Start with a hypothesis
Describe the market, time horizon, entry and exit logic, and risks. The AI research agent turns your direction into a structured hypothesis.
Strategy research
Turn an idea into a testable strategy.
Idea development is the first stage of the factory. AI research and coding agents turn a hypothesis into strategy code; tests and reviews guide later revisions.
See the strategy factory in motion.
Illustrative example · not a live result
Watch AI agents build and review a strategy.
AI agents develop the idea, write code and assess the test results. Review findings can send the strategy back for another iteration.
Watch the path, not just the endpoint.
An illustrative equity curve reveals setbacks along the way. Inspect drawdown and trading costs before interpreting a result.
Keep open positions in the picture.
Compare account equity with closed-trade balance. An open position can change equity before a trade closes.
- Describe
- Build & verify
- Backtest
- Stress-test
- Paper-trade
EquityClosed-trade balanceDrawdown
Net return—
Max drawdown—
Checkpoint—
Build with context
An AI coding agent implements the strategy using available data and configured venue rules. Automated verification and an AI QA agent check it before backtesting.
Learn from each run
Keep hypotheses, iterations, results and review decisions together. Your workspace learns from its own runs; paid plans also use the platform research corpus.
A research brief with a clear next step.
Start with the reason a strategy might work. Define what would disprove it, which data it needs, and the costs it must overcome. The factory records the resulting hypothesis and its development history.
Direction
Market, instrument, timeframe
Hypothesis
Expected edge, rules, risks
Reviewable work
Code, checks, iterations
Give the AI research agent a question it can test.
A specific brief makes it easier to judge both the implementation and the outcome. Define the idea before looking for the most flattering result.
Specify the evidence that would make you stop. An unsuccessful experiment can still answer a useful question.
- Market & horizon
- BTC spot, hourly bars; study short-term mean reversion.
- Proposed mechanism
- Large deviations from a rolling average may partly reverse.
- Rules to define
- Entry deviation, exit condition, position size and maximum holding time.
- What could disprove it
- The effect disappears after costs, or depends on one unusual period.
Change one assumption. Learn what matters.
Keep the comparison focused so you can explain why a result changed. Use the run history to connect each revision to its evidence.
Scroll horizontally to see all columns.
| Experiment | Change | Question |
|---|---|---|
| Baseline | Use the original entry and exit rules. | Is there an effect after modeled trading costs? |
| Robustness check | Try nearby entry thresholds. | Does the effect survive a small parameter change? |
| Risk revision | Add a maximum holding period. | Does it limit long losing positions without removing the effect? |
Example research design; these are not reported runs. Repeatedly selecting the best result can overfit the same history.
Interactive illustration
Look for stability around a setting.
A broad response across nearby settings gives you a different research question from an isolated peak. Move the lookback setting to compare both patterns.
Read the neighborhood
At 25 bars, the isolated peak looks strongest. Its neighboring settings weaken sharply. Check whether the selected result depends on one fortunate choice.
Broad responseIsolated peak
Illustrative Sharpe ratio
- Broad response
- Isolated peak
Questions, answered.
More questions and answersDoes every idea become a strategy?
No. Data limitations, failed checks, poor economics or review findings can cause a run to iterate or be abandoned.
Can I inspect public examples first?
Yes. Browse public strategies and their available reports before opening your own workspace. Marketplace code access depends on the publisher and your access rights.
What makes an idea testable?
A testable idea names an expected market behavior and gives measurable rules for entering, sizing and exiting a position. It also identifies the data needed and the evidence that would disprove the proposed edge. State those assumptions before comparing results.
How is the generated strategy checked?
Verification includes checks on the code, declared instruments and supported behavior, followed by simulation and QA review. The available reports show findings that need attention. These checks help identify implementation problems; they cannot establish that every economic assumption is correct.
Can I see what changed between attempts?
The strategy report and iteration history connect attempts with their results and review feedback. Read which assumptions or parameters changed before comparing performance. Source-code visibility follows the access rules for the strategy.
Does the research use lessons from earlier runs?
The research context includes earlier successes and failures available to the workspace. Free plans use their own run history; paid plans can also use the platform research corpus. Access to that corpus does not grant access to every other user’s private strategy code.
Start with a question worth testing.
Explore published strategies, or open a workspace to build, test and monitor your own with AI agents.
Stratmill is an AI trading strategy platform, not financial advice or a broker. Backtest and paper results are hypothetical. Trading involves risk of loss.