QuantumEdge vs Human Traders: AI Trading Competition Results

Bybit's AI vs Human competition puts automated bots against human traders on the same leaderboard, with real money. You need 1,000 USDT minimum, at least 10 trades per day, and maximum 15x leverage.

I entered QuantumEdge in March 2026. Here's what the data looks like after three weeks.

Competition Rules

• Minimum balance: 1,000 USDT
• Minimum trades: 10/day
• Maximum leverage: 15x (QuantumEdge uses 6x default)
• Competition period: 30 days
• Leaderboard: public, ranked by PnL%

Symbol Performance Breakdown

AVAX and BNB were cut after the first week — 25% and 33% win rates respectively. The signal model just doesn't work well on those symbols. Removing them immediately was the right call.

The Circuit Breaker in Action

The circuit breaker (halt after 3 losses in 30 minutes) triggered 4 times during the competition. All 4 times, it prevented what would have been 7-10 loss streaks:

// All 4 circuit breaker triggers and what followed:
// March 8,  11:30 UTC — choppy BTC price action, saved ~$80
// March 12, 08:15 UTC — ETH news spike, saved ~$120  
// March 15, 14:45 UTC — low liquidity session, saved ~$60
// March 19, 09:00 UTC — post-FOMC volatility, saved ~$150

Estimated savings from circuit breaker: ~$410 in losses avoided

ATR Stops vs Fixed Stops

Before ATR-based stops I used fixed 2% stops. The improvement after switching:

• Stop-out rate: 34% → 19% (less getting stopped on normal noise)
• Average winner size: +$4.10 → +$6.80 (bigger wins because stops give more room)
• Average loser size: -$5.20 → -$4.10 (tighter stops on low-vol sessions)

The tradeoff: ATR stops mean position sizing is more complex since risk per trade varies. I standardise by targeting $50 risk per trade and adjusting size accordingly.

What the Bot Gets Wrong

Honest assessment after watching it trade live for 3 weeks:

• Trending markets: The bot performs best in ranging/oscillating markets. In strong trends it often takes the wrong side after a pullback.
• News events: No macro awareness. Fed meeting, CPI prints, major exchange news — the bot doesn't know they're happening and gets caught on the wrong side.
• Session transitions: Asian → European → US handoff creates fake breakouts that the signal model occasionally chases.

What's Next: ML Signal Enhancement

I'm training a scikit-learn model on ~18 months of historical trade data to score setups by similarity to past winners. The goal is to filter out the bottom 25% of signals — the ones with marginal scores that drag down overall win rate.

Initial backtests suggest this could push win rate from ~65% to ~72% by volume, which would be a meaningful improvement in competition terms.

Live dashboard: rdai.in/QuantumEdge/