How we test, what we run, and what we found
Every strategy idea is tested against rules written down before we look at the data, against random entries and with real trading costs. Most ideas fail, and we publish those too. This page shows what runs in our apps today, how a module is tested, our best results so far, and when each open question gets its answer.
Numbers as of 2026-10-17, exported from the research system - not typed by hand.
The apps, as they are
Trading Copilot BETA
Nine rule-based checks (fundamental, technical, sentiment, macro, order flow, regime, correlation, quant and risk) and 25 built-in strategies vote on every signal. The Consensus Score is the share of votes that agree - not a probability of profit. Every signal is recorded with its outcome.
How the Copilot works →IEB on MT5 DEVELOPMENT
Runs on our own demo accounts at IC Markets, Pepperstone, FTMO and MetaQuotes. Gold FVG (on three accounts), overnight drift and QHMR send demo orders; 4 modules run as paper observers; 240 DeCasteljau executor arms trade on two demo accounts until their keep-or-switch-off verdict; and the Meta selector scores every signal in shadow mode.
Explore IEB →Gold FVG (H1) - promising, not proven
A module that trades fair-value gaps on gold's 1-hour chart. It sends orders to two demo accounts, and it is the best live result we have.
Why it is not proven yet
- Three weeks of trades give t below 2; our bar for a validated module is t above 3.
- Both accounts take the same signals, so the second account is not independent evidence.
- Our replay of the same rule over several years of gold history, traded with the live rules, could not separate it from random entries: +0.005R per trade, t 0.29 (research question RQ-0052).
The pre-registered forward test decides: at least 30 trades, t of 2 or more, and the random-entry check. Verdict around 24 November 2026.
The Meta selector
Meta does not create signals. It scores every signal from every module and keeps only the ones it rates highest - a filter on top of all modules. If it learns which conditions make a signal work, every module benefits, which is why it is our main research line.
Version 2 runs in shadow mode: it decides before the outcome is known and records the decision, but it does not place or block orders.
Version 3 is being rebuilt on several years of history from 100 module setups, under a sealed pre-registration (RQ-0049): train on 2022-2024, validate on 2025, then one locked test. The history build finishes around 19 October 2026.
Scored 4,116 signals from 33 modules and selected 35 (0.9%).
| Group | Signals | Wins | R per signal |
|---|---|---|---|
| Selected | 35 | 54% | +0.089R |
| Rejected | 4,043 | 46% | -0.012R |
The selected signals did better, but 35 is far too few to call (t = 0.57), and all came from one module (prim_fvg_h1). First formal readout at 30 selections of the current run, around 29 October 2026.
Every module that runs, with its numbers
Signal record: R from each module's own entry, stop and exit prices, before costs. R is the result in units of the stop distance. Demo = orders on a demo account; paper = observed only. Short samples prove nothing either way.
| Module | Signals | Wins | R / signal | t |
|---|---|---|---|---|
| Gold FVG (H1)IC Markets · demo | 18 | 72% | +0.52R | 1.95 |
| Gold FVG (H1)Pepperstone · demo | 23 | 70% | +0.35R | 1.88 |
| Overnight driftIC Markets · demo | 10 | 50% | -0.05R | -0.15 |
| QHMR quiet-hour fadeIC Markets · demo | 17 | 18% | -0.45R | -3.43 |
| FVG (H1), 9 symbolsIC Markets feed · paper | 115 | 61% | +0.20R | 2.47 |
| Mean reversionIC Markets feed · paper | 3 | 33% | +0.17R | 0.17 |
| Institutional VWAPIC Markets feed · paper | 13 | 38% | -0.46R | -2.41 |
| VPA-SMC panel v2IC Markets feed · paper | 49 | 31% | -0.15R | -0.77 |
Period: 28 Sept - 16 Oct. Every signal is also published on Telegram before its result is known.
Six rules every idea has to pass
- 01
Written down first
The question, the data, the rules and the pass mark are written before any data is read, and sealed with a SHA-256 fingerprint. Later changes are dated amendments, never edits - on 17 Oct 2026, for example, USDCAD left four FX tests because our broker has no minute data for it before 2025.
- 02
One bar for everyone
Clustered t > 3.0, and at least 4 of 5 chronological windows, and at least 70% of symbols agreeing in sign - fixed before any module was run.
- 03
Beat random, not just zero
A module must beat random entries taken at the same times, with the same stop and target. A strategy that is positive only because the market trended has no edge.
- 04
Real costs
Spread, commission and slippage measured at each broker are charged. Many ideas that look positive before costs die here.
- 05
Tested as traded
Results are replayed with the live rules: when an order can be placed, where the stop and target sit after the fill, one position at a time.
- 06
A final test, used once
A last period of data stays untouched and is used once. If the idea fails there, it fails - no second try on the same data.
Everything we tested, including what failed
- 55Ideas and modules tested
- 9Candidates - real, below the bar
- 1Still collecting data
- 45Failed - published with numbers
No module has cleared the full bar yet. Publishing the failures is how we avoid selling noise: a strategy that failed our tests does not send signals to anyone.
The candidates
- QHMR - quiet-hour fade, market vs limit (IC Markets)
FORWARD TEST (demo, paired market + limit). Entries beat random by +0.15 R (t 5.8) but net -0.08 R at market costs; this measures whether limit entries close the gap.
- Overnight Drift
8,457 SPY and 6,913 QQQ sessions. Filtered Sharpe 0.90/1.09, retention 1.00/0.96. OOS t 2.36/2.47 against a 2.73 threshold - BELOW it; carried by retention and a 30-year prior, not by clearing the bar.
- Mandated Month-End Rebalancing Flow
First token to clear Gates 1-3 and its binding control. Edge +30.06 bp, 95% CI [+1.89, +58.23] excludes zero, z +2.63 vs the sign-permutation null (99.3rd pct), MaxDD -3.96% vs -15.37% for the passive control. FAILS Gate 4 on one era with t -0.24. Indeterminate, not refuted.
- Primitive: FVG (H1)
REAL-BUT-UNTRADEABLE. 10/12 positive (corr-adj t=+2.02, p=0.0193), but edge/toll 0.31x covers toll on 0/12.
- Mean Reversion
Real but weak. Passed both kill tests; below the bar.
- Institutional VWAP
NOT BACKTESTED in this project. It is an execution benchmark, not an anomaly - its value is the reference level, not a directional claim.
- FVG H1 - XAUUSD only (IC Markets)
FORWARD TEST of the one net-positive FVG cell (XAUUSD at IC, +0.07 bp net, t<1.91). IC Markets venue arm.
- FVG H1 - XAUUSD only (Pepperstone)
FORWARD TEST of the one net-positive FVG cell (XAUUSD at IC, +0.07 bp net, t<1.91). Pepperstone venue arm.
- FVG H1 - XAUUSD only (FTMO)
FORWARD TEST of the one net-positive FVG cell (XAUUSD at IC, +0.07 bp net, t<1.91). FTMO demo venue arm (1514960358, armed 2026-10-15).
When each open question gets its answer
- 17 Oct 2026Four new pre-registered tests run today: FX in New York hours, the weekend gap, drift before big releases, and premarket patterns.
- ~19 OctMeta V3: the history build finishes; training and validation follow, then the locked test.
- ~26 OctKeep, observe or switch off: a verdict for each of the 240 DeCasteljau executor arms.
- ~29 OctMeta V2: first formal readout at 30 selected signals.
- 5 NovOvernight drift: does the edge survive real costs in live trading?
- ~24 NovGold FVG forward test verdict (IC Markets).
- ~30 NovQHMR forward test: first readout of market versus limit orders.
- ~10 DecGold FVG on the FTMO demo account.
Dates with ~ are estimates; each result is published here and on the evidence register when it is in, whatever it says.
We publish methods and results; the exact rules and parameters of the modules stay private. Forward signals - recorded before their results - are measured in the Copilot's Signal Lab and posted on Telegram.