Every trading pitch has an AI angle now: signal generators, bots that trade while you sleep. But how much of this technology may improve trading decisions, and how much is marketing? Research is starting to separate the real benefits from the hype.
Do Funds Using AI Really Earn More?
Some of them, by a small margin. A 2026 paper in the Review of Financial Studies compared hedge funds that adopted generative AI against those that did not. The adopters earned 2 to 4 percentage points more per year in abnormal returns, meaning returns above what risk exposure alone would explain.
The same test on non-hedge investment firms showed nothing, so the AI was not the cause. The winners already had data teams and AI specialists and added a tool to a working process.
Why Do Impressive AI Results Keep Collapsing?
Two failure modes repeat. The first is contaminated data. A 2017 study reported a random-forest strategy earning 1.54% alpha per month, alpha being return above the benchmark. A 2025 reassessment found the model had been fed information that was not public yet on the dates it supposedly traded. Strip out that look-ahead bias and the 1.54% vanished. Plain linear regression then made more money.
The second is cost. A 2026 study by Esakia and Goltz tested machine-learning strategies ignoring trading costs. Once turnover and price impact were counted, net results turned sharply negative. Price impact is your own order nudging the price against you.
A signal worth a few basis points disappears inside a wider spread. This is why paper testing sits between backtest and live money, and demo accounts replicating live pricing show trading costs first.
Why Is Execution the Safer Place to Use AI?
Nasdaq’s Dynamic M-ELO order type weighs over 140 parameters and resets its holding period every 30 seconds. Nasdaq’s own testing reported fill rates up 20.3% and mark-outs down 11.4%.
Fill rate is how often your order gets executed. Mark-out is how far the price moves against you after it does. Both are measurable within seconds.
That is the whole difference. Execution gives you a narrow target and instant feedback. Price prediction gives you months of waiting before you learn you were wrong.
One caveat: these figures come from the exchange itself rather than an independent reviewer.
What Changes If You Trade Your Own Money?
Downloading the same model does not hand you the fund’s edge. Their timestamped data and validation process stay with them.
There is a ladder here, and each step up costs more when it goes wrong:
- explaining a concept or a filing
- checking your research code
- testing a strategy you defined
- ranking which trades look best
- placing orders with real money
The last two rungs carry most of the risk.
Automation is old news. Expert advisors have run rule-based logic on MT4 and MT5 for years, and both are standard at global brokers like JustMarkets. The broker applies extended swap-free by default across its account types, so an algorithm holding positions overnight has one less financing cost to model. Negative balance protection sets a floor on the account, which becomes relevant before any automation touches live money.
Risk Warning: Trading Forex and CFDs involves significant risk and can result in the loss of your invested capital.
