forexbot

A trading research lab and a set of paper-trading bots. It began as an MT5 forex bot; most of the work, and everything now running, is on crypto perpetuals (Bitget and Binance data), with forex, gold and NASDAQ tested and mostly killed. Built to be inspectable, not a black box: most of the repo is the record of what was tested and why it died.

No real money is in anything, and nothing here is advice or a forecast. A backtest is what a rule did on history; the evidence is the forward paper record, and real money is gated by a decision rule written before that record existed.

Where it stands

docs/STATUS.md is the one page for the current state: what is running where, on which code, the paper result next to its backtest and drawdown, the real-money gate, and the dates that matter. It carries the date it was last verified and a one-command re-check (bash deploy/verify_status.sh).

Then: CLAUDE.md (the goal, the validation rules, the traps that have produced wrong answers, the hard safety gates, and a dated session log), docs/22-decision-rule.md (when real money may go in), and docs/README.md (the full index of the research notes).

Principles

• Paper and demo until proven. ALLOW_REAL = False in every bot that could place a real order; only a human edits it.
• No credentials in the repo. Exchange keys come from the environment only; MT5 attaches to an already-logged-in terminal.
• Honest backtests. Fees on every trade (12bp round-trip), a held-out last 40% never tuned on, point-in-time universes that include dead coins, funding charged, hindsight shrunk, and a prediction recorded before each test.
• Kills are permanent unless the measurement itself was broken (docs/02-what-failed.md, docs/21-graveyard-audit.md).

Layout

CLAUDE.md            orientation: goal, validation rules, traps, safety gates, dated log
docs/                STATUS.md (current state), README.md (index), research notes 00-26
backtest/            every experiment; blend.py holds the deployed constants
news/                news / sentiment / positioning engine and the crowd-squeeze paper book
combo_bot.py         the combination bot (dry / Bitget-demo; live is gated), with final_books.py, wick_live.py
blend_paper.py       the seven trend paper books;  combo_paper.py  machine v2;  wick_paper.py  crash bids
pair_live.py         the daily + capitulation pair and the capit2 shadow book
dashboard/           the live dashboard (web app);  dashboard_push.py  feeds it from the VM
bot/                 the original MT5 layer: broker, risk manager, execution, strategies
deploy/              systemd units, deploy scripts, verify_status.sh
tests/               offline tests on fake exchanges (run each as a script: python tests/test_combo_bot.py)
logs/                state, trade records, and the saved output of every experiment
strategy_analysis/   validation protocol (the authority on testing), the trader study, cached data (gitignored)

Checking on it

bash deploy/verify_status.sh     # read-only: VM services, code identity, live equity vs the day-45 line
python health.py                 # PC-side: is anything wedged or dead

MT5 smoke test (the original forex layer, optional)

1. Install MetaTrader 5, log into a demo account, enable Algo Trading.
2. pip install -r requirements.txt
3. Copy config/config.example.yaml to config/config.yaml and fill in the login/server.
4. python check_connection.py prints account info and candles.