Is Claude AI any good at creating trading strategies? We backtested it

Contents

1. How did we set up the test?
2. What strategies did Claude come up with?
3. What did the backtests show?
4. Where did Claude go wrong?
5. The parameters came from habit, not from the market
6. One date range can tell you anything
7. A profit can be one good stretch
8. It didn't ask what we wanted
9. Which steps still need a human?
10. What does this mean if you use AlphaPrime?
11. Frequently asked questions

Claude is good at turning a trading idea into exact rules you can backtest, and not good at picking rules that make money on the first try. On 2026-10-28 we asked it for three futures strategies and ran each one through AlphaPrime's hosted backtester. Two of the three lost money after costs. The third made money, for a reason that had little to do with the rule.

That's the short version. The longer one is more useful, because the way it went wrong is the same way most of us go wrong. Asking an AI model for a strategy is a newer way into algo trading, and the tests afterwards stay the same.

How did we set up the test?

We gave Claude one plain brief and let it do the rest. Claude ran in the chat with the AlphaPrime connector added, on a normal Claude plan, with no API key. The brief was: "Give me three rule-based futures strategies I can backtest right now. Be specific enough that there is nothing left to interpret."

The connector's backtest tool is free and uses no credits. It covers 17 futures markets on 5-, 15-, 30- and 60-minute and daily bars, and on the day we ran it the data went from 2018-09-02 to 2026-10-14. Costs are fixed and on the heavy side. For ES the tool charged $33.50 commission per side, which already includes 2.5 extra ticks, plus one tick of slippage. Results come back in US dollars per contract.

We didn't touch the rules before the first run. Whatever Claude wrote is what got tested.

What strategies did Claude come up with?

Three textbook ideas, written as exact conditions. A strategy rule is "exact" when two people reading it would place the same trades. All three passed that bar straight away, which is more than I can say for a lot of strategy ideas posted on forums.

#	Market and bars	Entry (long)	Exit	Stop
1	ES, 60-min	Close above the 200-bar SMA and RSI(14) crosses above 30	RSI(14) crosses above 70	20 points
2	NQ, 15-min	EMA(20) crosses above EMA(50) and ADX(14) above 25	EMA(20) crosses below EMA(50)	40 points
3	GC, daily	Close above the highest high of the previous 20 days	Close below the lowest low of the previous 10 days	None

One detail it got right without being asked. For the gold breakout it compared today's close with the highest high of the 20 days before today, not including today. Leave today in and the close can never be above the high, so the rule never trades. Small thing. It's also exactly the kind of off-by-one error that quietly ruins a hand-written backtest.

What did the backtests show?

Two lost money after costs and one made money. Here are the numbers exactly as the tool returned them.

Put plainly, AlphaPrime's backtester had Claude's ES idea losing $17,878 over 109 trades. Profit factor is gross profit divided by gross loss, so anything under 1.0 lost money overall. Row 1b is a rerun we asked for after seeing the first result; more on it below. The gold breakout is the only winner, and I wouldn't trade it.

Where did Claude go wrong?

Not where people expect. It didn't invent indicators, and it didn't peek at future data. The connector only lets a condition look at the current bar and up to 10 bars back, so there was nowhere to hide a look-ahead mistake even if it had made one.

The mistakes were about judgment.

The parameters came from habit, not from the market

RSI 14 at 30 and 70. A 200-bar average. EMA 20 and 50, ADX above 25. Those are the defaults you'd find in any indicator manual, and Claude picked them because they are the defaults. Nobody tested them on ES 60-minute bars first. Neither did Claude.

The stops were worse. A 20-point stop on ES is $1,000 per contract, and on 60-minute bars the market can travel that far before an RSI exit at 70 ever arrives. Strategy 1 won 22.94% of its trades. A mean-reversion idea that loses three trades out of four is a sign that the stop, not the idea, is deciding most outcomes.

So we changed one number. Same rules, stop at 40 points instead of 20. The win rate went from 22.94% to 40.59%, the loss over the whole period shrank from $17,878 to $8,692, and the profit factor rose from 0.808 to 0.931. Still a losing strategy. But one parameter Claude chose by feel had cut the loss roughly in half, which tells you the first result was mostly about the stop.

One date range can tell you anything

The NQ result is the one I'd show anyone who trusts a single backtest. Its monthly equity sank to -$50,162.50 by June 2022, climbed to +$16,983.50 by June 2024, then slid to -$22,701.50 by September 2026. Run it from January 2023 to mid-2025 only and you'd have a lovely equity curve and a strategy you'd want to trade. Same rules. Different window. Walk-forward analysis in AlphaPrime's strategy validation tests the same rules across successive windows, so a curve that only looks good in one stretch, like January 2023 to mid-2025 here, shows up.

Claude didn't pick that window, to be fair. The tool uses its fixed range. But if you ask an AI to "find a good strategy" and let it choose the dates, this is the trap.

A profit can be one good stretch

The gold breakout made $65,356.80, so it looks like Claude's one success. Look closer. It traded 34 times in eight years and spent 48.7% of the time in the market, which is close to simply holding gold half the time. Its monthly equity was -$30,818.80 in August 2025 and +$104,729 in February 2026. Nearly all of the profit arrived in one gold rally, and the maximum drawdown of $58,169.60 is almost as big as the profit itself.

I think 34 trades is too few to tell a rule apart from the gold price. Claude didn't flag that. It reported the profit factor of 1.441 like any other number.

It didn't ask what we wanted

No question about account size, how much drawdown we could stand, or whether we trade during the day. It went straight to rules. That's what the brief asked for. It's also why I think a strategy an AI writes from memory is a starting point, never a candidate.

Which steps still need a human?

The parts that need a decision about you, not about the data.

• What counts as good. Profit factor, drawdown, number of trades: you set the bar before you look at results, or you'll move it afterwards.
• Whether it holds up. One backtest proves very little. Shift the parameters, try a related market, raise costs, and look at data the rules were not built on. Our guide to robustness testing for trading strategies goes through each check.
• Whether you'd actually trade it. A strategy that loses 20 trades in a row will be switched off by its owner long before the statistics say it's broken.

Claude is useful around all three. It explains what each number in a report means, and it will happily rerun a variant when you ask. It just shouldn't be the one deciding.

What does this mean if you use AlphaPrime?

Asking Claude for strategies one at a time is the slow way. In AlphaPrime the same chat can start a strategy generation run instead: you describe the market and what you're after, and AlphaPrime searches its library of nearly 100,000 trading rules and keeps only the strategies that pass your filters, including out-of-sample tests. That run uses credits, so we didn't start one for this article. Important actions such as starting a run always wait for your OK.

The AlphaPrime with Claude page shows how to connect. It uses your own Claude plan, and the backtest tool we used here costs nothing.

If you try it, start where we did. Take one of Claude's strategies, change only the stop, and run it again. If one number moves the result that much, which of the other defaults would you still trust?

Frequently asked questions

Is Claude AI any good at creating trading strategies?

It is good at writing exact, testable rules from a plain idea. In our test two of its first three strategies lost money after costs, mostly because it chose textbook parameters and stops without testing them, and the third made money mainly in one gold rally. Treat its ideas as a starting point and backtest every one.

Can Claude backtest a strategy by itself?

Not on its own. With the AlphaPrime connector added, Claude can send exact rules to AlphaPrime's backtester and read the results back, for free, on 17 futures markets.

How much does a backtest through the connector cost?

Nothing. The backtest tool is free and uses no credits. Strategy generation runs are the part that uses credits, and they start only after you confirm them.