We specify a proprietary trading firm in which every trader is a self-learning AI agent, and in which the firm, rather than any single agent, is the unit of design. Our agents have careers. Each is admitted to a school, educated in market theory, examined on scenarios it cannot have seen, hired into a cohort, deployed first to replay and then to paper trading, reviewed daily, and then cut or promoted. The most successful become teachers, and the next generation is educated on the firm's validated research and the case files of its best traders.
The paper makes no performance claim. Its most important deliverable is a clean answer to whether self-learning agents beat frozen copies of themselves at all.
1The idea
A human trader is not born with an edge. They read, they are taught, they are put on a desk with a small amount of risk, and they make mistakes in front of people who are paid to notice. Over months and years they either develop a repeatable process or they are let go. A proprietary trading firm is the institution that runs this process at scale. It doesn't need every trader to succeed. It needs a reliable way to tell, early and cheaply, which traders have skill.
Our thesis is that this institution can be rebuilt with AI agents as the traders, at a far faster pace, and that doing so produces more than a possible source of returns. It produces a scientific instrument: a firm that can run controlled experiments on its own staff.
Most "AI hedge fund" designs split one decision across fixed roles that debate it. Here the unit of design is the career and the firm.
2A career, not a role
Agents begin as candidates with no track record. They pass through school and examinations, trade on replayed and then live paper markets, and are reviewed every day. Capital, privileges and teaching rights flow to those whose skill is statistically established, and away from those whose skill is not. The best become teachers, so each generation is educated on what the last one proved.
An agent's behaviour at month twelve should differ from its behaviour at month one for reasons we can name, because everything it has learned is written down.
3Scoring skill, not luck
Trading outcomes are dominated by noise. Profit and loss cannot separate skill from luck on any horizon a firm can afford to wait, so a population of agents ranked by profit is a machine for promoting luck.
The firm therefore scores agents on pre-registered, probabilistic forecasts under strictly proper scoring rules, corrected for the number of agents being tested. Profit is kept as lagging confirmation rather than the primary measure.
4Clean evaluation
A large language model has absorbed most of market history before it ever trades. On dates inside its training window, apparent forecasting skill partly reflects recall of the outcome. The only honest test is data that post-dates the model's training cutoff.
Agents are therefore evaluated only on fresh data, through a data layer that enforces point-in-time access, and that scarce data is protected rather than spent on tuning.
5Experiments only agents allow
Because an agent's knowledge lives in versioned text rather than in model weights, the firm can do three things no human firm can.
Frozen controls
Snapshot an agent at graduation and keep a copy that never learns, to measure whether learning helps at all.
Lesson ablation
Clone a trader, remove one lesson, and measure its causal effect on the same market at the same time.
Counterfactual replay
Re-run any past decision with one input changed, to audit what actually drove it.
6Research programme
The paper closes with eight falsifiable hypotheses. Each one is useful to the firm whichever way it comes out.
| ID | Hypothesis | Primary test |
|---|---|---|
| H1 | Learning agents beat their frozen controls on forecast skill. | Paired daily difference in skill score, learning agent against its own snapshot. |
| H2 | Different curricula produce less correlated agents. | Correlation of forecast errors within and across schools. |
| H3 | The morning meeting changes forecast quality. | Skill before and after the meeting, same agents, same days. |
| H4 | Cohorts taught by teachers outperform other cohorts. | Time to graduation, and skill at a fixed tenure. |
| H5 | The value of a lesson can be measured by paired ablation. | Ablation effect against later results of agents that adopt the lesson. |
| H6 | Forecast skill predicts later profit. | Rank correlation between early forecast skill and later profit. |
| H7 | Agents' stated reasons are only partly faithful. | Stated rationale against removal and perturbation tests. |
| H8 | Agents add information beyond the firm's own model. | Skill score against the house model, by market regime. |
Every experiment is pre-registered with its metric, sample size and analysis, uses paired designs wherever possible, and is reported whether or not it finds anything.
7Scope and status
- The first market is spot gold, with front-month gold futures as the reference contract, chosen for its liquidity and its clear macro drivers.
- Execution is paper-first. Small live capital comes only after agents are promoted on forecast skill.
- Everything in the paper is a design proposal. It makes no claim about performance.
The full paper is available on request. Ask for a copy.