Imagined Markets. Real Consequences.
Almagest
We build world models of financial markets that answer one question: what happens if I act? A world model is a learned simulator. Given the state of a market and a proposed action, it imagines what follows, so decisions can be rehearsed before they are made. And because a decision is only as good as its reasons, every imagined future can be opened up and explained.
Ptolemy’s Almagest predicted the motion of the planets for fourteen centuries. Its model of the sky was wrong, but its predictions were held to the measurements. We hold market models to the same standard: judged only by what they predict, on data they have never seen.
Most machine learning in finance tries to forecast prices. We model something else: the consequence of an action. Every order consumes liquidity and moves the market it trades in. That response is mechanical, depends on the state of the market, and can be learned.
An action-conditioned world model takes the market’s state and a proposed decision, and imagines the path that follows. A desk can ask it the counterfactual: what would this trade have cost in a different market? An agent can learn inside it, rehearsing thousands of scenarios before it acts once.
The most useful thing a simulator can tell you is where it stops being right. We build models that know the edge of their own experience, and that fall back to simple, proven laws when the market moves beyond it.
Almagest will develop applications where the cost of a wrong decision is real: execution, pre-trade cost estimation, stress testing and risk.
5 November 2026, a real market day. Take a simple “buy the dip” rule, run it through the eighteen most plausible worlds that could follow, see where it makes money, where it loses, and why, and then fix it.
Why interpretability, and why finance first
A model that predicts well but cannot say why is a liability wherever money is at stake. We build world models whose imagined futures can be opened up: which part of the state drove a prediction, which action caused which consequence, and where the model has stopped being on familiar ground.
We are starting with finance because it is where world models and interpretability meet most naturally. Every action has a measurable consequence. Every decision is recorded. The cost of a wrong or unexplained decision is immediate and real, and the people who carry that cost, and those who regulate them, need to know the reason.
We think this is how one should trade: know what your action will do to the market before you take it, know why the model believes it, and know when it does not know.
Questions a world model can answer
What would this trade have cost?
Replay a past order in a different market state and see the price path it would have caused.
How should this order be sliced?
Plan the size and timing of each child order under a learned model of impact.
Learn before acting
Train decision policies on imagined market paths instead of on live capital.
What if the market breaks?
Push a book through a liquidity crisis and find where its assumptions fail.
We share one belief: understanding a market starts with knowing what your own actions do to it.