We Asked an AI to Build 5 Crypto Trading Strategies: The Backtest Results
Can you trust a strategy an AI builds for you? The honest answer is: only after you backtest it, and this article is the proof. The AI trading strategy backtest results below come from CoinQuant's AI strategy builder, which created five strategies from plain-English prompts, exactly as any user would, before every one was tested on real market data with fees included.
The results are published raw: four winners and one loser. That mix is the point. An AI that generated five perfect strategies would be lying to you, and the one losing strategy is the best demonstration in this article of why validation is the product. Every strategy below is a named entry in the CoinQuant Strategy Library, so each prompt, each strategy, and each result can be reproduced.
The Prompts We Used
The five prompts were written the way a normal trader would talk to an AI assistant. No hidden parameters, no expert shorthand, no strategy keywords beyond what a casual trader knows.
| # | The Prompt (verbatim) | Resulting Strategy (library name) |
|---|---|---|
| 1 | "Buy BTC daily when EMA 50 crosses above EMA 200, sell when it crosses below" | AI EMA 50/200 1D Trend Strategy |
| 2 | "Buy BTC daily when price breaks the 20-day high, sell on the 20-day low" | AI Donchian 1D Breakout Strategy |
| 3 | "Buy BTC 4H when MACD crosses above its signal line, sell when it crosses below" | AI MACD 4H Momentum Strategy |
| 4 | "Buy BTC when RSI(14) on 4H crosses above 50, sell when it crosses below 50" | AI RSI 4H Cross Strategy |
| 5 | "Buy BTC 4H when price touches the lower Bollinger band, sell at the middle band" | AI Bollinger 4H Mean Reversion Strategy |
The AI builder validated every prompt into a backtestable schema, including the strategy name, asset, timeframe, and fee assumptions, with no coding and no additional instructions from us.
Test Setup
| Parameter | Setting |
|---|---|
| Asset | BTCUSDT (spot, Binance) |
| Timeframes | Daily for prompts 1-2, 4-hour for prompts 3-5 |
| Windows | Jan 2021 to Aug 2026 (daily), Jan 2023 to Aug 2026 (4-hour) |
| Fees | Binance standard taker, 0.1%, included in every result |
| Initial capital | $10,000 |
| Position | 100%, one position at a time, long only, no leverage |
AI Trading Strategy Backtest Results: Ranked
| Rank | Strategy | Total Return | Trades | Win Rate | Profit Factor | Sharpe | Max DD |
|---|---|---|---|---|---|---|---|
| 1 | AI EMA 50/200 1D Trend Strategy | +216.6% | 3 | 66.7% | 20.11 | 0.74 | 49.5% |
| 2 | AI Donchian 1D Breakout Strategy | +118.4% | 24 | 50.0% | 1.45 | 0.54 | 48.6% |
| 3 | AI MACD 4H Momentum Strategy | +34.3% | 291 | 33.3% | 1.05 | 0.41 | 44.8% |
| 4 | AI RSI 4H Cross Strategy | +28.1% | 401 | 20.9% | 1.04 | 0.37 | 58.5% |
| 5 | AI Bollinger 4H Mean Reversion Strategy | -8.2% | 141 | 66.0% | 0.95 | 0.04 | 38.4% |
The Winner: The AI's Simplest Idea
AI EMA 50/200 1D Trend Strategy returned +216.6% ($10,000 to $31,665), the best result of the five, from a prompt of a single sentence. The profit factor of 20.11 is extreme because the strategy traded only three times in 5.5 years: it bought the 2021-2022 cross, exited, re-entered after the 2023 recovery, and rode the cycle. One of its two winning trades did most of the work.
That result comes with two warnings that belong in the same paragraph. Three trades is not statistical evidence, and a 49.5% drawdown means the strategy sat through the 2022 crash fully invested. The strategy is a case study in how a simple, patient trend rule behaves, not a recommendation to deploy on three trades.
The Workhorse: Donchian Breakout
AI Donchian 1D Breakout Strategy is the most robust result in the set: +118.4% across 24 trades, a 50% win rate, profit factor 1.45, and Sharpe 0.54. The AI translated "break the 20-day high, sell on the 20-day low" into a textbook Donchian breakout, the same family that won the five-family comparison on Bitcoin, and it behaved like it: selective entries, large trend captures, and a 48.6% drawdown during the 2022 bear market.
The Middle: Two Thin Edges
AI MACD 4H Momentum Strategy returned +34.3% but needed 291 trades, and its profit factor of 1.05 is nearly breakeven. AI RSI 4H Cross Strategy returned +28.1% across 401 trades with a 20.9% win rate and a profit factor of 1.04. Both are the same profile: the AI followed the prompt faithfully, the strategy generated a lot of signals, and after fees the edge nearly disappeared.
These two are the most instructive rows in the table. An AI that builds what you ask for will build high-frequency strategies that bleed fees, and the backtest is the only place you can see that before it happens live.
The Loser: Why We Are Publishing It
AI Bollinger 4H Mean Reversion Strategy lost 8.2%. The AI did exactly what the prompt asked: buy the lower band, sell the middle. The strategy won 66% of its trades and still lost money, because the losses were larger than the wins and the 4-hour timeframe generated 141 trades of fees.
We are publishing the loser because it is the most honest result in this article. If every AI-generated strategy in this test had been profitable, the correct conclusion would be that the backtest was rigged. Instead, the losing strategy demonstrates the actual value of the product: the AI generates the idea, and the backtest separates the ideas that work from the ideas that sound good. The trader who deployed the Bollinger prompt without a backtest would be learning the lesson with real money.
What the Test Proves
Three conclusions come out of the five runs:
AI-generated strategies are like human strategies: most are mediocre, a few are good. Two of five were genuinely strong, two were marginal after fees, one lost money
The prompt quality determines the ceiling. The two best results came from the two simplest prompts with trend logic; the two worst came from high-frequency oscillator logic
Validation is non-negotiable. The difference between the +216.6% winner and the -8.2% loser was not the AI's competence, it was the backtest's existence
A fourth conclusion is worth stating for the flagship question behind this article. The full cycle, plain-English prompt to validated schema to completed backtest to metrics report, ran in minutes per strategy, end to end, with no code and no manual parameter entry. That is what an AI agent backtesting workflow looks like when the generation layer and the testing layer are the same product, and it is the workflow the AI agent backtesting tutorial teaches step by step.
This is the demo only a platform with a working AI builder can publish: the full pipeline, prompt to schema to backtest to metrics, run live and shown raw. The AI agent backtesting comparison explains why the agent layer changes the workflow, and the results above show what it produces.
The Practical Lesson
Ask the AI for a strategy, then treat the answer as a hypothesis, not a conclusion. Backtest it with fees included, read the profit factor and the drawdown, and only then decide. The five strategies in this article took minutes to generate and minutes to test, and the full evidence is public in the CoinQuant Strategy Library.
Disclaimer:
Key Takeaway
