Weeks of plumbing before the first result
Building a leakage-safe pipeline, a cost model and a validation harness from scratch for every idea is weeks of infrastructure before the question gets asked.
For trading analysts
Kill the idea early, or earn the right to build it.
A strategy is a versioned specification, not a live position. Validate it against its own trial count and its own transaction costs before anyone spends a sprint building the version that trades.
01A sprint with Quant Compose
From a raw price history to a kill-or-promote decision, without a single line of execution code.
Day 1
Features and labels
Build price, statistical, volatility or microstructure features, then label outcomes with triple-barrier labelling — which of profit-take, stop-loss or holding-period-expiry happened first — or meta-labels judging whether a bet a primary model already decided was worth taking net of cost.
Feature Engineering Feature preview for AAF.L: the engineered columns row by row. Day 2
Rules in the strategy builder
Turn a model's probabilities into explicit rules: threshold, direction, position sizing, stop-loss, take-profit, trailing stop, cooldown and a drawdown halt. The same specification compiles whether it came from the visual builder, a formula editor or the API.
Strategy Builder Strategy Builder, Signal tab: the final model, a long threshold of 0.55, minimum signal strength, optional entry and exit conditions, and the signal workflow from probability to net return. Day 3
Backtest net of cost
The event-driven engine — the only one modelling partial and no fills — charges commission at 5.0 basis points and slippage at 2.0 basis points by default, plus an optional borrow cost, on every executed change in exposure.
Backtesting Strategy capital against buy-and-hold over the held-out window, net of a 0.1% cost per unit turnover. Day 4
Cost sensitivity and the verdict
Re-run across cost multiples to find where the edge disappears, then read the Deflated Sharpe Ratio and Probability of Backtest Overfitting against the trial count the ledger actually recorded — not the one trial you liked best.
Backtesting The Robustness panel after validation: the verdict is insufficient evidence, the statistics it computed are listed with their sources, and Deflated Sharpe is reported as not available rather than estimated. Day 5
Kill it or promote it
Reaching validated requires a validation record showing the result survives its trial count; reaching approved requires a recorded human decision naming the evidence it rests on. Neither step can be set by an AI, and neither can be skipped.
Strategy Monitor Registered strategies with status and last-computed performance.
02Three problems it removes
Named specifically, not as a category.
In-sample optimism
A backtest tuned on the same data it is judged on will look better than it is — the gap only shows up out of sample, if anyone bothers to check.
No record of trials
Without a ledger, the tenth configuration tried looks identical to the first when it is presented — and its apparent edge is inflated by everything tried and discarded along the way.
03A gated lifecycle
Method that holds up when someone checks.
A strategy cannot talk its way to approved. Two of its stages require evidence a human, not a model, has to supply.
- Validated requires a validation record
- The lifecycle runs draft through research, backtested, validated, paper, approved, live, monitoring, paused and retired. Reaching validated needs a validation record showing the result survives its own trial count.
- Approved requires a recorded human decision
- Reaching approved needs a decision naming the evidence it rests on, recorded explicitly. An AI cannot set either of these two gates.
- Event-driven backtesting with partial fills
- The only engine in the catalogue that models partial and no fills, processing each bar in order through an order queue and cash ledger — the vectorised engine is faster but assumes fills at the modelled price.
- Transaction-cost sensitivity, not a single number
- Re-running across cost multiples shows where the edge disappears, rather than reporting one backtest at one assumed cost.
04Objections, taken seriously
The questions this audience actually asks.
Does it execute?
No, and that is the point. Nothing in the workspace places or routes an order — the health endpoint reports live order routing as disabled. This is the stage before an order exists, where a bad idea is meant to die cheaply.
Intraday?
The workspace works on daily bars. The bundled 2010 flash-crash episode carries an explicit caveat that daily bars cannot see an intraday liquidity event — a limitation stated, not hidden.
Can I export the strategy?
Yes — as a versioned specification document, the same object the visual builder, formula editor and API all compile to. It is a specification for review, not a connection to a broker.
05In beta, for you
What is live for trading analysts during the beta.
- The full strategy builder, event-driven backtest engine and cost-sensitivity analysis, at no cost
- Triple-barrier and meta-labelling in the feature and labelling catalogue
- Deflated Sharpe Ratio and Probability of Backtest Overfitting on your own trial ledger
- The gated lifecycle from draft through approved, with write-once validation verdicts
Other audiences, and further reading
Test the idea before it costs a sprint.
Beta researchers compose a strategy, backtest it net of cost, and read the verdict before committing engineering time.