How I Built an AI Crypto Trading Bot on Bybit (And It Actually Works)

I've been trading crypto since 2018. For the first few years I did what everyone does — stared at charts, got emotional, bought tops, sold bottoms. Classic. Around 2022 I started writing scripts to automate the boring parts. By 2024 I was running full bots. And in March 2026 I entered one of my bots — QuantumEdge — into Bybit's official AI vs Human trading competition.

This is the real story of how I built it, what works, and what nearly blew up my account.

Why I Built QuantumEdge

The honest reason: I got tired of losing money to emotions. A bot doesn't panic. It doesn't chase. It doesn't ignore its own stop-loss because "it'll come back." It executes the rules you wrote, every time, without ego.

The Bybit AI vs Human competition was the push I needed to actually finish it. Real money, public leaderboard, competing against other traders. Nothing focuses the mind like stakes.

The Tech Stack

QuantumEdge is built with Node.js and runs as a systemd service on a Linux VPS. The core components:

• Bybit Unified API — for market data, order placement, position management
• 15-minute OHLCV candles — the primary signal timeframe
• ATR (Average True Range) — for dynamic stop-loss and take-profit levels
• Multi-factor signal scoring — combining momentum, volume, and trend indicators
• SQLite — for trade logging, equity snapshots, performance tracking
• Circuit breaker — halts trading after 3 losses in 30 minutes

The Signal Logic

The bot checks each symbol on a 15-minute interval. For a trade to trigger, the signal score must exceed a threshold (currently 10/15). The score combines:

// Simplified signal scoring
function scoreSignal(candles, symbol) {
  let score = 0;
  
  // Trend alignment (EMAs)
  if (ema20 > ema50) score += 3;  // bullish structure
  
  // Momentum (RSI range)
  if (rsi > 45 && rsi < 65) score += 2;  // sweet spot
  
  // Volume confirmation
  if (volume > avgVolume * 1.5) score += 2;
  
  // ATR-based volatility filter
  if (atr / price > 0.008) score += 2;  // enough movement
  
  // MACD crossover
  if (macdCross === 'bullish') score += 3;
  
  return score; // trade if >= 10
}

Direction is determined separately — the bot looks at recent price structure and volume distribution to decide long vs short. A minimum direction margin of 3 points is required before committing.

ATR-Based Stop Loss and Take Profit

This was the biggest improvement over my early bots. Fixed percentage stops are fine, but the market doesn't care about your 2% rule. ATR-based stops adapt to actual volatility:

// ATR stop/take-profit calculation
const atr14 = calculateATR(candles, 14);
const stopLoss = entryPrice - (atr14 * 1.8);   // 1.8x ATR below entry
const takeProfit1 = entryPrice + (atr14 * 1.5); // partial exit
const takeProfit2 = entryPrice + (atr14 * 2.8); // full exit

On a volatile symbol like SOL, the stop might be 3-4% away. On a quiet session of LINK, it might be 1.2%. The bot adjusts automatically — no more getting stopped out by normal noise on a wild day.

The Circuit Breaker

This saved me from a few disasters. If the bot loses 3 trades in any 30-minute window, it halts completely for the rest of the session:

// Circuit breaker check
const recentLosses = trades.filter(t => 
  t.pnl < 0 && 
  t.closeTime > Date.now() - 30 * 60 * 1000
);

if (recentLosses.length >= 3) {
  console.log('Circuit breaker triggered. Halting.');
  process.exit(0); // systemd restarts next session
}

Without this, a bad signal model in choppy conditions can turn 3 small losses into 10. The circuit breaker means the worst case is manageable.

Symbols and Results

The competition runs on Bybit with a minimum 1,000 USDT balance, 10 trades/day minimum, and maximum 15x leverage. QuantumEdge uses 6x leverage by default.

Active symbols: XRP, DOGE, SUI, HYPE, NEAR, LINK, 1000PEPE, XAUT

Removed from the list after poor performance: AVAX, BNB — both had 0% win rate in early testing.

Current competition status: running live, competing in the AI vs Human leaderboard. Dashboard: rdai.in/QuantumEdge/

What I'd Do Differently

• Start with paper trading longer. I rushed to live too early and paid tuition.
• Symbol selection matters more than signal tuning. A great signal on the wrong symbol is worthless.
• Log everything. SQLite trade logs saved me countless hours of debugging.
• The circuit breaker is non-negotiable. Add it on day one, not after a 10-loss streak.

What's Next

I'm working on ML signal enhancement using scikit-learn to score setups based on historical pattern similarity. The goal is to filter out marginal signals that currently drag down the win rate.

The bot is open-source: github.com/rdhacker09/bybit-1v1-bot