Two days ago, I built my first trading bot — an EMA crossover strategy on MetaTrader 5. It was basic, but it worked. So I did what any engineer would do: I tried to make it better.

48 hours and 9 bots later, I learned the hardest lesson in algorithmic trading: every "improvement" you add is a new way to lose money.

The Setup

• Platform: MetaTrader 5 on Mac (Wine)
• Symbol: XAUUSD (Gold) on M1 charts
• Account: Demo with ~$1,035 starting balance
• Pair programmer: Claude Code for rapid MQL5 iteration
• Goal: Build a profitable scalping bot with consistent 2% daily returns

Phase 1: EMA Crossover (Bots 1-2) — The False Start

My first bot used EMA 14/28 crossover with a scoring system — trend detection, strength, pullback, momentum, breakout, and crossover signals. It had market mode detection (TREND vs RANGE) and dollar-based TP/SL.

The problems showed up immediately:

1. Spread killed everything. MetaQuotes demo has 30-40 point spreads on XAUUSD. My max spread filter was set to 25. No trades fired until I raised it to 50.
2. Break-even triggered too early at $0.80 — five trades exited at exactly $0.00, wasting perfectly good entries.
3. TP was too small ($1) vs SL ($3) — I needed a 74% win rate just to break even.
4. Overtrading. 35 trades in 3 minutes with a 3-second cooldown.

Results:

• Smart Pro v3.0: 35 trades, 57% win rate, -$7.50
• Smart Pro v3.2: 6 trades, 33% win rate, -$6.00

Lesson learned: EMA crossover strategies lag too much for M1 scalping. By the time the signal confirms, the move is already over. Trend-following does not work on 1-minute gold charts.

Phase 2: Mean Reversion (Bots 3-4) — Finding the Edge

I scrapped trend-following entirely and switched to mean reversion: buy dips, sell rips, close on small profit.

The logic was simple: gold on M1 charts tends to bounce. If price drops X points from a recent high, buy. If it rips Y points from a recent low, sell. Close everything at $1 profit.

Dip Buyer v1.0 — BUY Only

• Buy when price dips 60 points from recent high
• Average down once (max 2 positions)
• Close all at $1 profit or $10 loss

Result: 6 rounds, 6 wins, 0 losses, +$6.45

First time I saw green. The strategy worked.

Dip Buyer v2.0 — Added SELL (Broke It)

Added sell-the-rip to complement buy-the-dip. But the triggers competed — only the strongest signal fired. In an uptrend, the rip was always bigger than the dip, so the bot only sold.

Result: 12 rounds, 10W/2L, -$4.20

Fix: Made triggers independent. Dip and rip fire separately.

Dip Buyer v2.1 — Independent Triggers

Both BUY and SELL working simultaneously. 22 buys + 9 sells.

Result: 34 rounds, 31W/3L (91%), +$3.30

Getting somewhere. But all 3 losses came from averaged positions — rounds where the bot held 2 positions and the total loss was 2x.

Dip Buyer v3.0 — The Winner

Removed averaging entirely. One position at a time. Tightened parameters.

Result: 82 trades, 71W/11L (87%), +$34.99

This was the overnight run. I went to sleep with the bot running and woke up +$35. At 87% win rate with $1 TP / $5 SL, the math works: each loss costs 5 wins, and you win 87 out of 100 trades.

Phase 3: The Over-Engineering Spiral (Bots 5-9)

Here is where I made the classic engineering mistake. Dip Buyer C was making money, so I tried to make it "smarter."

Bot 5: Smart Reversal — The Sniper

Four entry filters: EMA 50 slope + dip size + RSI oversold/overbought + price below fast EMA. Plus a 3-minute timeout exit.

Result: 3 trades, 3 wins, +$5.00

100% win rate — but 3 trades in hours. It was too selective. Fine as insurance alongside Dip Buyer C, but it could not carry a session alone.

Bot 6: Precision Scalper — Adding Safety

Took Dip Buyer C and added "improvements":

• Tighter SL ($3 instead of $5)
• Waterfall detection (skip falling knife entries)
• Post-loss cooldown (60 seconds)
• Consecutive loss pause (3 losses = 5-minute break)
• 120-second time exit

Result: -$10

The waterfall filter was too aggressive — it blocked good entries along with bad ones. The tighter SL ($3) meant I now needed 75% win rate instead of the 83% I needed before. Marginal improvements in safety cost more in missed opportunity.

Bot 7: APEX Scalper — Maximum Complexity

This was the peak of over-engineering:

• Declining take-profit: Started at $1.15, dropped to $0.10 over time
• Confidence-based lot sizing: 0.02 lots on high-confidence, 0.01 on low
• Peak profit lock: If unrealized profit hit $0.50, lock it in
• EMA entry filter
• RSI entry filter
• Quick-cut logic: Exit at -$0.50 on low-confidence trades

Result: 68% win rate, -$20

Every "smart" feature backfired:

• Peak profit lock cut winning trades at $0.50 while losses stayed at $3. It was designed to protect gains but instead destroyed them.
• Quick-cut was supposed to limit losses to $0.50. Actual average loss: $1.35 due to execution delay + spread.
• EMA filter blocked good trades in range-bound markets where mean reversion works best.

Bot 8: Phoenix Scalper — The Worst

EMA-based entry, 10-second quick check, 3-position smart averaging.

Result: 24 trades, 38% win rate, -$30

The worst bot of all nine. Averaging amplified losses. EMA entry missed the best dip-buying opportunities. Everything that could go wrong did.

Bot 9: Back to Basics

Went back to Dip Buyer C. Added only two things: market-hours check and the 120-second timeout.

The Numbers Do Not Lie

The simplest bot with one filter — "did price dip enough?" — crushed every complex version.

~450 lines of code beat ~700 lines. One human observation ("gold bounces from dips") beat RSI, EMA crossover, and every technical indicator combined.

What I Learned from 200+ Live Trades

1. Simplicity Beats Complexity

Every "smart" feature was a new failure mode. Declining TP cut winners. Peak lock killed profits. EMA filters blocked good entries. Quick-cut failed due to execution delay.

The market does not reward complexity. It rewards discipline.

2. Test Live, Not on Paper

Real spreads (30-40 points), real slippage, and real execution delays kill strategies that look perfect in theory. Demo accounts with live market data reveal truth instantly. If your bot cannot handle real-world conditions, the strategy is broken.

3. Human Intuition Beats Indicators

"Gold always bounces from dips" — this one observation from watching the chart outperformed RSI, EMA crossover, MACD, and every technical filter I tried. The best algorithm was codifying a simple human observation, not stacking indicators.

4. No Averaging

All of my biggest losses came from averaged positions. When you hold 2-3 positions and the market keeps going against you, you lose 2-3x. Single position with a hard stop loss outperformed every DCA variant I tested.

5. R:R Ratio Is Everything

At 87% win rate, $1 TP / $5 SL gives +$48 per 100 trades. Change the SL to $3 and you need 75% — sounds easier but the tighter stop gets hit more often in volatile conditions, dropping actual win rate.

6. Kill Your Darlings Fast

I retired 7 bots in 48 hours. No ego, just data. Every bot had rich logging — entry price, exit price, hold time, spread, balance, consecutive losses. When the data says your clever feature hurts performance, delete it.

7. AI Accelerates, You Decide

Claude Code built all 9 bots — from idea to deployed EA in minutes. The iteration speed was incredible. But the winning strategy came from watching the chart, not from code. AI is the best pair programmer I have ever had, but it cannot feel the market.

Technical Notes

The Stack

• MetaTrader 5 on Mac via Wine
• MQL5 for all bot logic
• Claude Code as AI pair programmer for rapid iteration
• Bot files symlinked from project directory to MT5 Experts folder — edit, hit F7 to compile
• Background monitoring script parses MT5 Expert logs (UTF-16LE) and updates a live scoreboard

Key Parameters (Winning Bot)

Strategy:       Mean Reversion (Buy Dips / Sell Rips)
Take Profit:    $1.00
Stop Loss:      $5.00
Dip Trigger:    150 points from recent high
Rip Trigger:    150 points from recent low
Max Positions:  1 (no averaging)
Cooldown:       15 seconds
Time Exit:      120 seconds

Account Journey

The account went from $1,035 to a peak of $1,065, then back to $1,030 as I tested progressively worse bots. The irony: I would have been up $35 if I had just left the simple bot running.

What Is Next

Going back to what works. Dip Buyer C with minimal additions:

• Market-hours check (skip low-liquidity periods)
• 120-second timeout (if it does not bounce in 2 minutes, exit)
• Rich analytics logging for continuous improvement
• Target: 2% daily ($21) with 85%+ accuracy

Building in public. More updates coming.

This is part 2 of my algorithmic trading journey. Read part 1 for how I went from zero MQL5 knowledge to a working bot in one day.

I am Manjodh Singh Saran, a Senior Software Engineer building products across web, mobile, and algorithmic trading. Follow my journey on LinkedIn.