DEV Community

Ray
Ray

Posted on

I ran a self-hosted AI trading strategy lab for 30 days. Here is what happened.

I've been running TradeSight — a self-hosted Python app that runs AI strategy tournaments overnight — for about 30 days of paper trading. Here's the honest report.

What TradeSight does

You give it a set of indicators (RSI, MACD, Bollinger Bands, 15+ total). It backtests each one, runs them in head-to-head tournament rounds, picks a winner, then paper trades that strategy live via Alpaca API. Zero capital at risk — Alpaca's paper trading account is free.

GitHub: github.com/rmbell09-lang/tradesight

Two tournaments, two champions

Session 1: RSI Mean Reversion won. Score: 0.6238 over 4 rounds. Params: oversold=30, overbought=75, position_size=0.8, SL=5%, TP=6%. Seeded with real Alpaca historical data.

Session 2: MACD Crossover dethroned RSI. Score: 0.72 over 4 rounds. Consistent head-to-head wins.

MACD Crossover became the live paper trading strategy.

What actually happened in live trading

Paper portfolio: started at $10,000. Realized PnL: -$54.55 (-31%). Open positions at time of writing: QQQ and GOOG long.

Yeah. The tournament winner got smoked in live paper trading.

The lessons

Stop-loss implementation matters more than strategy selection. Early on I had a Phase 1 bug where proper SL wasn't wiring correctly. By the time I noticed, some positions had drifted well past acceptable loss. A strategy that wins in controlled tournament rounds will fall apart fast without hard SL enforcement.

Tournament winner ≠ live trading winner. MACD beat RSI in head-to-head backtests. But backtests are clean — no slippage, no regime changes mid-run. RSI Mean Reversion might have been the better live strategy. I don't actually know yet.

Confidence score tuning is iterative. The tournament assigns scores, but calibrating what threshold to trust requires more data than 2 sessions. I'm on session 3 now.

The thing that actually works

The tournament format. Running 15+ strategies against each other in elimination rounds is genuinely useful for narrowing the field. Even if the absolute winner isn't perfect, you're eliminating obvious losers fast.

RSI Mean Reversion surviving 4 rounds with real Alpaca seed data and specific params tells you something real about that parameter set in that market regime. That's more signal than a single backtest.

What's next

Session 3 is running. I added circuit breaker logic (kill switch if daily loss exceeds threshold) which was a Phase 1 gap. SL enforcement is fixed. Curious if tournament results stabilize once the bug surface shrinks.

Want to run your own? Star it on GitHub and try the paper trading mode — Alpaca account is free.

github.com/rmbell09-lang/tradesight

Top comments (0)