Private beta
AI research agents for systematic traders
Pragma takes a trading idea from hypothesis to code to validated backtest — end to end or one stage at a time, with guardrails built in.
Example hypothesis
- Hypothesis
- After the weekly opening gap on EURUSD, price retraces at least half of the gap within four hours.
- Rationale
- Weekend news is often over-priced at the open; liquidity returns gradually.
- Testable prediction
- Gap retracement happens more often than in a random-entry baseline.
- Data
- EURUSD 1-minute bars, weekly opens, five years.
- Acceptance criteria
- Beats the baseline after spread and commission; holds out-of-sample and across walk-forward windows.
- Status
- Rejected— edge disappears after costs.
Research is the bottleneck
Ideas pile up, tests don't.
Turning one idea into tested code takes days of manual work.
Backtests flatter.
Without costs and out-of-sample checks, results look better than they are.
Agents without limits drift.
Autonomous AI needs hard boundaries to stay useful.
From hypothesis to validated backtest
Hypothesize
Agents formulate a testable hypothesis with its rationale, the data it needs and acceptance criteria.
Implement
Agents turn the hypothesis into strategy code.
Validate
The strategy is backtested with realistic costs, then checked out-of-sample and walk-forward.
Run the full pipeline, or any stage on its own. Pick the model and provider for each agent.
- from hypothesis to backtest
- ~15 min
- hypotheses per run
- 30
Rigor is built in, not bolted on
Out-of-sample and walk-forward
Every hypothesis is tested on data it never saw.
Realistic costs
Spread, commission and slippage are part of every backtest.
Reproducible runs
Data, code and prompts are versioned, so any run can be repeated.
Bounded autonomy
Agents work within loop and budget limits.
Any model, any agent
Choose the model and provider for each agent. Pragma works with Claude (Anthropic), OpenAI models and other providers. Market data and execution run through MetaTrader 5.
Engineering notes
Validation that doesn't flatter: costs, out-of-sample and walk-forward
Why most backtests overstate results, and how Pragma's checks are built to reject ideas that only look good.
Why our research agents talk through a single API contract
Decoupled stages, outputs you can inspect, and the upfront design cost we accept in exchange.
Bounded autonomy: loop and budget limits for AI agents
Why autonomous loops drift, and why our agents stop at hard limits instead of retrying forever.
Who builds Pragma
Pragma is built by specialists in data science, machine learning, trading and automation: the disciplines a research pipeline depends on.
Data science
Honest baselines, held-back data and tests that have to hold out-of-sample.
Machine learning
Agents that turn an idea into a testable hypothesis and the hypothesis into strategy code.
Trading
Spread, commission and slippage counted from the first run.
Automation
A pipeline that runs end to end or stage by stage, inside loop and budget limits.
Join the private beta
We review requests weekly and reply by email.
Request access