Private beta guide¶
Thank you for testing Market Lab before its public release. This page takes you from installation to a finished, reviewable study in about an hour, without help. Please send us what was unclear, slow or wrong at each step.
What we ask you to do¶
- Install Market Lab and run the demo.
- Run one study: an Agent (with your own model, or a local model) against a rule and buy and hold, on a recording you choose.
- Review it: evaluation, decision timeline, PnL attribution and a model risk report.
- Tell us where you got stuck, using the questions at the end.
1. Install¶
Use the installer for your system, from the link we sent you:
| System | File | First start |
|---|---|---|
| macOS 12 or later (Apple Silicon) | Market-Lab-<version>-macos-arm64.dmg |
Open the image and drag Market Lab to Applications |
| Windows 10 or later (64-bit) | Market-Lab-<version>-windows-x86_64.exe |
Run it; it installs for your user only, without administrator rights |
| Linux (x86-64, glibc 2.35 or later) | Market-Lab-<version>-linux-x86_64.AppImage |
chmod +x the file and run it |
Check the download against SHA256SUMS. During the beta the Windows installer is not
yet signed, so Windows may warn before running it. For scripting, install the Python
package instead: pip install market-lab-research gives the
market-lab command, with no C++ build needed.
Start the App. To try it on a small fixture recording first, run market-lab desktop
--demo (from the Python package), or market-lab demo for the command-line check,
which creates a temporary workspace, trades the fixture and removes it again.
2. Run a study¶
- Data → Import: add a price-bar CSV (format) or an
ITCH file. To keep a model from recognising the period, disguise the recording
first (
market-lab disguise); see evaluation. - Agents → New Agent: start from a reference Agent or your own Python project. Under Model provider choose OpenAI (store your key under Settings → Model credentials) or Local server for vLLM, Ollama or llama.cpp on your machine, so no data leaves it (research controls).
- New experiment: choose the recording, the Agent and Turn-based pacing, which measures decisions without speed and runs faster than the market window.
- Repeat for the baselines: a rule run and a buy-and-hold (hold) run on the same recording, cash and costs.
Before paying for a real model, rehearse the Agent: set its model name to rehearsal
in a script, or add "rehearsal": "SYMBOL" to a market-lab experiment configuration.
To cap spending, add a budget to the Agent settings.
3. Review it¶
- Compare runs: choose the Agent's runs and the baselines; read the intervals, p-values and the deflated Sharpe ratio, not only the means.
- Results → Agent decisions: each order with what the Agent read before it.
- From Python or the command line:
pnl_attribution(alpha, beta, costs),diff_runs(why two runs differ) andmodel_risk_report(one reviewable document):
market-lab risk-report --profile PROFILE --runs AGENT_RUN RULE_RUN --output review.md
market-lab verify --profile PROFILE --run AGENT_RUN
Study tools explains every figure. A study can also run entirely from the command line; see the command-line guide.
Known limits during the beta¶
- Execution is simulated against recorded displayed liquidity; read the methodology before quoting results.
- Market Lab never sends real orders and has no broker connection.
- Windows and Linux installers are new; please report anything that differs from macOS.
- Team workspaces, roles and a shared database are not part of the beta.
Feedback we need¶
- Where did you first need help, and what did you try?
- Which step took longest, and was the wait explained?
- Did any result look wrong or impossible? Send the run's model risk report.
- What would you need before trusting a result from Market Lab in your own work?
- What is missing for your usual study?
Report problems with the run ID, the model risk report and, when the App failed, the
log from market-lab --log-level debug .... Do not send API keys or licensed data.
Section index · All documentation