AlgoStack: Teaching AI to Read Crypto Charts

AlgoStack started as a weekend experiment: what if you gave Claude a bunch of candlestick data and asked it to identify trading patterns? Turns out, it's surprisingly good at it — but getting from "surprisingly good" to "actually profitable" is where the real work begins.

The concept

Traditional trading bots use hard-coded rules: if RSI drops below 30, buy. If MACD crosses above signal line, buy. These work until they don't. Markets evolve, and static rules get eaten alive.

AlgoStack takes a different approach. It feeds Claude real-time market data — candlestick patterns, volume profiles, order book depth, and social sentiment — and asks for a structured analysis. The AI doesn't make trades directly. It produces a conviction score and reasoning, and the execution engine decides whether to act.

The data pipeline

Every 15 seconds, AlgoStack pulls data from multiple sources:

• Price data from Binance WebSocket streams (BTC, ETH, SOL, and 20 altcoins)
• Volume profiles aggregated over 1h, 4h, and 1d timeframes
• Order book depth — bid/ask walls and imbalances
• Social sentiment scraped from crypto Twitter and Reddit

All of this gets formatted into a structured prompt that Claude can parse. The prompt engineering here was the hardest part — too much data and the model gets confused, too little and it misses important signals.

What Claude sees

Each analysis request includes a 50-candle window (about 12 hours on 15-minute charts), the current volume profile, and any significant order book imbalances. Claude responds with:

• Pattern identification (head and shoulders, double bottom, etc.)
• Support and resistance levels
• Trend strength assessment
• A conviction score from -100 (strong sell) to +100 (strong buy)
• Written reasoning for the assessment

Early results

After two weeks of paper trading (no real money), AlgoStack is hitting about 58% accuracy on directional calls with a 1.3 profit factor. That's not amazing, but it's profitable — and it's only going to get better as I tune the prompts and add more data sources.

The biggest wins come from Claude identifying divergences between price action and volume — situations where the price is moving up but volume is declining, which often precedes a reversal.

What's next

I'm building a backtesting framework so I can run AlgoStack against historical data. The goal is to validate the strategy across different market conditions — bull, bear, and sideways — before putting real capital at risk. This is a SaaS play eventually, but I need to prove the edge first.