AI trading: what it is, how it works, and what the evidence shows
AI trading is the use of machine-learning, natural-language and adaptive computational systems to analyse market data, generate trading signals, size and route orders, and monitor positions. It is a subset of algorithmic trading, it is legal and regulated in Australia, New Zealand and the United Kingdom under technology-neutral rules, and the published evidence does not show it reliably beating a low-cost index after costs.
Published 27 August 2026 · Updated 28 August 2026 · AI Trading Book Editorial · Reading time about 18 minutes
- Definition. AI trading applies learning or language models to market and alternative data to forecast, signal, size, execute and monitor. Regulators regulate the resulting activity, not the technology label.
- Adoption is high, autonomy is not. 75% of 118 UK firms surveyed by the Bank of England and FCA already used AI; only 2% of use cases were fully autonomous (21 November 2024).
- Legal everywhere in scope. The FCA states it does not plan extra AI-specific regulation; ASIC and the FMA apply existing frameworks. Activity, not the algorithm, determines licensing.
- Accuracy is modest. Careful studies cluster at 55–65% directional accuracy. 90%+ claims usually indicate look-ahead bias.
- The dominant retail harm is fraud and leverage. 68% of Australian retail CFD investors lost money in FY2024; ASIC removed 11,964 scam sites in calendar 2025.
What is AI trading?
AI trading is the use of machine-learning, natural-language, optimisation or adaptive computational systems to analyse market and alternative data, estimate prices or risks, identify trading signals, size and route orders, and monitor open positions. Decisions range from human-assisted, where the system only produces an alert, to fully automated, where orders reach the venue without a human in the loop. The term describes a family of methods rather than a single product.
There is no universal legal category called "AI trading". Regulators normally regulate the resulting activity — dealing, advice, portfolio management, market access, financial promotion — and apply the same obligations whether or not a model learns. This is the single most important structural fact for anyone trying to work out which rules apply to a given tool.
A one-sentence version for a general audience: AI trading uses computer systems that learn from data to make, or help make, trading decisions.
AI trading versus algorithmic trading
Algorithmic trading is the parent category: any trading decision executed by computer according to a program. AI trading is the subset in which the program learns from data, interprets language, or adapts its own parameters. A moving-average crossover system automated on a broker API is algorithmic but not AI. A gradient-boosted classifier trained on order-book features is both.
The distinction matters for reading statistics. Figures such as "60 to 75% of volume is algorithmic" describe the parent category. They are routinely misquoted as evidence about AI specifically, which they are not.
Terminology and the semantic field
These terms are related but not interchangeable, and conflating them is the most common source of confusion in vendor marketing:
- Algorithmic trading / algo trading — computer-executed rules, learning or not.
- Systematic trading — rule-governed rather than discretionary; may use no AI at all.
- Quantitative trading — decisions derived from mathematical and statistical models; AI is one toolset within it.
- High-frequency trading (HFT) — a speed regime, measured in microseconds to milliseconds. Speed, not intelligence, is the defining feature.
- Robo-advice — automated recommendations or portfolio management for clients. A regulated advice activity in all three jurisdictions.
- Trading bot / expert advisor (EA) — retail-facing automation, often on MetaTrader or a crypto exchange.
- Copy or social trading — replicating another account's trades; can itself constitute regulated dealing or advice.
- Sentiment analysis / NLP in finance — one possible input, not a complete system.
- Agentic trading — tool-using models that plan and execute multi-step sequences.
- Execution algorithms — VWAP, TWAP, percentage-of-volume and implementation shortfall; these minimise trading cost rather than pick direction.
How AI trading works: the decision loop
Every serious system, retail or institutional, runs some version of the same nine-stage loop. Vendors differ mainly in which stages they hide from the user.
What happens at each stage
- Define the decision. Forecast a return, classify direction, detect an event, execute a parent order, rebalance, or constrain risk. A system with no single stated objective cannot be validated.
- Acquire data. Prices, order books, fundamentals, macro series, filings, news wires, social text, or proprietary alternative data such as card transactions and satellite imagery.
- Clean and timestamp. De-duplicate, align corporate actions, reconcile venues, and record when each fact actually became knowable. This stage is where look-ahead bias is either prevented or baked in permanently.
- Create features. Returns, volatility, technical indicators, embeddings, sentiment scores, regime labels, liquidity measures, portfolio exposures.
- Train or specify the model. Fixed rules, supervised learning, unsupervised clustering, NLP and large language models, reinforcement learning, or portfolio optimisation.
- Convert output into a trade. Threshold the signal, size the position, apply exposure and loss limits, choose an order type. A signal is not a strategy until this stage exists.
- Validate realistically. Chronological out-of-sample periods, fees, spreads, slippage, latency, rejected orders and survivorship-safe universes.
- Deploy gradually. Paper trade through the intended broker, then minimal capital, with drift monitoring, immutable logs and a human kill switch.
- Learn under governance. Retrain through an approved change process and compare live results with the expected distribution rather than silently editing production code.
Eight functional categories and their failure modes
The most useful way to classify AI trading is by what the system outputs, because each output class fails in a characteristic way. A production system usually combines several.
| Category | Primary output | Typical method | Main failure mode |
|---|---|---|---|
| Market analysis | Regime, trend, anomaly | Clustering, feature models | Regime instability |
| Price or return prediction | Numeric return or direction | Trees, neural networks, ensembles | Non-stationarity and data leakage |
| Signal detection | Buy, sell, hold, rank | Classification, pattern recognition | Multiple-testing false positives |
| News and sentiment | Event and sentiment score | NLP, LLMs, embeddings | Hallucination, timestamp and source errors |
| Automated execution | Orders and schedules | TWAP, VWAP, routing, reinforcement learning | Latency, market impact, runaway orders |
| Portfolio optimisation | Weights and hedges | Mean-variance, Black-Litterman, robust optimisation | Unstable inputs, crowded positions |
| Risk management | Limits, VaR, anomaly alerts | Forecasting, stress testing | Tail risk, model misspecification |
| Agentic trading | Multi-step research and action | Tool-using LLM agents | Excess authority, weak auditability |
Other useful cuts
By autonomy: signal-only, where a human places the order; semi-automated, where a human approves; fully automated, where the system trades unattended. The Bank of England and FCA survey found only 2% of UK financial-services AI use cases were fully autonomous, against 55% involving some automated decision-making — a useful corrective to the assumption that "AI trading" means unattended trading.
By model class: statistical rules; classical machine learning such as random forests and gradient boosting; deep learning such as LSTM and transformer architectures; reinforcement learning; and generative or agentic large language models.
By horizon: high-frequency (microseconds to seconds), intraday, swing (days to weeks), position (weeks to months), and portfolio management (months to years).
By asset class: equities, foreign exchange, crypto, futures and options. Automation is not uniform: the algorithmic share of FX orders rose from roughly 25% in 2006 to about 80% in 2016, making FX the most automated major market.
By deployment: hosted software-as-a-service, self-hosted engines such as the open-source LEAN, broker-integrated tools, and no-code builders.
How much of the market is automated?
Algorithmic execution, taken broadly, already dominates volume in developed markets, and has for over a decade. The share of that volume attributable specifically to learning models is not separately published by any authority we could verify.
Two cautions on reading this chart. First, definitions differ between sources, and none of these series is an official statistic. Second, "algorithmic" here includes simple execution algorithms that no one would call artificial intelligence. Treating a 70% algorithmic share as evidence that "AI runs the market" is a category error that appears constantly in vendor material.
Who is actually using AI, and how far
The strongest public baseline we could verify for adoption is the Bank of England and FCA survey of 118 firms, published 21 November 2024. It measures AI across financial services rather than trading alone, and its central finding is a gap between breadth of adoption and depth of autonomy.
Three readings matter. Adoption is near-universal but shallow: most use cases are decision support. Accountability is formally assigned in 84% of firms, yet only 34% report complete understanding of the AI they use — the governance gap ASIC warned about in its own review. And a third of use cases depend on third parties, with the top three cloud providers accounting for 73% of named providers, which turns vendor outage into a shared, correlated risk rather than an isolated one.
History and current state
No single person invented AI trading and no single date marks its arrival. It emerged from the convergence of electronic markets, quantitative finance, statistical learning and automated execution. Any account crediting one founder is marketing.
| Date | Event | Why it matters |
|---|---|---|
| 1970s–80s | Program trading; portfolio insurance implicated in the 1987 crash | First demonstration that automated strategies can amplify a decline |
| 2001–2005 | US decimalisation; Regulation NMS | Market-structure changes that made electronic and high-frequency strategies viable |
| 6 May 2010 | The Flash Crash; CFTC and SEC staff report dated 30 September 2010 | Established automated feedback and liquidity withdrawal as a systemic control problem |
| 1 August 2012 | Knight Capital loses approximately USD 440m in about 45 minutes from a faulty deployment | The canonical deployment-governance failure; the SEC later fined the firm USD 12m under the Market Access Rule |
| 2013–2015 | Bailey, Borwein, López de Prado and Zhu develop the Probability of Backtest Overfitting | Formalised selection bias in investment simulations |
| 3 January 2018 | MiFID II and RTS 6 algorithmic-trading regime takes effect | Created the systems-and-controls obligations that still govern UK algorithmic trading |
| February 2018 | FCA, "Algorithmic Trading Compliance in Wholesale Markets" | First detailed UK supervisory expectations for algorithmic governance |
| 2019–2021 | FCA PS19/18 (from 1 Aug 2019) and ASIC's product intervention order (from 29 Mar 2021) restrict retail CFDs | Leverage caps, margin close-out and negative-balance protection become law |
| 21 November 2024 | Bank of England and FCA publish the 118-firm AI survey | The strongest public UK adoption baseline available |
| 29 October 2024 | ASIC REP 798, reviewing 624 AI use cases across 23 licensees | Documented that adoption was outpacing risk and compliance frameworks |
| 12 March 2025 | IOSCO consultation report on AI in capital markets (IOSCOPD788) | Shifted supervisory focus from isolated models to lifecycle and third-party risk |
| 21 August 2025 | FCA multi-firm review of algorithmic trading controls, 10 principal trading firms | Concrete observations on testing, inventories and vendor understanding; no new rules |
| 27 August 2025 | ASIC consultation on modernising trading-system and automated-trading rules (CP 386) | Signals direction of Australian reform — a proposal, not final law |
| May 2026 | IOSCO final Supervisory Toolkit for AI Use in Capital Markets (FR/02/2026, IOSCOPD823) | Non-binding lifecycle supervision framework that explicitly addresses agentic AI |
The current phase, through August 2026, is best described as structured experimentation by regulators rather than new legislation. The FCA has run its AI Lab and, with NVIDIA, a "Supercharged Sandbox" announced on 9 June 2025, with testing from October 2025 and a first-cohort showcase on 28–29 January 2026. ASIC has moved in the opposite direction on the consumer side, intensifying enforcement against AI-branded scams. IOSCO has published a common vocabulary. None of this amounts to an "AI licence", and any vendor claiming to hold one is misdescribing its status.
Regulation in the UK, Australia and New Zealand
AI trading is legal in all three jurisdictions and regulated under technology-neutral rules. No regulator in scope maintains a bespoke AI trading statute. The obligations that bite are the ones attached to the activity being performed.
| Jurisdiction | Position | Key instruments | What triggers obligations |
|---|---|---|---|
| United Kingdom (FCA) | No extra AI-specific regulatory layer planned; existing frameworks apply | MAR 7A and onshored MiFID II RTS 6 (conformance testing, annual self-assessment, stress testing, kill switch); Consumer Duty; SM&CR; COBS 22.5 and PS19/18 for CFDs | Carrying on algorithmic trading, dealing, advice, portfolio management or direct electronic access as an authorised firm |
| Australia (ASIC) | Technology-neutral; modernisation of automated-trading rules under consultation | Corporations Act obligations including acting efficiently, honestly and fairly; Market Integrity Rules; REP 798; CFD product intervention order (leverage 30:1 to 2:1, extended to 23 May 2027); CP 386 as a proposal | Providing financial services (generally requiring an AFSL); operating market-participant trading systems |
| New Zealand (FMA, NZX) | Activity-based licensing; class advice always permitted | Financial Advice Provider (FAP) licence plus dispute-resolution scheme membership; DIMS requirements; NZX Participant Rules, edition dated 19 February 2026 | Providing regulated financial advice to retail clients; participant access to NZX markets |
Two practical consequences follow. If you trade your own capital through your own account, you are generally not providing a financial service, whatever software you use. The moment you take other people's money, give personalised recommendations, or promote a product, licensing applies — and the software's sophistication is irrelevant to that question.
Globally, the EU AI Act (Regulation (EU) 2024/1689, in force 1 August 2024) does not apply directly in the UK, Australia or New Zealand, but it shapes what international vendors build. Jurisdiction-by-jurisdiction detail sits in the UK page, the Australian page and the New Zealand page.
What the evidence says about performance
The independent record does not show AI reliably beating the market after costs. The Eurekahedge AI Hedge Fund Index returned 9.8% annualised from December 2009 to July 2024 against 13.7% for the S&P 500. A peer-reviewed review of 27 machine-learning equity experiments by Buczynski, Cuzzolin and Sahakian (2021) found "no conclusive evidence of any ML-driven investment funds delivering spectacular returns at scale".
On prediction, careful directional-accuracy figures cluster between roughly 55% and 65%. Reported accuracies of 80–90% or more are widely attributable to look-ahead bias, particularly where datasets are shuffled before splitting, which leaks future information into training. For retail vendors specifically, no audited, cost-adjusted, like-for-like live-return comparison exists, so any ranking of "best-performing AI bot" is fabricated by construction.
This does not make AI useless. Documented institutional benefits concentrate in fraud detection, anti-money-laundering, surveillance, execution-cost reduction and research throughput — areas where the objective is measurable and the counterfactual is clear. Full treatment on the evidence page and the accuracy page.
Advantages, stated carefully
- Data scale and speed. Systems ingest and act on volumes of heterogeneous data beyond human capacity, and act within microseconds where latency matters.
- Consistency. A rule executes identically every time, without hesitation, fatigue or revenge trading. The same property applies to a bad rule.
- Continuous coverage. Monitoring across hundreds of instruments and, in crypto, around the clock.
- Testability. A systematic strategy can be simulated before capital is risked — the single largest structural advantage over discretionary trading, and the one most often abused.
- Execution cost. Smart order routing and scheduling algorithms measurably reduce implementation shortfall on large orders.
- Non-trading applications. Fraud, AML and surveillance benefits are the best-documented in the supervisory literature.
Limitations, risks and documented harms
Model and methodology risk
Overfitting, look-ahead bias, survivorship bias and data snooping are the standing hazards. Their severity is structural rather than accidental: when analysts test very large numbers of variants and publish the best, selection bias alone produces impressive-looking results. Bailey and López de Prado's Deflated Sharpe Ratio exists precisely to correct for this, and Harvey, Liu and Zhu argued in the Review of Financial Studies in January 2016 that a newly claimed factor should clear a t-statistic above 3.0 rather than the conventional 2.0.
Regime change and concept drift then erode whatever edge survives. Slippage, spreads, commissions and market impact consume the rest. A gross edge of a few basis points per trade is not an edge at all once costs are modelled honestly.
Opacity and accountability
ASIC's REP 798 cited a licensee whose credit-scoring model was described internally as a black box, with no ability to explain which variables drove an outcome. The FCA's 2025 multi-firm review found the analogous problem in trading: firms using third-party algorithms sometimes relied on vendor simulation and lacked good technical understanding of how those algorithms were built. Outsourcing the code does not outsource the market impact.
Operational and concentration risk
Credential compromise, exchange or broker outage, stale data feeds and rejected orders are ordinary events, not tail events. Because a third of surveyed use cases run on third parties and the top three cloud providers represent 73% of named providers, failures tend to be correlated across firms rather than isolated. The Bank of England has warned that wider AI trading could produce correlated positions that amplify shocks.
Fraud: the largest documented consumer harm
In Australia, ASIC coordinated the removal of 11,964 phishing and investment-scam websites during calendar 2025, against 6,270 in the preceding twelve-month period, and more than 25,000 since the service launched in 2023. Fake brands marketed as AI or quantum trading platforms recur, alongside deepfake celebrity endorsements. ASIC's guidance is blunt: verify that the promoter holds an Australian Financial Services licence before providing personal details or money. See the scam patterns page.
Who AI trading suits, and who it does not
Potentially suitable: quantitative researchers and developers comfortable with data engineering; regulated firms with model-risk controls; active traders who want scanning and alerting rather than autonomous profit; and crypto users who genuinely understand exchange, custody, API and volatility risk.
Usually unsuitable: anyone seeking guaranteed passive income; anyone who cannot explain their own strategy in a paragraph; anyone using borrowed or essential funds; beginners starting with leverage; anyone unwilling to monitor a running system; and anyone who cannot verify licensing, custody arrangements and API permissions.
Myths and corrections
| Claim | Correction |
|---|---|
| "AI predicts the market with 90% accuracy" | Careful peer-reviewed work reaches roughly 55–65% directional accuracy; higher figures typically reflect look-ahead bias or shuffled time-series splits |
| "AI trading means fully autonomous trading" | Only 2% of surveyed UK financial-services AI use cases were fully autonomous; 55% involved some automated decision-making |
| "Bots make money while you sleep" | Between 68% and approximately 82% of retail CFD accounts lose money in the published regulator figures |
| "A strong backtest proves a durable edge" | Repeated selection across many variants manufactures apparent skill; live frictions and regime change remain |
| "Removing emotion removes risk" | Automation scales discipline and operational error at the same rate, and does not reduce leverage |
| "A free bot is free to trade" | Exchange-native bots still charge transaction fees, and every workflow adds spreads, data, broker and tax costs |
| "This platform is AI-regulated" | No such status exists. Regulators apply activity-based frameworks; the FCA has said it does not plan an extra AI-specific layer |
| "Vendor win rates are comparable" | No common audited live benchmark was found for any retail vendor set |
Beginner mistakes and their corrections
| Error | Why it fails | Correction |
|---|---|---|
| Optimising hundreds of variants and reporting the best | Selection bias creates false discoveries | Pre-register the hypothesis; use purged or walk-forward out-of-sample tests |
| Shuffling time-series data before splitting | Future regimes leak into training | Split chronologically and preserve information-arrival times |
| Ignoring delisted securities | Survivorship bias inflates historical results | Use point-in-time universes with corporate actions |
| Testing without fees or slippage | Small gross edges disappear in live trading | Model spread, commission, market impact, latency and rejections |
| Granting API withdrawal rights | Credential compromise becomes a custody loss | Trade-only keys, IP allowlists, subaccounts, rapid revocation |
| Going live at full size | Operational errors scale immediately | Paper trade, then a staged capital ladder with predetermined gates |
| Treating model prose as verified data | Language models can fabricate numbers and sources | Retrieve primary documents and validate structured fields |
| Buying on a win rate alone | Win rate omits payoff asymmetry and drawdown | Require net expectancy, maximum drawdown, turnover, tail loss and capacity |
| Assuming a regulated broker validates a third-party bot | Broker licensing does not certify bot performance | Evaluate the vendor, the code, the custody model and the promotion separately |
Mini-glossary
| Term | Definition |
|---|---|
| Alpha | Return beyond the chosen benchmark, after risk and cost adjustment |
| Backtest | Simulation of a strategy on historical data |
| Look-ahead bias | Use of information that was not available at the simulated decision time |
| Survivorship bias | Testing only instruments that survived to the present |
| Overfitting | Learning noise or peculiarities that do not generalise |
| Walk-forward testing | Repeated training on past windows and testing on subsequent unseen windows |
| Concept drift | Change in the relationship between inputs and outcomes |
| Slippage | Difference between the expected and the realised execution price |
| Market impact | Price movement caused by the strategy's own orders |
| Sharpe ratio | Excess return divided by return volatility; horizon and assumptions matter |
| Maximum drawdown | Largest peak-to-trough decline over a period |
| Latency | Delay between decision and execution, measured in microseconds in HFT |
| DEA | Direct electronic access to a venue through a participant |
| Kill switch | A control that immediately blocks or cancels automated activity |
| Agentic AI | AI that plans and executes multi-step tasks through tools or APIs |
Frequently asked questions
What is AI trading, in one sentence?
AI trading is the use of learning or language models to analyse market and alternative data and to generate, size, execute or monitor trading decisions, with autonomy ranging from alerts to unattended execution.
Is AI trading the same as algorithmic trading?
No. Algorithmic trading is the parent category of computer-executed rules; AI trading is the subset where the system learns or adapts. The FCA regulates the broader category through MAR 7A systems-and-controls requirements rather than through a separate AI category.
Is AI trading legal in the UK, Australia and New Zealand?
Yes. None of the three regulators bans the technology or has an AI-specific statute. Licensing follows the activity — dealing, advice, portfolio management, market access — not the software label.
Do I need a licence to run a bot on my own account?
Generally no, if you trade only your own capital through your own account. Obligations arise when you deal for others, advise, manage third-party money, promote a product, or obtain direct electronic access. Borderline arrangements need jurisdiction-specific legal advice.
Can AI predict prices accurately?
Modestly. Careful peer-reviewed models reach roughly 55–65% directional accuracy. Claims of 90%+ are widely attributed to look-ahead bias, especially where time-series data are shuffled before splitting.
Do AI hedge funds beat the market?
Historically not, on the published index evidence: 9.8% annualised for the Eurekahedge AI Hedge Fund Index versus 13.7% for the S&P 500 between December 2009 and July 2024.
How autonomous is AI in finance today?
Only 2% of use cases in the Bank of England and FCA survey of 118 firms were fully autonomous, against 55% involving some automated decision-making. AI in finance is overwhelmingly decision support with a named accountable human.
What share of trading is algorithmic?
Roughly 60–75% of overall volume in US equity, European and major Asian markets according to Select USA, with high-frequency strategies at approximately 50–55% of US equity volume in 2023. These describe algorithmic trading broadly, not AI.
What is the biggest risk for a retail user?
Fraud first, leverage second, model error third. Regulator figures put retail CFD loss rates between 68% and approximately 82%, and thousands of AI-branded scam sites are removed each year.
What is agentic trading?
Tool-using AI that plans and performs multi-step research and action rather than emitting a single prediction. IOSCO addressed it explicitly in its May 2026 supervisory toolkit; the characteristic risks are excessive authority and weak auditability.
Does the EU AI Act apply to UK, Australian or New Zealand firms?
Not directly. Regulation (EU) 2024/1689 has applied in the European Union since 1 August 2024 and shapes global vendor practice, but the three jurisdictions in scope rely on existing financial-services frameworks.
Are AI trading subscriptions expensive?
Retail tools in the compared set run from USD 0 to about USD 200 per month, with Trade Ideas Premium at USD 178 per month or USD 2,136 per year. Subscription price excludes broker commissions, exchange data, spreads, slippage and tax.
Is copy trading regulated?
It can be. Depending on how it is structured, copy trading may constitute regulated dealing or advice, and providers offering it in these markets generally operate under FCA or ASIC authorisations.
Can a language model write a profitable strategy for me?
It can write code and hypotheses; it cannot validate them. Everything that determines whether a strategy survives — point-in-time data, chronological splits, cost modelling, capacity, drawdown tolerance — happens outside the text generation, and language models will fabricate plausible numbers and citations if asked to supply evidence.
Conclusions
AI trading is a real and mature set of techniques, legal and regulated in Australia, New Zealand and the United Kingdom, and genuinely useful in narrow, measurable roles — execution cost, surveillance, fraud detection, research throughput. It is not a route to reliable excess returns for retail participants, and the published evidence has been consistent on that point for over a decade.
The defensible position is not "buy the smartest model". It is a narrow objective, point-in-time data, a simple non-AI baseline the model must beat net of costs, chronological out-of-sample validation, explicit cost and tail-risk modelling, regulated execution, restricted credentials, staged deployment and named human accountability. Edge comes from process, not from the label on the software.
For anyone deciding whether to start: verify the provider's licence with ASIC, the FMA or the FCA before anything else; assume leverage rather than the model is your dominant risk; treat every backtested or marketed performance figure as inflated until you have reproduced it out of sample with costs; use free tiers and paper trading before subscribing; and benchmark honestly against a low-cost index fund, which most active and AI strategies fail to beat after fees.
About this page
Compiled by the AI Trading Book editorial desk from four independent deep-research passes, then reduced to claims traceable to a named primary source. Two of the four source passes contained regulatory documents that could not be located in the issuing authority's own register; those claims were excluded rather than softened, and our editorial policy explains the test applied.
Published 27 August 2026; last updated 28 August 2026. Next scheduled review: November 2026. Found an error? Tell us — corrections are logged publicly.
Sources
- ASIC, REP 798 "Beware the gap: Governance arrangements in the face of AI innovation" and media release 24-238MR — 29 October 2024 — 624 use cases across 23 licensees; black-box example.
- ASIC, REP 828 — 20 January 2026 — 68% of retail CFD investors lost money in FY2024.
- ASIC, 20-254MR product intervention order and 22-082MR extension — 2020 and April 2022 — leverage caps 30:1 to 2:1; extension to 23 May 2027.
- ASIC, 26-063MR "ASIC ramps up action to protect consumers from AI-powered online investment scams" — 8 April 2026 — 11,964 sites removed in 2025; 6,270 prior period; more than 25,000 since 2023.
- ASIC, CP 386 "Proposed amendments to the ASIC market integrity rules: Trading systems and automated trading" — 27 August 2025 — proposal, not final law.
- Bank of England and FCA, "Artificial intelligence in UK financial services – 2024" — 21 November 2024, n=118 — adoption, autonomy, third-party and provider-concentration figures.
- FCA, "AI and the FCA: our approach" — first published 8 September 2025, updated 13 February 2026 — no extra AI-specific regulation planned.
- FCA Handbook, MAR 7A and onshored MiFID II RTS 6 — algorithmic-trading systems-and-controls requirements.
- FCA, "Multi-firm review of algorithmic trading controls: high-level observations" — 21 August 2025 — 10 principal trading firms.
- FCA, PS19/18 and Handbook COBS 22.5 — from 1 August 2019 — CFD leverage caps, close-out, negative-balance protection, quarterly loss-percentage disclosure.
- FMA, "Financial Advice Provider (FAP)" — page updated 1 July 2026 — licensing and dispute-resolution requirements.
- NZX, Participant Rules — rules dated 19 February 2026 — participant obligations on NZX markets.
- IOSCO, CR/01/2025 (IOSCOPD788) — 12 March 2025 — AI use cases, risks and challenges.
- IOSCO, FR/02/2026 Supervisory Toolkit for AI Use in Capital Markets (IOSCOPD823) — May 2026 — lifecycle supervision; agentic AI.
- CFTC and SEC staff, "Findings Regarding the Market Events of May 6, 2010" — 30 September 2010.
- Bailey, Borwein, López de Prado and Zhu, "The Probability of Backtest Overfitting" — revised 2015; and "Pseudo-Mathematics and Financial Charlatanism", Notices of the AMS, 2014.
- Harvey, Liu and Zhu, "…and the Cross-Section of Expected Returns", Review of Financial Studies 29(1) — January 2016 — t-statistic above 3.0.
- Buczynski, Cuzzolin and Sahakian, International Journal of Data Science and Analytics 11(3) — April 2021 — review of 27 machine-learning equity experiments.
- Eurekahedge AI Hedge Fund Index, as reported by IG — 21 November 2024 — 9.8% versus 13.7% annualised, December 2009 to July 2024.
- Select USA and QuantifiedStrategies — algorithmic share of volume; FX order series 2006 to 2016.
- SEC market-structure reporting — 2023 — high-frequency share of US equity volume.
- Bank of England, "Financial Stability in Focus: Artificial intelligence in the financial system" — 9 April 2025 — correlated positions amplifying shocks.
- Regulation (EU) 2024/1689 (EU AI Act) — in force 1 August 2024.
Informational research only. Nothing on this page is personal financial, legal, tax or investment advice, or a recommendation to trade any instrument.