From research papers to tested ideas.
We read the literature, examine the theories behind it and turn claims into experiments. Research agents help us reproduce studies and test their assumptions. The output is a record of the evidence, including the results that do not hold.
A paper is the beginning of the work.
Before developing a model, we need a clear question, a plausible explanation and a test that someone else could repeat.
- UnderstandIdentify the claim, the proposed mechanism and the conditions under which the theory should hold.
- ReproduceReconstruct the data, code and published experiment. Record missing details and implementation choices.
- ChallengeVary the assumptions, compare simple alternatives and test other periods. For trading ideas, account for costs and feasible execution.
- DocumentKeep the sources, configuration and findings together. Negative and inconclusive results also shape the next question.
Agents that carry the research forward.
Our agents read papers, reconstruct experiments and run backtests. The system evaluates its research strategies and retains useful changes to how it explores, using recorded experience to inform the next experiment.
Two loops. One research process.
The agent extracts the hypothesis, instruments, sample period and procedure, turning the paper into an explicit experiment specification.
The feedback loop improves the research procedure. Held-out evaluation stays separate from the experience used to select policy updates.
This loop improves the research procedure. Training the weights of a language or market model is a separate task, covered in Models.
The questions connect.
Our reading spans company fundamentals, macroeconomics, market structure and machine learning. The map connects these topics; it helps organise questions rather than rank investment opportunities.
What we learn, written down.
Literature reviews and working notes, with sources, assumptions and the scope of each result.
Give the idea a model.
Research defines the question and its limits. The next step is to train a model or specify an algorithm, then evaluate it on data kept outside development.