How it works.

One engine runs every strategy in the model portfolio by fixed rules, and nine tests stand between an idea and the portfolio. Here is each in full: one trade as the engine logged it, then the tests as formulas.

t = SR \cdot \sqrt{T}

Confidence grows with time.

The engine, trade by trade

One engine runs every strategy in the model portfolio. Here is one backtest trade, as the engine logged it and as it looked on its chart.

engine.logmodel account, backtest replay, 14 to 17 February 2022

14 Feb 00:00bar4-hour bar closed # the strategy reads its own chart
14 Feb 00:01signallong # model portfolio, dial 1.0×
14 Feb 00:01sizeon model equity, inside the risk budget
14 Feb 00:01orderentry MARKET filled # model account
14 Feb 00:01orderstop STOP working # placed on fill
14 Feb 00:01ordertarget LIMIT working # placed on fill
14 Feb 00:01ledgerposition open matches model account
17 Feb 02:36filltarget LIMIT filled # target reached
17 Feb 02:36orderstop cancelled # one-cancels-other
17 Feb 02:36ledgerflat matches model account

One trade from a backtest replay of the model account, as the engine logs it, 14 to 17 February 2022. Sample data, not results: see how results are shown.

The same trade on its 4-hour chart: the entry on 14 February, the stop below, the target above, reached on 17 February. Sample data, not results.

The path of every trade

The five steps of every trade: signal, size, stop and target, publish, reconcile. If you switch copying on, an independent third-party trade-copying provider (“the provider”) copies each position into your account as it happens, sized by your dial, and closes it when the model closes it.

Nine tests before a strategy joins

We publish how each strategy is tested, not just that it passed. A strategy is judged on years it never saw, and its result is marked down for every version tried before it.

[Salvatore to confirm: all nine in force for every strategy in the portfolio]

0Confidence grows with time
$t = SR \cdot \sqrt{T}$
Signal over noise, measured over longer: the same edge over more time, or over more independent trades, is harder to explain by luck.
1The entry is tested alone
$E_{N} = \frac{\sum_{i} \frac{{MFE}_{i, N}}{{ATR}_{i}}}{\sum_{i} \frac{{MAE}_{i, N}}{{ATR}_{i}}}$
Before any exit exists: N bars after each signal, how far did price go our way against how far it went against us?
2Judged on years it never saw

in-sampleout-of-sample 1out-of-sample 2
$WFE = \frac{1}{K} \sum_{k = 1}^{K} \frac{R_{OOS, k}}{R_{IS, k}}$
Built on the first window, then judged on K later windows it never touched: return out of sample as a share of return in sample.
3It has to work on related markets
$\Pr [pass ≥ k of m ∣ no edge] = \sum_{i = k}^{m} (\binom{m}{i}) p^{i} {(1 - p)}^{m - i}$
A rule with no edge seldom passes on several markets at once. Related markets move together, so the true chance of passing by luck is higher than the formula gives: the formula is a lower bound, and the test is read with that in mind.
4Small changes to its settings
$ρ = \frac{\underset{θ^{'} \in N (θ)}{median} S (θ^{'})}{S (θ)}$
a peaka plateau

Every setting nudged up and down. Nearby settings should score almost as well: a plateau, not a peak.
5Thousands of unlucky versions of its past
$D_{95} = Q_{0.95} [\max_{t} (1 - \frac{E_{t}^{(b)}}{\max_{s \leq t} E_{s}^{(b)}})]$
Trades reshuffled and a share removed at random, across thousands of runs (b). We plan around the bad tail of that spread, not the backtest.
6Thousands tried, luck priced in
$DSR = Φ [\frac{(\hat{SR} - {SR}_{0}) \sqrt{T - 1}}{\sqrt{1 - {\hat{γ}}_{3} \hat{SR} + \frac{{\hat{γ}}_{4} - 1}{4} {\hat{SR}}^{2}}}]$
N tried1001,0002,000DSR99.7%97.2%95.6%

The result must beat the best that N tries with no edge would show, skew and fat tails included. Bailey and López de Prado (2014).
7The bad tail, measured directly

${ES}_{0.95} = E [r_{t} ∣ r_{t} \leq q_{0.05}] = - 4.3 %$ ${\hat{γ}}_{3} = + 0.63$

The average of the worst 5 of 101 months, not a single number at a cut-off. A positive skew: the largest months were gains.
8Not the same bet twice
${\hat{SR}}_{c} > {\hat{ρ}}_{c, P} \cdot {\hat{SR}}_{P} {\hat{ρ}}_{c, P} = corr (r_{c}, r_{P})$
It joins only if its Sharpe ratio beats its correlation with the portfolio times the portfolio’s own. It must add edge, not repeat a bet.
9Sized to its worst case, then sized together
$f_{B} = \max {f : D_{95} (f) \leq B}$
Each strategy gets the largest size f whose 95th-percentile drawdown stays inside a fixed budget B. Then the whole portfolio is tested the same way, all strategies at once.

The pipeline in nine formulas, v1.0, October 2026. The figures in it come from the sample series. Past and simulated performance is not a reliable indicator of future results. See how results are shown.Sample data – not results

About 1 in 200 candidates passed every test

Each dot is one candidate. Sample pipeline run, not a count of anything live.

A thousand candidate strategies tested, five passed every test. Illustration from the sample pipeline run.

The pass marks are fixed in writing before a strategy is tested. [Salvatore to confirm: publish the pass marks]

Questions about the engine

Anything else: ask on the introduction call.

Why these markets?

Forex, stock and futures markets are deep and trade most of the day, and going short is as easy as going long. Futures are cleared by an exchange: the same contract for everyone.

What if the engine fails mid-trade?

It saves its state, so a restart picks the model portfolio’s open positions back up. Anything it cannot reconcile stops new trades from being published and alerts our team.