DayTrader — Hybrid Paper Trading Algorithm

A paper trading bot that combines technical analysis with AI-powered news sentiment to make buy/sell decisions using fake money and real market data.

Trades only execute when both signals agree: strong technicals and confirming news sentiment. This two-layer filter reduces false signals and keeps the bot from acting on noise.

How It Works

Alpaca Market Data (free IEX feed)
        │
        ▼
┌─────────────────────┐
│  Technical Analysis  │  RSI, SMA crossover, MACD, Bollinger Bands, Volume
│  (strategy.py)       │  → weighted score from -1.0 to +1.0
└────────┬────────────┘
         │ strong signals only (|score| ≥ 0.25)
         ▼
┌─────────────────────┐
│  News Headlines      │  Fetched from Alpaca's Benzinga-sourced News API
│  (news.py)           │  → last 24 hours, up to 6 per stock
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│  Claude Sentiment    │  Claude Haiku classifies headlines as
│  (claude_strategy.py)│  BULLISH / BEARISH / NEUTRAL
└────────┬────────────┘
         │ trade only if technicals + sentiment agree
         ▼
┌─────────────────────┐
│  Portfolio Manager   │  Fractional shares, position limits,
│  (portfolio.py)      │  cash management, CSV logging
└─────────────────────┘

Setup

Prerequisites

• Python 3.10+
• Free Alpaca account (for market data + news)
• Anthropic API key (for Claude sentiment filter)

Install

pip install -r requirements.txt

Configure API Keys

Create a .env file in the project root:

ALPACA_API_KEY=your_alpaca_key
ALPACA_SECRET_KEY=your_alpaca_secret
ANTHROPIC_API_KEY=your_anthropic_key

Usage

# Single scan (--force runs outside market hours)
python engine.py --force

# Live paper trading (scans every 15 min during market hours)
python engine.py --live

# Dry run — full pipeline with real API calls, but no trades committed
python engine.py --force --dry-run

# Backtest over historical data (rule-based only, no sentiment)
python engine.py --backtest
python engine.py --backtest --backtest-days 365

# Check portfolio status anytime
python status.py
python status.py --full

Configuration

All settings live in config.py:

Setting	Default	Description
STARTING_CASH	$100	Paper trading balance
MAX_POSITIONS	5	Max stocks held at once
MAX_POSITION_PCT	25%	Max portfolio allocation per stock
WATCHLIST	25 stocks	S&P 500 blue chips + tech + growth
BAR_TIMEFRAME_MINUTES	15	Intraday bar size (5, 15, 30, 60)
STRONG_BUY_THRESHOLD	0.25	Min score to trigger Claude check
CLAUDE_MONTHLY_BUDGET	$8.00	Hard cap on API spending per month
CLAUDE_MODEL	claude-haiku-4-5	Model used for sentiment
BENCHMARK_TICKER	SPY	Buy-and-hold comparison index

Scoring Weights

Each indicator contributes to a combined score. Weights sum to 1.0:

Weight	Value	Indicator
WEIGHT_RSI	0.25	Relative Strength Index
WEIGHT_MA_CROSSOVER	0.25	SMA short/long crossover
WEIGHT_MACD	0.25	MACD histogram
WEIGHT_BOLLINGER	0.15	Bollinger Band position
WEIGHT_VOLUME	0.10	Volume vs 20-day average

Project Structure

DayTrader/
├── engine.py            # Main entry point — orchestrates the full scan loop
├── config.py            # All tunable parameters in one place
├── strategy.py          # Technical indicators + weighted scoring
├── claude_strategy.py   # Claude sentiment filter (BULLISH/BEARISH/NEUTRAL)
├── news.py              # Alpaca news headline fetcher
├── data_source.py       # Alpaca market data client (bars + quotes)
├── portfolio.py         # Portfolio tracking, fractional shares, trade logging
├── budget.py            # Monthly API spend tracker with hard cap
├── benchmark.py         # SPY buy-and-hold comparison (alpha tracking)
├── decision_log.py      # Per-signal audit log (why each trade was taken/skipped)
├── status.py            # Quick portfolio summary CLI
└── requirements.txt     # Python dependencies

Key Design Decisions

• Two-layer filter: Technicals alone generate too many false signals. Claude reads actual news headlines and only confirms trades with clear material catalysts. No trade executes without both layers agreeing.
• Categorical sentiment, not confidence scores: Claude returns BULLISH/BEARISH/NEUTRAL — no fake confidence percentages. A 3-way vote is harder for the model to fudge and easier to audit.
• Budget-capped AI: Hard monthly spending limit on Claude API calls. When the budget is hit, all sentiment checks return NEUTRAL (safe default — no trades execute).
• Fail-safe defaults: API errors, missing news, budget caps — all result in NEUTRAL sentiment, which blocks trades. The bot never acts on incomplete information.
• SPY benchmark: Every scan logs portfolio return vs SPY buy-and-hold. If alpha is consistently negative, the strategy isn't working.

Important Notes

• This is a learning/research tool, not financial advice.
• Alpaca free tier uses IEX data (slight delay, subset of full market) — fine for paper trading.
• Backtest mode uses yfinance and is rule-based only (no sentiment) since historical news archives aren't available.
• The decision log (logs/decision_log_YYYY-MM.csv) records every strong signal evaluation for post-hoc analysis.