Why 38 Out of 40 AI Trading Bots Lost to a Simple 50/200 Moving Average (And How to Actually Use AI in MQL5)
Everyone in the retail and quant space right now wants an AI chatbot to write them a "Holy Grail" Expert Advisor. You type in a prompt, paste the code into MetaEditor, run a quick backtest, and expect magic.
Except a recent empirical benchmark study (FIDES) actually put this to the test. Researchers evaluated 40 different trading strategies generated by leading AI models, and the results are a massive wake-up call for anyone blindly trusting LLM-generated EAs:
📊 The Benchmark Results: AI Bots vs. Simple Baselines
- 38 out of 40 AI-built trading strategies lost to a simple "Buy and Hold" baseline.
- Even more humbling: a basic 50/200-period Simple Moving Average (SMA) crossover—one of the oldest, simplest indicators built into MetaTrader—beat the average Sharpe ratio of every single AI model tested.
- The funniest detail from the paper? In their own self-assessments, 32 of the 40 AI strategies confidently claimed they beat the market—when only 2 actually did.
If you have spent serious time coding in MQL4/MQL5 and running real capital in the markets, this shouldn’t surprise you at all.
Why Chatbots Build Terrible Execution Logic
An LLM is a text predictor, not a market practitioner. When you ask an AI to design a trading strategy from scratch, it doesn't understand market microstructure, sudden volatility expansion on XAUUSD, spread widening during news events, or how real-tick slippage destroys naive scalping logic.
Instead, it gives you a tangled mess of overfitted indicator rules that might look smooth on a careless history test, but immediately fall apart when market regimes shift from trending to compression.
Where AI Actually Gives MQL Developers an Edge
Does this mean AI is useless for algorithmic traders? Not at all. I use it every day in my development workflow—just never to invent entry and exit signals out of thin air.
As an MQL developer, real alpha comes from separating pattern research & engineering from deterministic execution. Here is where AI actually saves hours of grunt work behind the scenes:
- Data Cleaning & Feature Engineering: Writing Python or MQL5 utility scripts to parse massive tick history, clean up corrupted bars, and export structured datasets.
- Regime & Stress Diagnostics: Analyzing how an existing EA's parameters hold up when volatility suddenly compresses or expands across multiple timeframes (e.g., M5 execution vs. H1 trend filters).
- Faster Code Refactoring: Structuring boilerplate MQL5 classes, custom on-chart UI panels, array handling, and debugging compiler warnings in minutes instead of days—freeing up mental energy for actual risk math.
The Takeaway for Traders & EA Builders
When real orders hit the broker server, the market doesn't care how clever your AI prompt was. It only cares about your execution logic, your drawdown control, and whether your system recognizes the active market regime.
That exact philosophy—relying on strict mathematical regime filtering and true volume/divergence confirmation rather than black-box AI guessing—is why I build all my tools on 100% deterministic logic. Whether I'm engineering multi-timeframe regime strategies for XAUUSD or tracking institutional volume flows with OBVCD Pro, the rule never changes: let AI speed up the R&D lab, and let proven math protect the live account.
👉 Explore my published indicators, trading tools, and ongoing quantitative research on my MQL5 profile: mql5.com/en/users/naamo (@naamo)
💬 For those of you building EAs or custom indicators in MQL5 right now: where are you drawing the line between AI assistance and your core trading logic?
📚 References & Further Reading:
- FIDES: A Concordance Protocol for LLM-Generated Trading Strategies (arXiv: 2608.23308): https://arxiv.org/abs/2608.23308
- Summary Breakdown (Only 2 of 40 AI Trading Strategies Beat Buy-and-Hold): https://algotrader.ch/blog/is-ai-trading-legit/
