graded it. The findings below come from that work.

What we found

The language model explains. It doesn't forecast.

In June 2026 we froze one rule for the whole system: evidence becomes observations, observations update a state estimate, a simulation turns the state into a probability, and only then does a language model explain the number. A language model never produces the probability itself. The rule is enforced in code, not just written down.

Five graded forecasts prove nothing

Our first published scorecard covered 5 graded outcomes: an 80% hit rate and a Brier score of 0.2728. Simple constant baselines beat it. A high hit rate on a small sample was mostly luck, and we now report sample sizes next to every score.

1,440 agents were really about one

We simulated a fleet of 1,440 trading desks. Their effective number of independent opinions was 1.12. It tracked how many independent pieces of evidence went in, not how many desks there were. Desks reading the same evidence add confidence, not information. That matters for my own favorite idea, that many small agents with partial views can out-forecast one big one: the diversity has to come from the evidence, not from personas.

Honest grading is slow

Only about 121 to 800 outcomes a year in our area can be graded honestly, meaning forecasts sealed before the fact and scored against criteria fixed in advance. At that rate, proving calibration takes years.

So the product is the grading

Sealing forecasts before outcomes, grading forward only and publishing the effective sample size are the parts that last. A big fleet of agent traders is a hypothesis we can't yet afford to test properly.