AI Crypto Portfolio Manager
A 33-day live experiment: an autonomous AI agent managing a $500 paper trading portfolio, making buy/hold/sell decisions every 15 minutes using real Binance market data and Claude Haiku.
SCREENSHOTS
Research Question
Can an AI agent autonomously manage a paper trading portfolio to a +10% return target within 24 hours using only real-time market data and no human intervention?
Video Walkthrough
Key Results
3,188 AI decisions
1,579 paper trades
756 closed positions
22.5% win rate
-0.44% avg return per trade
-4.32% max drawdown (circuit breaker held)
+10.02% best season (Jaguar)
3/34 seasons beat BTC benchmark
Architecture
Case Study
The Experiment
Over 33 days (March 8 – April 10, 2026), an autonomous AI agent ran 38 consecutive 24-hour trading experiments — each starting with $500, with a single mission: grow the portfolio to $525 (+5%) before time ran out.
Every decision was logged. Every trade was recorded. Every reasoning step was stored. The system made 3,188 AI decisions and executed 1,579 paper trades across 20+ coins with zero human intervention.
What We Built
4 microservices running continuously on Railway:
- Market Data — Binance WebSocket streaming live prices for 20+ coins
- Strategy — Claude Haiku evaluating RSI/MACD signals every 15 minutes
- Executor — Risk rules, circuit breaker, paper trade execution
- Dashboard — Live SPA showing every decision and trade in real time
A circuit breaker halted trading if any season dropped 8% — protecting the experiment from catastrophic loss.
What the Data Shows
The experiment ran 38 seasons. 34 completed. Here's the honest result:
Win rate per trade: 22.5% (170 wins / 756 closed positions)
Average return per trade: -0.44%
Seasons beating BTC buy-and-hold: 3 out of 34 (8.8%)
Max drawdown: -4.32% — circuit breaker never triggered
Best season: Jaguar +10.02% (13 minutes, single momentum bet)
The AI consistently picked altcoins (SOL, FIL, APT, OP) over BTC/ETH. In trending markets this is the losing side of the risk curve. The three worst-performing coins — OPUSDT (3% win rate), FILUSDT (11%), APTUSDT (12%) — account for the majority of losses.
The 4 seasons that hit +10% all used the same pattern: a single concentrated momentum bet that caught a pump in 13 minutes. This is not a repeatable strategy.
What Worked
The risk management held up. Over 33 days of continuous autonomous operation, the worst loss was -4.32%. The circuit breaker design and position sizing rules prevented any catastrophic outcome.
The infrastructure was solid. 4 microservices ran continuously with no human intervention for over a month on Railway.
What Didn't Work
The AI has a structural altcoin bias. Restricting the watchlist to top-10 coins only would likely produce materially better results. The confidence scoring also anchored to round numbers (40% of decisions at exactly 60% confidence) — the signal can't be trusted as a trade quality filter.
The Dataset
The real value of this experiment is the data:
- 3,188 AI decisions with full reasoning text (1.4MB)
- 1,579 executed paper trades with entry/exit prices
- 3,252 hourly portfolio snapshots
- 38 named seasons with full performance records
This dataset documents how Claude Haiku reasons about live crypto signals — what it gets right, what biases it has, and exactly what to fix.
Key Findings & Learnings
Confidence scores anchor to round numbers — not a reliable trade filter
Circuit breaker design works — 33 days, max loss contained at -4.32%
Target-hit wins were momentum catches, not repeatable strategy
Buy-and-hold BTC outperformed the AI in 91% of seasons
Restricting watchlist to top-10 coins is the most actionable fix