Every edge we trade and every study we publish goes through the same process. Here it is, step by step.
The data is divided before anything is tested
Test a strategy on the same history it was built on and it will always look good. So before any result is seen, the data is split into separate periods, and each one answers a single question.
2019 – 2022DiscoveryThe only period where an idea can be explored and shaped.
2023 – 2024ConfirmationRules frozen, then checked.
Jan 2025 – Jun 2026Out-of-sampleNever used to build anything. Tested once.
Jul 2026 → todayLiveCompared against the backtest and monitored for anything that looks off.
The split is enforced in our software, not left to discipline. Loading the out-of-sample data during research raises an error unless it is deliberately unlocked for the final test.
From an idea to a live edge
Ten steps, in order. Most ideas fail at step 4, and an idea that fails any step stops there. It isn’t adjusted and sent back through.
The idea
1
Write the idea down firstPre-registrationThe question, the measurement, the expected direction and the comparison are recorded before any data is touched. The plan is never edited to fit the result.
2
Code it into exact rulesSystematic rule specificationEntry, stop, target and exit, written so a computer executes them identically every time. A rule that needs judgement can’t be tested.
Real, or luck?
3
Measure it against chanceMatched control, bootstrap confidence intervalsEvery result is compared to a control, such as the same trade taken at a matched but meaningless moment. Only the difference counts, and it’s reported with a confidence range.
4
Discover, then confirmHold-out confirmationIdeas are shaped on 2019–2022 only. With the rules frozen, they run on 2023–2024. Most ideas fail here.
5
Check every stretch of historyCombinatorial purged cross-validation (CPCV)2019–2024 is re-split fifteen different ways. An edge has to hold up in most of them, not just on average.
6
Count the attemptsDeflated Sharpe ratio (DSR), PBOTry enough variations and one will look good by chance. The more we try, the higher the bar a result has to clear.
Real costs
7
Real costs, real fillsTransaction cost modelling, tick-level fillsCommission and slippage on every entry and exit, at the size actually traded. One-minute bars describe a pattern; the verdict is made on tick-by-tick data.
Final exam
8
One out-of-sample testSingle-shot out-of-sample testThe frozen strategy is run once on January 2025 to June 2026. If it fails, it’s done. A second attempt on the same data is no longer a test.
Going live
9
Live must match the backtestLive–backtest parityBefore it trades, the live code is replayed over history and must reproduce the tested trades one for one.
10
Compare live to the backtestLive performance monitoringOnce an edge trades, its live results are compared against what the backtest says should have happened, and the whole system is monitored for anything that looks off. An edge that drifts from its tested behaviour is reviewed and can be removed.
The studies on our Research page use the same controls and the same costs. Our own edges go through every step.
What this does and doesn’t mean
Passing this process is strong evidence that an edge was real over the history we tested. It is not a guarantee that it will keep working. Markets change, and any edge can stop working.
Backtested and simulated results are hypothetical. They are not live trading and past performance doesn’t predict future results. Where we publish a number, we say whether it came from a backtest, a simulation or live trading.
See the process applied.
Popular strategies put through these tests are on the Research page. Our own edges’ full track records, after costs, are on Performance.