AI Trading: What’s Real and What’s Hype
AI trading is the same label on at least four different products. The mechanism behind it can be genuine machine learning trained on broad cross-sectional data, sentiment scoring on news feeds, ordinary linear regression with a rebrand, or research-productivity tooling that does not drive any allocation decision.
Those four do not produce the same kind of edge. They do not deserve the same kind of fee. And telling them apart is the work this page is built for.
The U.S. SEC fined two firms a combined $400,000 in March 2024 for marketing AI capabilities they could not back up. The first regulated cases of AI-washing are now public record. Anyone evaluating an AI trading manager in 2026 has a duty to know which side of that line a firm sits on.
This page is for that judgment.
What does AI trading actually mean?
Ask ten managers what AI trading means and you get six different answers. The label sits on top of at least four distinct uses of machine learning inside a trading process. They behave differently in live markets, and they deserve different fees.
Those four are not the same business. A model that improves order placement has a very different risk profile from a model that drives portfolio exposure. One helps implementation. The other shapes the return stream itself.
The most common version of the question reaches us as a tool question: what ChatGPT can and cannot do in a trading process. That evaluation now has its own page: ChatGPT trading: myths, reality, uncomfortable truths.
Regulators have started to put the line in writing. The EU AI Act entered into force on 1 August 2024, with phased application running through August 2027. FINMA Guidance 08/2024 on artificial intelligence in financial services was published on 18 December 2024, with named expectations on model risk, data quality, explainability, and accountability.
Switzerland and the EU now both treat AI in finance as something a firm has to be able to describe, not just claim.
Why so much money is going into AI trading
The promise is real. Modern markets emit more usable information per second than human-only research can absorb consistently: price moves, order-book behavior, cross-asset correlation, corporate language, macro releases, sentiment streams. AI methods earn their keep when the signal is faint, noisy, and spread across many variables that no analyst can hold in mind at once.
Three conditions have to hold for that promise to pay. The data has to be relevant and clean enough to support the model, the model has to produce edge net of trading costs, and the firm has to be able to describe what the model does, when it retrains, and what stops it. Without all three, AI is expensive trial and error.
When the conditions do hold, the academic record is genuinely supportive. Gu, Kelly and Xiu showed in the Review of Financial Studies (2020) that machine learning improves cross-sectional U.S. equity return prediction, with gains concentrated in non-linear methods such as neural networks and tree ensembles. Israel, Kelly and Moskowitz framed the harder version of the same question in the Journal of Investment Management (2020): ML helps in finance, but the gains are bounded by low signal-to-noise, regime change, and capacity limits.
That second paper is the one serious managers cite when they are being honest. The first one is the one marketing decks reach for when they are not.
- Israel, Kelly and Moskowitz (2020) argue that machine learning does help in finance, but the gains are constrained by low signal-to-noise ratios, non-stationary markets, and capacity limits that no model can engineer away.
- Gu, Kelly and Xiu (2020) found that non-linear ML methods improve return prediction in large cross-sectional U.S. equity datasets, with neural networks and tree ensembles outperforming linear baselines.
- The FCA’s 2019 research note on machine learning in UK financial services documented real ML use across trade pricing and execution, with mature deployment depending on supervision as much as on model quality.
- López de Prado’s Advances in Financial Machine Learning (2018) is the working text for serious teams. It barely appears in retail-facing AI-trading content.
When AI trading works, and when it doesn’t
Most disappointment with AI trading does not come from a dramatic model collapse. It comes from the slow loss of edge between research and live trading. Live evaluation has to start at that gap, not at the model’s research benchmarks.
Begin by separating the claim from the mechanism. The marketing language for AI trading and the disclosed methodology often do not describe the same thing. A claim of “deep learning alpha” can sit on top of a logistic classifier with a dozen technical features, and a claim of a “self-learning algorithm” can sit on top of a manually retrained linear model with new parameters every quarter.
The pattern is consistent enough that it has a name. The CFA Institute calls it AI-washing, and published a 2025 framework for investors to test for it.
| Marketing claim | Mechanism that often sits underneath |
|---|---|
| “AI-powered trading” | Linear or logistic model with rolling-window retraining on technical features. |
| “Deep learning alpha” | Two- or three-layer feed-forward network on a small feature set, trained without modern out-of-sample discipline. |
| “Self-learning algorithm” | Manually retrained statistical model with parameter updates approved by a human at fixed intervals. |
| “Sentiment AI” | Off-the-shelf NLP scoring of news headlines, fed into a conventional ranking model. |
| “Adaptive AI strategy” | Regime-switching rules with thresholds tuned in-sample and lightly cross-validated. |
| “Neural network strategy” | Tree ensemble or gradient boosting with a neural-network label on the marketing deck. |
Once the mechanism is clear, the live-trading risks are easier to read. Three of them do most of the damage.
Turnover quietly kills good ideas. A model can identify short-term mispricings with attractive predictive accuracy on paper. In live trading, high turnover forces the strategy to cross the spread repeatedly, pay commissions, and move the market in less liquid names. The signal survives. The returns do not.
Capacity is usually overstated. A strategy that works on $20 million of capital can decay as assets grow into the hundreds of millions. Crowding, liquidity limits, and slower fills turn an elegant small-book result into an average institutional product. AI does not exempt a strategy from the physics of capacity.
Regime dependence hides in attractive backtests. Many ML models are quietly trained on one dominant market structure: falling inflation, central-bank liquidity support, stable factor leadership, narrow volatility ranges. Once the regime shifts, signal decay can show up faster than retraining can catch.
WorldQuant’s 2025 quant contest drew 80,000 participants, roughly double the prior year. AI is lowering the cost of generating ideas faster than it is creating differentiated edge for any one of them.
“Most of what is sold as AI trading is not AI — it is ordinary quant with a wrapper, and the wrapper exists because ‘AI’ raises fees.”
What good AI trading managers do differently
The firms worth taking seriously usually sound less dramatic than the firms working hardest to sell AI. That is the first signal. They describe AI as one component of a wider research process, not as the process itself.
Strong teams can explain why a model exists, what information it uses, how often it retrains, what constraints surround it, and how human review works at each step. They know where the model is fragile. They do not pretend otherwise.
Bridgewater publicly frames its work as a systematic investment process integrating human judgment with machine intelligence, not as autonomous AI. Two Sigma’s own materials describe AI success in investing as a matter of channeling capability wisely, not switching it on. The hybrid framing is closer to operational truth than the autonomous one.
These firms also separate research skill from deployment skill. Building a model that predicts is one job. Validating it before live capital, watching its behavior in production, controlling the version history, and stopping it when its assumptions break are entirely different jobs.
The CFA Institute’s 2024 Investment Model Validation guide treats validation as fiduciary discipline, not a tech function. That framing matches what serious teams describe in due diligence.
In the AI trading decks we review, the firms that describe their work as one input among several tend to be the ones whose live results match their research most closely. The firms that describe their AI as the engine tend to surface in our notes for the opposite reason.
In our review work, fewer than one in ten AI trading providers can describe their model oversight at the level the CFA Institute’s 2024 guide treats as standard. The conversation usually ends within ten minutes.
| Question area | What strong tends to sound like | What weak tends to hide |
|---|---|---|
| Source of edge | Specific, bounded, economically plausible | Broad claims about AI finding hidden patterns |
| Out-of-sample testing | Walk-forward results, named difficult periods, live-versus-paper comparison | One long backtest, no live record |
| Trading costs | Documented assumptions and realized slippage data | “Costs are conservative” without numbers |
| Model oversight | Retraining schedule, documented changes, named stop rules | Portfolio manager judgment overrides as needed |
| Capacity | Explicit limits and degradation analysis at higher AUM | “Highly scalable” implied, never demonstrated |
- The U.S. SEC charged Two Sigma with failing to address known access-control vulnerabilities in its investment models, weaknesses raised internally for years before being fixed.
- The firm repaid clients $165 million and paid a $90 million civil penalty, with enhanced supervision and remediation requirements.
- The lesson is structural. Even at one of the most quantitatively sophisticated firms in the world, model oversight is fiduciary work, not a tech-team checklist.
- A clean backtest cannot reveal whether the firm running the model has the supervision rules to catch a known weakness. Only diligence and disclosed enforcement records can.
Red flags: what weak AI trading providers hide
This is the section many readers actually need. The pattern is consistent enough across providers that it can be listed.
Black-box confidence without process clarity. If a team cannot explain the economic logic behind a model, do not assume the model is smarter than you. Assume the explanation is weaker than it should be.
Backtests with suspicious smoothness. Real strategies have scars. If every line is beautifully stable, ask what was filtered out, excluded, or optimized away.
Loose language around adaptation. “Self-learning” can mean disciplined online updating with named retraining rules. It can also mean undocumented parameter drift that nobody is tracking.
No hard discussion of trading costs. Any AI trading presentation that spends more time on accuracy statistics than on turnover, slippage, and market impact is incomplete.
Heavy emphasis on the technology stack, light emphasis on investment evidence. Impressive infrastructure is not the same as persistent edge.
A common misunderstanding sits underneath all of this. People treat complexity as evidence of sophistication. More often, it is evidence that the manager has not distilled what actually drives returns.
The conversation that usually separates serious AI trading firms from weaker ones is the one about what stops the model. A serious team has thought about it, written it down, and tested it. A weak team treats the question as if no one had ever asked it before.
This is not the failure mode of a few bad actors. It is the failure mode of most providers in this category, which is why so few of what we review ends up listed.
- The U.S. SEC settled charges against Delphia and Global Predictions on the same day, 18 March 2024, for false and misleading claims about their AI capabilities. Combined penalties: $400,000.
- Global Predictions had marketed itself as “the first regulated AI financial advisor.” The SEC found that claim was not supported by the firm’s actual methodology.
- The CFA Institute’s June 2025 AI Washing framework asks four diligence questions: who leads the AI work, how are models validated, how is overfitting controlled, how is drift monitored. Most marketing decks answer none of them.
- The pattern visible across enforcement and review is consistent. Firms that cannot describe their AI in operational terms are usually firms whose AI is something other than what is described.
Questions to ask before you invest in AI trading
A serious diligence meeting should leave the manager slightly uncomfortable. The aim is not to test buzzwords. It is to find the actual weak points, on the four areas where AI trading typically breaks.
A question we have learned to ask before any architecture discussion: “What would make you stop using this model?” The answers separate the marketers from the practitioners faster than almost anything else.
Research integrity. How was the original hypothesis formed before the model search widened? What data was excluded, and why? How does the firm prevent repeated specification testing from manufacturing false confidence?
Which periods or market types hurt the model most, and what did the firm change in response? One worthwhile follow-up, drawn from the CFA Institute’s 2025 work on synthetic data: does the firm test models against scenarios beyond the realized historical path, or only against what the market happened to do?
Model oversight. Who approves model changes? How often is retraining permitted, and against what criteria? What happens when live performance diverges from the expected range?
Can the firm produce a version history with the rationale for each major change?
Implementation quality. What is realized turnover at current AUM, and how does it compare to research assumptions? How do paper costs compare with live slippage, by venue and by asset class? What are the true capacity limits by instrument?
What pre-trade controls prevent bad orders or runaway exposure, and who can pull the kill switch?
Business credibility. Where does the team’s actual skill sit: research, deployment, infrastructure, or marketing? How dependent is the process on one or two key individuals? What evidence shows the firm can maintain the strategy through a downturn, not just describe it during a bull run?
These questions usually move the conversation further than asking which machine-learning architecture the firm uses. The architecture rarely separates the strong from the weak. The answers to these questions almost always do.
Common questions about AI trading
No. AI trading sits inside the broader quant universe. Quant trading includes many rule-based and statistical approaches that do not use modern AI at all, while AI trading specifically uses machine learning or related methods inside one or more parts of the trading process.
That distinction matters when comparing fees. A claim of AI-driven alpha is not automatically more sophisticated than a claim of quant-driven alpha. In some cases it is the same alpha with a different label.
It can, in narrow conditions. The academic record, including Gu, Kelly and Xiu (2020) and Israel, Kelly and Moskowitz (2020), shows that machine learning methods can improve return prediction in cross-sectional equity data. The improvements are real, bounded, and sensitive to capacity, costs, and regime change.
Sweeping claims of consistent double-digit AI-driven outperformance are usually either backtest artifacts or selection-biased presentations of one strong period. Serious managers describe the boundary conditions. Marketing decks do not.
Three checks usually do most of the work. Ask the manager to describe the actual statistical method underneath the AI label, in plain terms. Ask who leads the AI work and how it is validated, against what criteria.
Ask for the live-versus-research divergence. A strong AI process can usually show what its models predicted, what actually happened, and how the firm responded.
The CFA Institute’s 2025 AI Washing framework formalizes these checks. The U.S. SEC has now turned the same questions into enforcement actions.
AI-washing is the practice of marketing investment products or capabilities as AI-driven when the underlying methodology does not justify the claim. The first regulated cases were settled by the U.S. SEC on 18 March 2024 against Delphia and Global Predictions, for combined penalties of $400,000.
For investors, AI-washing is now a real diligence risk, not a theoretical one. The CFA Institute’s June 2025 framework gives a structured way to test for it.
Technically, general AI systems can be connected to trading workflows. Several public pilots have shown this with varying levels of success. That is very different from being suitable for sustained, real-money trading.
A general-purpose AI agent has no model validation history, no documented oversight, no capacity analysis, no live track record, and no fiduciary obligations to its user. It is closer to an experiment than a product.
The conditions that make AI trading serious are clear: disclosed methodology, validation, oversight, and live evidence. These are not what general AI agents currently offer.
Two names dominate this question now: ChatGPT, and increasingly Claude. We put both through the same test rather than take the hype at face value. For the full tool-by-tool review, see our ChatGPT trading page, and the same evidence review for Claude AI trading.
What we feature, and what we don’t
The AI and machine-learning “strategies” pitched to us almost never clear the bar this page describes, so we do not list them. That is deliberate. We do not believe AI-driven trading is proven enough yet to put in front of an investor as something to trust; the live evidence is not there.
What we feature is the opposite. The Review covers proven strategies with a real live track record, judged on risk before performance is ever discussed. That selectivity is the point of this site. The premium is in the filter, not the catalogue, and most of what crosses the desk does not pass.
Get in touch →