We tested ChatGPT's trading strategies
Four ChatGPT-labelled Bitcoin strategies lost money in a dated backtest. Review the recorded results, missing settings, and limits of the experiment.
A dated August 6, 2026 repository record says four strategy implementations selected from a ChatGPT response all lost money in one two-year BTCUSDT test. The recorded returns ranged from −7.31% to −44.47%, while the same record puts buy and hold near +15% for the window.
That result needs an evidence limit beside it. The original ChatGPT transcript, model version, generated strategy code, and machine-readable run outputs were not retained. We did not rerun the study for this editorial update. The numbers below survive as a dated publication and commit record, but the experiment is not independently reproducible from the repository today.
The question and the surviving protocol
This is the prompt preserved in the original article:
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.
The wording asks for a level of consistency that historical research alone cannot establish. “Statistically significant in one study,” “profitable in one backtest,” and “consistently profitable in the future” are different claims. A language model can repeat the first two as prose without producing evidence for the third.
| Method field | Surviving record |
|---|---|
| Service and model | Labelled ChatGPT; exact model/version unknown |
| Prompt date and full response | Exact request time unknown; original response unavailable |
| Selected candidates | Four named rule families; full answer and selection rationale unavailable |
| Market and bars | BTCUSDT, 1-hour |
| Test window | August 6, 2024 to August 6, 2026 |
| Venue | Unknown |
| Engine revision | Unknown |
| Position sizing | Unknown |
| Fees, spread, slippage, funding, and fill settings | Described as realistic in the publication; numeric settings unknown |
| Parameter tuning | Publication says none; source needed to verify this was not retained |
| Separate final OOS period | None recorded |
The original article says a fifth idea, futures cash-and-carry, was excluded because the required data were unavailable. Without the full response, we cannot tell whether other candidates were proposed, discarded, or rewritten. That missing selection record matters whenever a small subset stands in for a model's overall answer.
The recorded scoreboard
| Strategy label | PnL | Sharpe | Win rate | Trades | Max drawdown |
|---|---|---|---|---|---|
| MACD with ADX Regime Confirmation | −7.31% | −2.58 | 17.6% | 17 | 40.1% |
| Ensemble Time-Series Momentum | −8.34% | −0.71 | 31.3% | 496 | 13.6% |
| Volatility-Scaled Donchian Breakout | −15.05% | −1.20 | 34.6% | 208 | 18.2% |
| EMA Crossover with ADX and ATR Filters | −44.47% | −1.44 | 34.4% | 587 | 52.8% |
These are simulated historical outputs, not live trades. The CFTC's discussion of hypothetical performance limitations explains why a simulation can omit material market factors and why costs and losses must be presented with performance claims.
What each row contributes
MACD with ADX regime confirmation
The published row records −7.31%, a −2.58 Sharpe ratio, 40.1% maximum drawdown, and 17 trades. The ADX condition appears to have produced a low trade count, but 17 observations leave wide uncertainty. The aggregate record does not let us audit signal dates, position size, or how the large drawdown arose.
Volatility-scaled Donchian breakout
The Donchian row records −15.05% across 208 trades, with a −1.20 Sharpe and 18.2% maximum drawdown. Calling it a breakout rule describes its design; it does not establish that false breakouts caused the loss. That would require a trade-level attribution the surviving record does not contain.
Ensemble time-series momentum
The ensemble row records −8.34% across 496 trades, with a −0.71 Sharpe and 13.6% maximum drawdown. It traded much more often than the MACD row. Without the exact fee and fill settings, the result cannot separate signal loss from execution drag or show which lookback contributed what.
EMA crossover with ADX and ATR filters
The moving-average row records −44.47% across 587 trades, with a −1.44 Sharpe and 52.8% maximum drawdown. It is the lowest-return row in this table. The result describes this implementation in this window; it does not prove that all EMA, ADX, or ATR combinations lack an edge.
What the experiment can answer
The dated record supports one narrow statement: the four listed implementations all lost money in the recorded BTCUSDT window under the unpublished numeric settings. It does not establish that ChatGPT originated each rule, that the code matched the unavailable response exactly, or that AlphaProve's AI translation was correct. It also cannot compare current ChatGPT models, different prompts, other assets, or a separately held-out period.
The varied trade counts still point to a useful research question. A generated rule is only a hypothesis until its code, data, costs, and selection process are inspectable. Familiar indicator names and citations in an answer do not replace a backtest, and one backtest does not replace chronological forward-walk analysis.
Where AI can still help
Turning a precise rule into reviewable code can save implementation time. That benefit is separate from whether the rule has an edge. A safer division of work is to bring a falsifiable hypothesis, ask the model to express it without adding unstated conditions, review the generated strategy, and then test it under a documented protocol.
For a reproducible version of this study, retain the full model name, request time, prompt and response, every candidate and rejection, final code, data source, date boundaries, sizing, and each execution setting. Predefine how a winner will be selected and whether a final untouched period will be used.
AlphaProve's strategy-authoring reference documents the inputs available to generated code, while the getting-started methodology explains the run settings that should accompany a result. Those records are the difference between an interesting dated story and an experiment another researcher can inspect.