Research that ships to your account.
The Desk is where strategies are born before they ever touch a live account. Discover factors, train models, validate them honestly with walk-forward testing, and experiment with reinforcement-learning agents — then promote what survives straight into the Council and Backtest. Research and production aren't two worlds here; they're one pipeline.
Every idea, tracked and testable.
Runs are versioned, scored on out-of-sample data, and comparable at a glance. The board below is an illustrative view of how experiments are organised.
| Run | Approach | Metric | State |
|---|---|---|---|
| #07-a | Momentum factors | out-of-sample Sharpe | PROMOTE |
| #07-b | Gradient-boosted model | walk-forward | review |
| #06-c | Mean-reversion | cross-validated | ARCHIVE |
| #06-a | RL execution agent | reward curve | training |
A research bench with guardrails.
Powerful enough to discover something real, disciplined enough to stop you fooling yourself. Honest validation is the default, not an afterthought.
Find the edge
Search for signals across price, fundamentals and alternative data, with tooling that flags overfitting and multiple-testing traps before you trust a result.
Build & tune models
Train classical and machine-learning models on your features, with reproducible runs and clear metrics — not a one-off notebook you can't rerun.
Validate honestly
Walk-forward and cross-validation that respect time, so a strategy is judged on data it never saw — the only test that counts.
RL strategy R&D
Experiment with reinforcement-learning agents for entry, exit and execution, in a sandbox before anything reaches real capital.
Reusable features
Engineer a feature once and reuse it across experiments and models — consistent inputs from research all the way to live.
Ship to production
When a strategy survives out-of-sample, promote it directly into a Council preset or the Backtester — no rewrite, no translation.
The path a strategy takes.
Every promoted system walked this path first — which is exactly why you can trust it on the chart.
Hypothesise
State the idea and the market it should work in. A clear thesis is what makes a result interpretable later.
Build features
Assemble inputs from the Data Fabric into reusable features, with leakage checks built in from the start.
Train & validate
Fit models and score them on held-out data. Honest cross-validation keeps optimism out of the numbers.
Walk-forward
Roll the test through time to see how the edge holds as conditions change — the closest thing to live before live.
Promote
Ship the survivor into a Council preset or the Backtester, and watch it on your own account.