When Trump tweets a tariff, when Powell opens his mouth, when a missile hits a tanker in the Strait of Hormuz — futures move in the first 30 seconds. Squawk Box is built to be in the trade before the headline finishes loading on Bloomberg.

A full-stack, end-to-end news trading pipeline running locally on your own machine:

1. Ingests breaking news from Telegram channels, Google News, Yahoo Finance, FinancialJuice, and SEC EDGAR — concurrently, in real time.
2. Filters the noise with a hand-tuned regex matrix that knows what actually moves ES, NQ, BTC, gold, and oil.
3. Reasons about each surviving headline with a locally-hosted DeepSeek R1 14B LLM (via Ollama) — extracting direction, instrument, confidence, urgency, and magnitude as structured JSON.
4. Executes trades automatically through a paper-trading simulator or live Binance Futures, with tight stops, trailing exits, position sizing tiers, and daily loss circuit breakers.
5. Broadcasts everything to a live web dashboard over WebSockets — with a text-to-speech squawk so you literally hear the news as it breaks.

Most "news trading" repos stop at sentiment scoring on Twitter. This one is the whole stack:

• Source-aware filtering — FinancialJuice headlines (pre-curated by pros) get a fast lane; raw RSS runs the strict regex gauntlet.
• Structured LLM signals — no vibes-based sentiment. Every signal is a typed LLMSignal with direction, confidence, urgency, magnitude, and reasoning. R1's <think> chain-of-thought is stripped before parsing.
• Two-tier urgency — critical headlines (Fed, Trump, geopolitics) jump the queue; everything else gets batched.
• Real risk management — per-instrument cooldowns, conflicting-signal suppression, trailing stops, max-duration auto-close, daily loss caps. Not a toy.
• Local-first — your model, your data, your machine. No OpenAI key, no subscription.

# 1. Clone
git clone https://github.com/Kiankk/squawk-box.git
cd squawk-box

# 2. Install
pip install -r requirements.txt

# 3. Pull the model
ollama pull deepseek-r1:14b

# 4. Configure credentials (.env)
cp .env.example .env   # then edit
#   TG_API_ID=...
#   TG_API_HASH=...
#   OLLAMA_URL=http://localhost:11434
#   EXECUTOR=paper            # or "binance"

# 5. Run the full stack
uvicorn server:app --reload
# open http://localhost:8000

# Or run the terminal-only squawk box
python main.py

Everything tunable lives in trading_config.py:

MIN_CONFIDENCE      = 0.75   # LLM confidence floor
HIGH_CONFIDENCE     = 0.85   # bump to medium size
ULTRA_CONFIDENCE    = 0.93   # max size
MAX_OPEN_POSITIONS  = 3
MAX_DAILY_LOSS_USD  = 500
DEFAULT_STOP_LOSS_PCT      = 0.5    # tight scalp stops
DEFAULT_TAKE_PROFIT_RATIO  = 2.0    # 1:2 R:R
ENABLE_TRAILING_STOP       = True

Position sizing scales with model confidence — small, medium, or large.

• A live web dashboard showing every headline as it lands, color-coded by category.
• LLM verdicts streamed in: SHORT DXY · conf 0.85 · IMMEDIATE · TRADEABLE — "China sets yuan midpoint at strongest level since 2023".
• Open positions with live P&L, trailing stops, time-to-auto-close.
• Equity curve, win rate, daily trade count, and market mood — all on one screen.

This is a research project. News trading is brutal, the LLM can be wrong, and live execution against Binance Futures will lose you real money if you misconfigure it. Default mode is paper trading. Leave it there until you know exactly what every knob does. Not financial advice.

• Twitter/X firehose integration
• Multi-model voting (R1 + Llama + Qwen ensemble)
• Backtesting harness against historical headline archives
• Discord/Telegram alert webhooks
• Browser extension for one-click headline injection

Issues and PRs welcome. If you find a category of headline this misses, open an issue with the source and timestamp — the filter matrix gets sharper every iteration.