From Research to Publication

Stage 01 — Research

AI Research

Every day, AI agents scan more than 1,000 sources — quantitative finance papers, GitHub repositories, trading blogs, academic journals, and community forums — in search of actionable strategy logic.

Each candidate strategy is analysed by LLMs to extract entry/exit rules, risk parameters, and the underlying market hypothesis. A confidence score is computed; only strategies scoring 65 or above advance to code generation.

Minimum confidence score: ≥ 65 / 100
Deduplication via Postgres (seen_urls + semantic hash)
Output: structured JSON with full strategy specification
DeepSeek / Qwen LLMs · n8n orchestration
01
02
Stage 02 — Generation

Code Generation & Security Audit

Strategies passing the research gate are converted into production-ready MQL5 Expert Advisor code and a companion Python implementation, both generated by a code-specialised LLM from the structured strategy JSON.

Before any file is saved, an automated security audit scans it for dangerous patterns: hardcoded credentials, unrestricted position sizing, missing error handling, memory leaks. Only files scoring 8.0 out of 10 or higher pass.

Security audit score: ≥ 8.0 / 10
Output: .mq5 + .py stored in S3 with versioning
Audit failure → status audit_failed + Slack alert
Claude Opus API · AWS S3 · Notion metadata
Stage 03 — Validation I

Python Backtest

The Python implementation is executed on an on-demand Linux VPS using VectorBT or Backtrader. This stage is deliberately fast and inexpensive — its job is to eliminate the bulk of underperforming strategies before they reach the more resource-intensive MT5 tester.

Testing uses several years of OHLCV data at the strategy's native timeframe. Results are evaluated against fixed thresholds. The VPS is terminated immediately after the test run to minimise cost.

Sharpe Ratio: ≥ 1.0
Win Rate: ≥ 50%
Eliminates approximately 80–90% of remaining candidates
VectorBT / Backtrader · Hetzner on-demand VPS
03
04
Stage 04 — Validation II

MT5 Strategy Tester

Strategies surviving the Python backtest gate are compiled and run in MetaTrader 5's Strategy Tester using every-tick simulation on real historical tick data. This is the most demanding gate — it accounts for spreads, slippage, commission, and swap as they actually occurred on a live broker feed.

The test runs on an on-demand Windows VPS (Contabo) which is booted before the run and shut down immediately after. All results are uploaded to Notion and S3 for archival.

Win Rate: ≥ 55%
Maximum Drawdown: ≤ 20%
Profit Factor: ≥ 1.2
MT5 Strategy Tester · Contabo Windows VPS · every-tick simulation
Stage 05 — Forward Test

30-Day Paper Trading

Strategies passing the MT5 tester gate are deployed on a live MetaTrader 5 demo account and allowed to trade for a minimum of 30 calendar days. This forward test uses real-time market prices, actual broker spreads, and realistic slippage — conditions that no historical simulation can fully replicate.

Paper trading monitors the EA in real time. Any breach of the drawdown ceiling or profit factor floor during this period results in immediate rejection, regardless of earlier results.

Duration: ≥ 30 calendar days
Profit Factor maintained: ≥ 1.2
No account ruin, real spreads & slippage applied
MT5 Demo Account · real broker feed · automated monitoring
05
Stage 06 — Published

Available in the Shop

Only strategies that survive all five preceding stages — research scoring, security audit, Python backtest, MT5 Strategy Tester, and 30-day paper trading — are published in the QIS shop.

Every published EA comes with its complete validation report: all quality gate results, test parameters, historical performance charts, and a clear statement of which metrics come from simulation versus forward testing.

Full validation report included with every purchase
All source metrics labelled: backtest / forward test / paper trade
Fewer than 5% of discovered strategies reach this stage

Quality Gates at a Glance

All thresholds are fixed and applied automatically. No human override. No exceptions.

Gate Stage Threshold Action on Failure
Research Confidence Score 01 — AI Research ≥ 65 / 100 REJECTED — strategy archived, status rejected
EA Generator Trigger 01 → 02 Transition ≥ 65 / 100 Score 60–64 → archived only, no code generated
Security Audit Score 02 — Code Generation ≥ 8.0 / 10 REJECTED — code blocked, Slack alert, status audit_failed
Python Sharpe Ratio 03 — Python Backtest ≥ 1.0 REJECTED — status rejected
Python Win Rate 03 — Python Backtest ≥ 50% REJECTED — status rejected
MT5 Win Rate 04 — MT5 Tester ≥ 55% REJECTED — status rejected
MT5 Maximum Drawdown 04 — MT5 Tester ≤ 20% REJECTED — status rejected
MT5 Profit Factor 04 — MT5 Tester ≥ 1.2 REJECTED — status rejected
Paper Trading Duration 05 — Paper Trading ≥ 30 days Insufficient data — test continues until threshold met
Paper Trading Profit Factor 05 — Paper Trading ≥ 1.2 REJECTED — blocked before live publication

See the survivors

Browse Verified Expert Advisors

Every EA in our shop has cleared all 7 quality gates. Browse by style, pair, or performance metrics.

Risk Disclosure: Passing all quality gates does not guarantee future profitability. Historical backtesting — including every-tick MT5 simulations — relies on past market data that may not reflect future conditions. Spreads, liquidity, and market structure change over time. Paper trading on a demo account does not carry the psychological or execution pressures of live trading. Always test any automated system on a demo account before committing real capital. Read our full Risk Disclosure before making any purchase.