AI Backtesting: Which Tools Really Test Your Strategy

An AI can backtest your strategy when the product behind the chat box owns historical price data and an engine to run your rules against it. A general assistant with neither runs nothing at all, and the vendors who sell AI backtesting say so as plainly as anyone.

ChartLabs has no AI backtesting. The platform is manual by design.

Can an AI backtest my trading strategy?

Yes, when "AI" is the front end of a testing engine. No, when it is a chat window with nothing behind it.

TradeZella's AI backtesting page states the first half: "AI backtesting and automated backtesting are the same thing. In TradeZella, the automated backtesting engine is AI-powered end to end." The AI reads your sentence and turns it into rules, an ordinary engine runs those rules over historical data, and the AI reads the results back to you.

TradeTest.ai heads a whole section "ChatGPT can't backtest your strategy. We can."

Both are vendors with something to sell, and they agree on the mechanism: the testing is done by an engine over data, not by a language model.

Which AI tools actually run a backtest, and which only talk about one?

Every one of these tools has its own engine except ChatGPT, where the model writes code and something else runs it. Every line comes from the tool's own page.

Tool	What its own page says it does	Markets it names
TradeZella	You describe the strategy in plain English, the engine runs every matching trade, and Zella AI reads the full trade log	futures, stocks, crypto, forex
TradeTest.ai	Describe a strategy in natural language, get charts, metrics and robustness checks	crypto, forex, stocks, gold and silver
TrendSpider	Sidekick builds, tests and refines strategies for you, and a tested strategy can be deployed as a bot	US equities, options, US futures, ETFs, crypto, forex
CoinQuant	An AI builds the structured logic, then runs it on Kaiko data with fees included	crypto, Bitcoin back to 2017
TradrLab	Describe an idea and get a ready-to-test strategy, plus data dashboards	forex, crypto, indices, commodities, metals
Capitalise.ai	"Easily create, test & automate Trading scenarios using everyday english"	none named
ChatGPT	Writes the backtest code you run somewhere else	forex and US stocks, in the guides' examples

Two of them also run strategies live, which is a different job from testing one. Capitalise.ai, whose site says it has been acquired by Kraken, is mostly about automated execution, and TrendSpider turns a tested strategy into a bot. CoinQuant draws the line between the two itself: "An AI trading bot executes orders live. AI agent backtesting validates whether a strategy has an edge before anything trades it."

If you are reading a ranking, check what it measures. Liberated Stock Trader puts TrendSpider first with a "Backtesting Suitability Score of 4.85", and those scores are the author's own lab tests across scanning, backtesting and an "AI & Algo Index" rather than backtesting alone.

Can I use ChatGPT to backtest a trading strategy?

The two camps answer differently, and the difference is who supplies the data.

TradeZella's FAQ answers "No." Its reason: "A general AI assistant like ChatGPT or Claude has no historical price data for your instrument, no engine to execute trades against that data, and no record of your actual trading ... It will produce numbers that sound plausible and are invented."

DailyForex answers the same question with a qualified yes: "It can backtest a strategy in some circumstances, though the trader must upload data for the backtest." The two are closer than they look. Both say the model has no price data of its own. They differ on whether handing it your own data counts, and DailyForex goes on to recommend the other route anyway: "I recommend that instead, traders use ChatGPT to generate code for a backtest to run on platforms such as TradingView or MetaTrader."

The code route has its own failure mode, and the pages that teach it say so. DailyForex: "AI is far from 100% accurate, and there is a reasonable possibility of errors in the code, which can lead to incorrect conclusions", and "Do not blindly trust the results without verification." QuantInsti shows the Python that ChatGPT produced for a mean reversion test, then says of it: "It's not a fully functional or optimised backtesting script. Backtesting requires expertise in financial data analysis and programming." TradeSearcher's guide routes the same way: generate the idea in the chat, then code it into a backtesting platform.

What does "describe your strategy in plain English" actually test?

It tests the engine's reading of your sentence, so look at the trades before the summary.

TradeZella's own fourth step: "Run it and scan the trades first. Before reading any summary stat, check 10 to 15 individual trades." It suggests the questions too: "Do the entries look real? Any trades in pre-market or overnight you would never actually take?"

What does not survive the translation is judgement. Rules like waiting for a level to hold, or taking an entry only if structure agrees, are the ones traders struggle to write down, and a test that quietly drops them is testing a different strategy from yours. That is the argument in our own manual vs automated backtesting article, and it applies to the plain-English route exactly as it applies to code.

What can an AI backtest not test?

Your execution. The engine follows the rules perfectly every time, which is the one thing you will not do.

Our own article puts it plainly: "Automated does not survive contact with you." TradeZella's comparison table agrees from the other side of the argument: on the row headed "Best for", its manual column reads "Discretionary strategies and execution practice".

Costs are the second gap, and it is a fixable one. CoinQuant: "Backtests that ignore trading costs overstate returns." That matches what we see in backtesting accuracy: thin-edge strategies are the ones whose edge disappears once commission, spread and slippage are counted.

Data is the third. One of these vendors names the same source we use: TradrLab says "We use raw exchange-level feeds from Dukascopy to ensure our reports reflect the real market, in real time." We source from Dukascopy too, tick-level where it is available, but we clean it to remove anomalies before building 1-minute candles, because an untouched feed carries spikes nobody could have traded.

Does running 10 or 15 AI variations find an edge or a curve fit?

The run cannot tell you. What settles it is testing the winner on data you did not pick it on.

The speed is real and the vendors are open about it. TradeZella: "In one afternoon you can run 10 to 15 versions and find the configuration with the highest expectancy and lowest drawdown." TrendSpider: "Test hundreds of strategies in the time it would take to test a few by hand", where Sidekick "tests multiple ideas in parallel, working in the background", and "When an idea looks promising, Sidekick refines it and tests it again." The same page offers to "Bring the power of hindsight into your corner by optimizing your strategies with 50 years of historical data."

Speed does nothing about curve fitting. The answer is to retest the winner on data it has not seen, and several of these sources say so in their own terms.

For Traders, a prop firm reviewing twelve tools, puts it directly: "Backtests are trivially easy to overfit, and vendor performance claims are rarely out-of-sample." It refused vendor numbers on the same grounds, saying "Every backtest curve a vendor shows you is cherry-picked by definition", and crediting only results "we could reproduce ourselves on out-of-sample data". Its rule for an AI-written strategy: "Any AI-generated strategy needs walk-forward testing on unseen data before it touches simulated capital".

DailyForex gives the practical version: "Avoid over-optimization. If a backtest is successful, rerun it on a different period to ensure it has similar results", and more bluntly, "Essentially, never rely on a single backtest!"

TradeTest.ai answers it inside the product: "Every qualifying backtest is re-tested on data it never saw, split across time periods, stressed with higher fees, and resampled thousands of times". It does not say what qualifies. TradeSearcher takes the lighter route, listing "Overfitting Alerts" and suggesting you ask the chat for simpler rules with fewer conditions.

TradeTest.ai's results board is headed "The best-performing strategies tested on TradeTest.ai right now" and described as "Anonymous, unfiltered". One of its features, "Always Reliable Results", reads: "If something goes wrong, AI automatically fixes it and retries. You only see successful results." Another feature on the same page says "We don't cherry-pick winning strategies." Both sentences are theirs.

Where ChartLabs fits

ChartLabs is manual, so the engine question does not arise. The curve-fitting question does, in its hand-testing form: run a strategy over the same stretch of chart until it looks good and you have fitted yourself to that stretch.

Sessions come in two types for that reason, training and verification, and the analytics split by session type so you can read them separately. The changelog describes that split as a way "to see how you really do on data you haven't traded before".

How many trades should an AI backtest include?

Two numbers circulate, and they count different things.

TradeZella answers for the backtest itself: "Aim for at least 50 trades as a minimum, with 100 or more giving stronger statistical confidence." For Traders answers for the step after it, forward testing on a demo account, where it asks for "30+ trades minimum" before any capital is involved.

TradeZella says "generating a 100-trade sample takes 40 to 60 hours" by hand, and its table gives the same row as minutes for the engine. The engine buys you the sample, and the hours buy you the screen time.

Where does manual, bar-by-bar testing still belong?

In the two jobs the AI route leaves open: checking that the engine traded what you meant, and practising the trade yourself.

The first is recommended by the AI pages themselves. DailyForex, on validating ChatGPT's code: "I recommend comparing the trades in the backtest with manual spot checks, i.e., manually test sample periods of the strategy and compare the results to the backtest results, and ensure the two sets of results are identical." For Traders ends its own workflow the same way: "Manual sanity-check every rule against price action you understand."

The second is conceded by the vendor with the most to lose by conceding it. TradeZella: "Once a strategy passes, switch to manual replay to practice executing it under pressure."

That second half is the part ChartLabs does. You step through historical candles one at a time, place the trade by hand with its stop and target, and every trade is logged with the chart state you had when you took it. Commission and slippage are set per symbol, so the P&L carries what the trade actually cost. Paper trading in live conditions is the step after that.

A backtest tells you how a strategy behaved on data. Markets stay uncertain, and no test changes that.

One plan at $20 a month or $200 a year, with a 7-day free trial and no credit card required to start. Start the free trial.