The autonomous TopstepX futures bot where selection beats prediction.
NQ · ES · RTY · YM · GC — 3-minute bars — fork it, fine-tune it, make it better.

---

🚀 What is AIGO Tradebot?

AIGO Tradebot is a fully autonomous futures trading system for the TopstepX evaluation platform. It ingests live 3-minute bars for five index/metals futures, detects signals across 10+ strategy families, scores every signal with a trained XGBoost model, and pushes only the highest-quality subset through a selection funnel — because after 5+ years of measured research, the funnel is the edge, not the predictor.

flowchart TD
    A[📡 Live 3-min bars<br/>NQ · ES · RTY · YM · GC] --> B[🧠 Strategy engine<br/>10+ families: BOS, CISD/OTE, EMA, ORB, Keltner, Supertrend...]

    B --> C[🎯 XGBoost probability<br/>proba 0-1 per signal]
    C --> D[🔻 Selection funnel<br/>floor 0.40 · ceil 0.65 · chop 2.0]
    D --> E[🛡️ LLM veto sidecar<br/>7B model · advisory]
    E --> F{Edge monitor<br/>+ consistency halt}
    F -->|✅ healthy regime| G[⚖️ Risk sizing<br/>vol-gated · ATR]
    F -->|🛑 regime broken| K[🔇 Halt - no trade<br/>protect the account]
    G --> H[📤 Broker execution<br/>TopstepX API · brackets]
    H --> I[📒 Attribution agent<br/>+ edge monitor replay]
    I --> J[🧪 Pre-registered specs<br/>specs/ - kill what fails]
    J --> C

    style A fill:#0f1d33,stroke:#22d3ee,color:#e2e8f0
    style C fill:#0f1d33,stroke:#4ade80,color:#e2e8f0
    style D fill:#0f1d33,stroke:#a3e635,color:#e2e8f0
    style H fill:#0f1d33,stroke:#22d3ee,color:#e2e8f0
    style K fill:#2a0f12,stroke:#f87171,color:#fecaca

Loading

🏆 Why it's built this way

The research record in specs/ is the real product — every idea gets a pre-registered spec with a success bar and a kill rule, a blind out-of-sample run, and a written verdict. GO. KILL. INCONCLUSIVE. No post-hoc tuning. No survivorship stories. Headline findings, verified on 5+ years of the author's own data:

Finding	Number
Direction-prediction ceiling (every model family tried)	46–53% — a dead end
Best live signal correlation ever measured	r ≈ 0.27
Raw signal engine win rate (unselected — loses money)	~29%
Funnel-selected subset win rate	~47% at +0.58R, PF 2.11
Volatility/regime correlation	r ≈ +0.39 — the one real signal
09:30–12:00 ET entry window vs all-day (out-of-sample)	+0.455R / PF 1.93 vs +0.027R / PF 1.03
Worst month's damage cut by edge-monitor halt rules	94%

The bot does not chase 70% win rates. It targets a positive-expectancy funnel and tells the truth when an idea dies. That honesty is why it's worth forking.

🔧 Quick start

git clone https://github.com/Sundar1k/AIGOtradebot.git
cd AIGOtradebot
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env        # add your TopstepX credentials
python bot.py               # paper-first: sim_broker / paper_trade.py

Trained artifacts ship in models/ and ppo_exit/policies/. Market data is not distributed — regenerate with python backfill_bars.py using your own account, or start research with keyless_feed.py (free, delayed, no-signup feed). Full config knobs: see .env.example and config.py.

🧬 Make it YOURS — fine-tune & progress it

This project is an open invitation. The entire point of publishing the code AND the research is that you can take it further than one person can:

• Fine-tune the models — the XGBoost/Chronos models in models/ and the PPO exit policies in ppo_exit/ are trainable artifacts. finetune/ training pipelines and retrain scripts show the recipe (weights too large to ship; train your own).
• Add a strategy — implement in strategies/, validate through the selection_validator/ harness, prove it beats the funnel baseline.
• Run a pre-registered experiment — copy the spec protocol from specs/: write the success bar BEFORE the test, run it blind, publish the verdict. Kills are celebrated, not hidden.
• Open a PR — bug fixes, new gates, better docs, whatever. PRs are welcome. If your change lifts the funnel's out-of-sample expectancy, it deserves to be in.

Contribution protocol (the short version)

1. Fork the repo.
2. State your hypothesis and your kill rule up front (see specs/ for the format — steal it).
3. Test point-in-time, out-of-sample. No look-ahead, no post-hoc tuning.
4. Report the numbers either way. A clean KILL is a good outcome.

⚖️ License & commission

• Author: Sundar1k. This evolved project is licensed under the AIGO Tradebot Public License v1.0 (LICENSE): attribution required, plus a 5% commission on Net Trading Profits if you use this software to generate trading income — quarterly, in BTC to bc1q3xqpc603l80vwcn9dr9d8g7rdevl42mldwvlpn.
• Plain-language terms: COMMISSION.md.
• A handful of unmodified files originate from the MIT-licensed algoTraderBot by John Cruz and remain MIT — see NOTICE.md.
• Trading futures involves substantial risk of loss. Past performance, including everything in specs/, is not a guarantee of future results. Nothing here is financial advice.

---

_{AIGO Tradebot · selection over prediction · made with 🔥 and honest kill rules}