Trading Expert
AI-powered trading advisor for semiconductor and tech stocks. Uses DeepSeek API to analyze news sentiment and supply chain cascades, sending buy/sell/hold alerts to your Android phone via Telegram.
Architecture
News → Prefilter (VADER) → DeepSeek API → Signal Scorer → Telegram Alert
│ │
└── Supply Chain Graph ←──────┘
(cascade effects)
- News ingestion: RSS feeds (unlimited, free) + NewsAPI + GNews
- Fast filter: VADER sentiment + urgency keyword detection
- Deep reasoning: DeepSeek API with function calling for structured output
- Cascade engine: Directed graph of 27 semiconductor companies — traces how events propagate through the supply chain
- Signal scoring: Weighted combination of AI sentiment (50%), news volume (20%), source credibility (15%), price confirmation (15%)
- Delivery: Rich Telegram messages with emoji, cascade effects, risk factors, and action recommendations
Quick Start
Prerequisites
Setup
# 1. Clone the repo
git clone https://github.com/Sinnup/trading-expert.git
cd trading-expert
# 2. Configure environment
cp .env.example .env
# Edit .env with your API keys:
# DEEPSEEK_API_KEY=sk-...
# TELEGRAM_BOT_TOKEN=...
# TELEGRAM_CHAT_ID=...
# NEWSAPI_KEY=...
# GNEWS_API_KEY=...
# 3. Install dependencies
uv sync
# 4. Start the system
docker compose up -d
# 5. Verify
uv run trading-agent --help
CLI Commands
uv run trading-agent analyze --ticker NVDA # Analyze one ticker
uv run trading-agent watch # Start intraday monitoring
uv run trading-agent daily # Run daily summary
uv run trading-agent signals --ticker NVDA # Recent signals
uv run trading-agent portfolio # Paper trading P&L
uv run trading-agent backtest # Signal accuracy stats
Telegram Commands
Once the bot is running, message it on Telegram:
/status — Current signal summary/signals NVDA — Last 5 NVIDIA signals/portfolio — Paper trading P&L/mute 4h — Snooze alerts/unmute — Resume alerts/threshold 0.8 — Set minimum alert score/help — Show all commands
Tracked Companies (27 tickers)
US Semiconductor Designers
NVDA, AMD, INTC, AVGO, QCOM, MRVL, MU, TXN, ADI, ON, MCHP
Equipment Makers
ASML, AMAT, LRCX, KLAC
Foundries
TSM, SMIC, 688981.SS (Hua Hong)
Memory (Asia)
005930.KS (Samsung), 000660.KS (SK Hynix)
Hyperscalers / Big Tech
AAPL, MSFT, GOOGL, AMZN, META
Mobile / IP / Auto
2454.TW (MediaTek), ARM, IFNNY (Infineon), STM
Project Structure
trading-expert/
├── pyproject.toml # uv project config
├── Dockerfile # Multi-stage Docker build
├── docker-compose.yml # redis, worker, beat, app
├── .env.example # API key template
├── config/
│ ├── settings.yaml # Global settings
│ ├── companies.yaml # Ticker universe + supply chain
│ └── thresholds.yaml # Alert thresholds + keywords
├── trading_expert/
│ ├── cli.py # Click CLI
│ ├── fetchers/ # News + price data ingestion
│ ├── analysis/ # Prefilter, DeepSeek, signals, cascade
│ ├── models/ # SQLAlchemy ORM (5 tables)
│ ├── notifications/ # Telegram bot + formatter
│ ├── portfolio/ # Paper trading + backtesting
│ └── tasks/ # Celery async tasks
├── tests/ # Unit tests (43 passing)
├── IP plans/ # Architecture docs
├── .claude/agents/ # Agent role definitions
└── FEATURE_MANAGER.md # Backlog + changelog
Database Schema
5 tables in SQLite (auto-migrates to PostgreSQL):
| Table | Purpose |
|---|
articles | Every fetched news article with sentiment scores |
signals | Buy/sell/hold signals per ticker |
signal_outcomes | Accuracy tracking at 7/30/90 days |
paper_trades | Virtual buy/sell execution records |
portfolio_snapshots | Daily portfolio value snapshots |
Considerations
Data Quality
- NewsAPI free tier returns only 100 articles/day and has 15-minute delay. Not suitable for sub-minute trading.
- RSS feeds are free and unlimited but limited to sources that provide them (Reuters, CNBC, Seeking Alpha).
- VADER sentiment is fast but not perfect on financial text. It serves as a pre-filter; DeepSeek is the real analyzer.
- yfinance data may have delays of up to 15 minutes during market hours.
API Costs
- DeepSeek: ~$0.14 per 1M input tokens, ~$0.28 per 1M output. With ~20 escalated articles/day, cost is approximately $0.05-0.15/day.
- Telegram Bot API: Free, unlimited.
- NewsAPI: Free tier (100 req/day). Paid starts at $449/month.
- GNews: Free tier (100 req/day). Paid starts at $5/month.
- Total estimated cost: $2-5/month (free tier APIs + DeepSeek usage).
Limitations
- Not financial advice: This system provides analysis and signals based on news sentiment. It is not a substitute for professional financial advice.
- No real trading: The system does not connect to brokerage APIs. Trades must be executed manually via GBM or your broker.
- Paper trading only: The portfolio tracker simulates trades. Real execution has slippage, fees, and liquidity constraints.
- Signal latency: News may already be priced in by the time the system processes it, especially with RSS delays.
- Model risk: DeepSeek and VADER can misinterpret news. Always read the reasoning before acting.
- Concentration risk: All 27 tracked companies are in tech/semiconductors. If the sector has a systemic shock, all signals may be correlated.
- No backtest history yet: Signal accuracy can only be measured after 7+ days of live operation.
Batch Processes
| Process | Frequency | What it does |
|---|
| Intraday fetch | Every 15 min (market hours) | Fetch news → filter → analyze → alert |
| Daily summary | 4:30 PM ET (market close) | Aggregate day's signals, send Telegram report |
| Outcome check | Daily at 6:00 AM | Check past signals for accuracy (7/30/90d) |
| Portfolio snapshot | Manual or via daily task | Save portfolio value for performance tracking |
Deployment
Local (macOS)
docker compose up -d
Cloud (any Linux host)
# 1. Clone repo + copy .env
# 2. Update DATABASE_URL in .env to PostgreSQL if using cloud DB
# 3. docker compose up -d
# Identical operation — no code changes needed.
Without Docker
# Start Redis manually, then:
celery -A trading_expert.tasks.celery_app worker --loglevel=info &
celery -A trading_expert.tasks.celery_app beat --loglevel=info &
Development
uv sync # Install all deps
uv run pytest # Run all tests (43 tests)
uv run trading-agent --help # CLI help
Agent Roles
See .claude/agents/ for role-specific instructions used during development:
architect.md — System architecture decisionsdeveloper.md — Coding standards and patternsqa.md — Testing strategyversioning-expert.md — Git workflow and commit conventionsdevops.md — Docker, deployment, environmentsecurity-reviewer.md — Secrets, dependencies, code security
Versioning
Follows Conventional Commits:
feat(analysis): implement DeepSeek signal analysis
fix(fetchers): handle RSS timeout gracefully
docs(readme): add deployment instructions
