We tested ChatGPT's trading strategies

Every day, people ask ChatGPT for trading strategies, and it always delivers: confident rules, a nod to the academic literature, even a disclaimer about when each one fails. What almost nobody does is take those answers literally and run them against real market data with honest fills and fees.

So we did, and the results turned into an opinion about where AI belongs in a trading workflow.

The setup

Here is the prompt we gave ChatGPT, word for word:

give me the best algorithmic trading strategies for bitcoin ohlc, that are statistically proven to provide consistent profits over both a short and long period of time.

Read it back slowly. It asks for strategies that are statistically proven, consistently profitable, and robust across both short and long horizons. ChatGPT answered without hesitation.

We took four answers that were concrete enough to test and implemented each one exactly as described, without tuning anything or quietly improving the rules on the way in. AlphaProve's AI chat translated each description into engine code in a couple of minutes, and every strategy ran through the same backtest: BTCUSDT on the 1-hour timeframe, August 6, 2024 to August 6, 2026, two full years, with realistic fills, fees, and slippage charged on every trade.

One more idea from the list, a futures cash-and-carry, never made it to the start line because it needs a futures data feed the engine doesn't have. ChatGPT will happily propose strategies your infrastructure cannot execute, which is its own small lesson about taking implementation advice from something that doesn't know your stack.

One benchmark before the scoreboard: over that same two-year window, Bitcoin itself went from about $56,000 to about $64,600. Simply buying and holding returned roughly +15%.

Here's how the "statistically proven" strategies did:

Four strategies, four losers, in a market that gained 15% on its own. The best Sharpe ratio of the group was -0.71.

MACD with ADX regime confirmation

ChatGPT's pitch: a MACD-plus-ADX system beats a naked MACD crossover because the ADX condition suppresses signals in weak trends. Best for medium-term directional movement.

The result: -7.31% over two years, Sharpe -2.58, and a win rate of 17.6% across just 17 trades, with a 40.1% max drawdown along the way.

The ADX filter did suppress signals. Seventeen trades in two years is close to dormant, and 17 trades is still not a sample you can trust. But look at what the filter let through: three winners out of seventeen, and at the worst point the strategy sat 40% below its peak. A system that risks losing 40% of the account on the way to a -7% finish isn't selective. It's slow and wrong at the same time.

Volatility-scaled Donchian breakout

ChatGPT backed this one hardest: Bitcoin trading-rule research, it said, has found especially strong results for trading-range breakout rules, and broader crypto research supports trend-following. Best for large directional trends; the named weakness, repeated false breakouts during sideways markets.

The result: -15.05%, Sharpe -1.20, a 34.6% win rate over 208 trades, max drawdown 18.2%.

Two years is long enough to contain the large directional trends this strategy says it needs. Bitcoin ended the window 15% higher. The Donchian system still lost 15%, a 30-point gap to doing nothing, because the false-breakout tax it warned about compounded across every sideways stretch and consumed more than the real moves paid. The volatility scaling kept the drawdown moderate, which is worth something. The pick with the strongest claimed research backing underperforming buy and hold by 30 points is the story here.

Ensemble time-series momentum

The pitch here was that time-series momentum is repeatedly documented in crypto returns, with effectiveness that "varies by regime and implementation". Best for persistent medium- and long-term trends; the stated weakness, momentum reversals after sharp trend changes.

The result: -8.34%, Sharpe -0.71, from 496 trades at a 31.3% win rate with a 13.6% drawdown.

This is what a slow bleed looks like at scale: a trade roughly every day and a half, for two years, never a blowup, just fees and small losses paid with machine-like consistency. An ensemble of three momentum lookbacks is three flavors of the same bet, and across regimes that included real trends, the bet still didn't pay. The regime caveat ChatGPT attached was accurate. It just isn't a caveat you can trade around when the model has no way to tell you which regime you're in.

EMA crossover with ADX and ATR filters

ChatGPT's pitch: moving-average strategies have significant support in high-frequency Bitcoin research, but results depend heavily on parameter selection and walk-forward design. Best for established trends; weakness, lagging entries and whipsaws.

The result was the worst of the four on almost every axis: -44.47%, Sharpe -1.44, a 34.4% win rate, a 52.8% max drawdown, and 587 trades, the most of any strategy here.

Five hundred eighty-seven trades of lagging entries into moves that were already ending. At the trough, more than half the account was gone. ChatGPT had warned that this strategy is correlated with the MACD system, "so running both may add little diversification", and the two-year data agrees in the grimmest way: both trend-filter cousins lost, and this one lost six times as much. The model also said "results depend heavily on parameter selection", which is a polite way of saying the idea has no edge of its own, only parameters that sometimes get lucky.

The pattern, and the argument

The prompt asked for strategies statistically proven to deliver consistent profits over short and long periods. Two years of honest data returned four losses between -7% and -44%, in a market that rose 15%. Every failure mode that showed up was one ChatGPT had described itself, unprompted, in the same breath as the recommendation.

That is how these models work. MACD crossovers, EMA systems, Donchian breakouts, and momentum rules are among the most thoroughly documented trading ideas on the internet. That makes them core training data, which is why a language model can describe them, code them, and recite their weaknesses fluently. It also makes them the most crowded ideas in existence. A model that generates the statistically likeliest answer will hand you the strategies whose edges have been arbitraged thinnest, along with an accurate eulogy for each one.

So we'd split the job in two.

Idea generation is the wrong half to delegate. An edge, if you have one, comes from something the crowd hasn't priced: a market quirk you noticed yourself, or an anomaly in data you collected. A "give me a strategy" prompt can never hand you an idea free of the crowd's prior, because the crowd's prior is what the model is made of.

Implementation is the right half. Watching plain-English descriptions become correct, testable engine code in minutes was genuinely impressive, and it's the part of this experiment that would have taken a hobbyist weeks five years ago. Turning a rule into code, wiring the indicators, running an honest simulation, reading the tearsheet: that is technical labor with a right answer, and AI is excellent at labor with a right answer.

The workflow this experiment supports with numbers rather than vibes: bring your own hypothesis, from a source that doesn't share the model's bias, and use the AI to build and test it faster than you could alone.

Honest caveats

This was one asset and one two-year window, run deliberately without parameter sweeps or forward-walk validation, because the point was to test what ChatGPT handed us exactly as handed. The MACD system's 17 trades remain too few to judge on their own. None of this proves these rule families can never work when tuned and validated properly. What it shows is that the ideas as delivered, by the most popular idea machine in the world, lost money for two straight years while the implementation worked flawlessly. That asymmetry is the takeaway.