AI-Assisted EUR/USD Trading System

The goal, in the words of the project's program README:

"Goal: find and validate a repeatable, net-positive process, judged only by forward paper results after costs, starting from a $10 paper account. No orders of any kind (no live, no demo). Read-only cTrader demo API. No profit is promised."

Source: program/README.md

The first version was a single EUR/USD rule, described in the main README:

"'Large 5m range': on each completed M1 candle, if range_5m (high-low of the last 5 candles) > 0.000326, go SHORT at the next candle's open. 0.01 lot, 5-pip stop, 120-minute maximum hold, one position at a time, $10 virtual account."

Source: README.md

A second, simpler simulator (newbot-paper) tested an EMA(20)/EMA(50) crossover on 1-hour EUR/USD bars with a 2×ATR stop and 4×ATR target, also from a $10 fake balance.

Source: newbot-paper/README.md

Architecture

• ctrader/market.py: builds 1-minute candles from live spot quotes on the cTrader Open API demo host (read-only requests only).
• trading/live_paper.py: the first live paper trader. It simulated trades in memory and logged them to CSV.
• program/engine.py: the long-running paper engine for the 21-day program: quotes, fill simulation, risk limits, plug-in strategy modules, review cycles.
• program/hypotheses/<ID>.py: plug-in rule modules, each pre-registered in an append-only journal before its backtest.
• program/risk.py: the hard limits in code, checked against RISK_LIMITS.md by a SHA-256 hash. The engine refuses to start if the file changes.
• program/stats.py, daily_summary.py, final_report.py: forward statistics and the promotion test.
• ops/: supervisor scripts (single instance lock, crash restart with backoff, heartbeat watchdog).
• program/backtests/: pre-registration, backtests and screening scripts.
• newbot-paper/: a standalone EUR/USD paper simulator using only the Python standard library and Yahoo Finance hourly data.

The deterministic risk rules (from RISK_LIMITS.md)

• Paper only: no live or demo orders. The engine refuses to start if any program file references an order, amend or close-position API message.
• Account and sizing: paper equity starts at $10.00. Size is the symbol's minimum volume (0.01 lot FX) while equity < $20.
• Per-trade risk: every trade has a hard stop set at entry that is never moved further away. Max risk per trade $1.10 while equity ≤ $11.00, then min($1.10, 10% of equity). No entry if the spread is > 3× normal. Used margin ≤ 70% of equity.
• Portfolio: at most 2 open positions; total open risk ≤ 15% of equity; flat by Friday 20:45 UTC.
• Daily loss limit: no new entries for the day if equity falls to ≤ 80% of the start-of-day equity.
• Kill switch: halt everything if equity < $4.00 or drawdown from peak > 50%. The program never resumes after a halt.
• Per-hypothesis kill: after ≥ 15 forward trades with profit factor < 0.8, or 8 losses in a row.
• Costs always applied: fills at live bid/ask, slippage, $3.00 per lot per side commission on FX, swap; if a stop and a target are hit by the same quote, the stop wins.
• Promotion rule: ≥ 30 closed trades and ≥ 14 days, net P&L > 0 after costs, profit factor ≥ 1.3, beats a random-direction control at the 95th percentile, t-stat ≥ 2.0, max drawdown < 35%. Even then, no orders; a human decides.

Source: program/RISK_LIMITS.md (full text in the source viewer below)

1. Frozen backtest of "Large 5m range" (EUR/USD)

Columns as in the file. Cost is in percent; returns are in percent.

Source: data/reference/nonoverlap_frozen_results.csv (rounded to 3 decimals)

2. Live paper trading of "Large 5m range" (2026-10-05 to 2026-10-06 UTC)

Commission was not modelled in this trader ("commission (~0.6 pip round trip on this Pepperstone raw demo) is not modelled"). The trader was then stopped (README: "the old short-only paper trader is STOPPED").

Source: data/live_paper_trades.csv (totals computed from the file), README.md

3. newbot-paper sanity backtest (EMA crossover, EUR/USD 1h, $10 start)

Its forward paper account: 1 closed trade (stop loss, -$2.95), balance $7.05 from $10.00, with one position still open at the last run on 2026-10-08.

Source: newbot-paper/backtest_results.txt, newbot-paper/data/trades.csv, newbot-paper/data/run.log

4. 21-day paper program: forward trades

Strategy LBO v1 (GBPUSD London opening-range breakout) was the only hypothesis that reached forward testing. 16 other hypotheses were rejected by the pre-forward screen.

Source: program/journal/trades.csv, program/reports/daily_2026-10-14.txt, program/journal/hypotheses.jsonl

5. LBO v1 history backtest (2022-01-03 to 2026-10-06, real bid/ask ticks)

Source: program/backtests/lbo_history_100plus.md (fixed 0.01 lot; equity may go negative on paper)

6. The strategy search

After the first rules failed, the AI-assisted research moved to a continuous loop of testing new strategy ideas, each written down before its backtest was run.

• Early rounds (to 2026-10-06): EUR/USD, other FX pairs and symbols, and a fixed bracket (10-pip stop / 20-pip target, about -$1 / +$2 per trade at 0.01 lot). "None of the 7 candidates is good enough to deploy." The bracket: "no variant has an edge." Other symbols: "44 pre-declared symbol-variant tests. None reaches significance." Including news-time strategies, the early rounds reached 80 tests: "Even so, no strategy becomes significant."
• Later rounds (2026-10-18 to 2026-10-19): the search ledger records 167 more research rounds (Rounds 7 to 171), 316 variants in all, covering FX, stock indices, gold and other commodities, crypto, rates and calendar effects.
• Method: historical backtests with in-sample / out-of-sample splits, after costs, against a random-direction control. The pass bar was a two-sided Bonferroni t-stat over the cumulative number of variants tested (at least 2.5); it rose to |t| ≥ 3.74 after Round 142.
• Rule: paper trade only strategies that first pass a historical backtest (the LBO history test used 100+ trades).
• Result: every round in the ledger is marked FAIL or NOT CONFIRMED. None passed every test.

The closest candidates (all failed)

Sources: research/README.md, research/SEARCH_LEDGER.md

• Costs erased the backtest edge. "Large 5m range" had a profit factor of 1.27 with zero costs, but it fell to 0.92 at 0.01% cost per trade.
• It lost in live paper trading. 22 of 29 trades hit the 5-pip stop, and the $10 account fell to $5.84 in two days, without commission even being charged.
• The best candidate failed on untouched data. The LBO backtest verdict: "LBO v1 does NOT survive the backtest." Out-of-sample: "Win rate 29.6% vs a 36.4% break-even, PF 0.73, net -$86.35, t = -3.04." The program was then stopped: "2026-10-18: program STOPPED at B's request after LBO v1 failed the 2022-2026 history backtest."
• $10 is too small to manage risk. From the LBO report: "In both periods the account freezes after its first 3-6 trades (from $10 to ~$6.4-6.6), so it could not even trade its way back." From newbot-paper: "The 1% risk target is not achievable at $10 … True 1% risk needs roughly $200+ of equity."
• Nothing else passed. Every later research round in the search ledger is marked FAIL or NOT CONFIRMED.
• The decision. On 2026-10-20 the project was stopped because no strategy showed a robust out-of-sample edge. No real-money orders were ever placed.

Sources: data/reference/nonoverlap_frozen_results.csv, data/live_paper_trades.csv, program/backtests/lbo_history_100plus.md, program/README.md, newbot-paper/README.md, research/SEARCH_LEDGER.md

eurusd-trading-system.zip: a sanitized copy with two folders, newbot-paper/ and eurusd-trading-system/. Credentials and account numbers were removed; see PUBLIC_COPY.md.

Easiest: newbot-paper (Python 3, no packages, no account)

unzip eurusd-trading-system.zip
cd newbot-paper
python3 backtest.py 6mo        # backtest on Yahoo Finance EURUSD=X hourly data
python3 run.py                 # start a fake $10 paper account in ./data

Its README uses absolute paths from the original machine; run the commands from inside the folder instead.

Full system: eurusd-trading-system (needs a cTrader demo account)

cd eurusd-trading-system
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
cp .env.example .env           # fill in your own cTrader Open API DEMO credentials
.venv/bin/python ops/demo_connection_test.py   # read-only demo check
python3 program/risk.py --selftest            # check the risk limits

Some scripts expect data files and paths from the original machine (/workspace/...) that are not included, so not everything will run as-is.

Paper trading only. Educational demo, not financial advice.