AI and algorithmic trading glossary

80 terms across eight categories, each with its own anchor so it can be linked and cited directly. Where a term has a regulatory definition, that definition is used rather than the industry one — the two frequently differ, and the difference usually matters. Where a term is marketing rather than substance, the entry says so.

Published 27 August 2026 · Updated 28 August 2026 · AI Trading Book Editorial

How the main terms nest

Four words are used interchangeably in casual discussion and describe different-sized sets. Getting this wrong is the source of most confusion about what regulators cover and what evidence applies to what.

Where the terms sit in a system

Most of the vocabulary attaches to one stage of a pipeline. Reading it in order makes the categories below easier to navigate.

What the vocabulary is made of

The distribution is itself informative. Nearly half the terms describe either validation or risk control — the parts of the discipline that decide whether anything else matters — while prediction, which dominates the marketing, accounts for a much smaller share.

Core concepts

What the field calls itself, and what those labels actually mean.

Agentic AI

Systems that plan and act over multiple steps toward a goal rather than responding to single prompts. Addressed by IOSCO in a May 2026 report (IOSCOPD823, FR/02/2026). The open questions are governance — accountability, stopping mid-sequence, reconstructing decisions — rather than capability. More →

AI trading

The use of machine learning, deep learning, natural language processing or related computational methods to analyse market data, generate forecasts or signals, manage risk and in some systems send orders. It is a marketing and industry term, not a regulatory category — no financial regulator in the UK, Australia or New Zealand defines it. More →

Algorithmic trading

Trading in which a computer algorithm automatically determines individual parameters of orders — whether to initiate, timing, price, quantity or subsequent management — with limited or no human intervention. This is the regulatory definition carried into MAR 7A in the UK; it captures machine learning and hand-written rules identically. More →

Automated trading

A broad, non-regulatory term for any trading executed by software rather than by a person clicking. Wider than algorithmic trading, since it includes systems that merely transmit orders a human specified.

High-frequency trading

A narrow subset of algorithmic trading identified in the MiFID-derived framework by three characteristics together: infrastructure minimising latency such as co-location, system determination of orders without human intervention, and high intraday message rates. All HFT is algorithmic trading; most algorithmic trading is not HFT. More →

Quantitative trading

Trading driven by statistical or mathematical models rather than discretionary judgement. Predates machine learning by decades and does not require it.

Robo-advisor

A licensed service that builds and rebalances a portfolio for a client, typically from index funds, based on a risk questionnaire and a rules-based allocation model. Usually not machine learning despite the marketing. More →

Systematic trading

Trading that follows a defined rule set consistently rather than case-by-case judgement. Can be executed manually; the defining feature is the rule, not the automation.

Models and methods

The techniques underneath, and where each is genuinely used.

Deep learning

Machine learning using neural networks with many layers. In trading, most commonly applied to sequence prediction and to extracting structure from order-book data.

Ensemble

Combining several models so that their errors partly cancel. Often improves stability at the cost of interpretability.

Explainable AI

Methods for interpreting why a model produced an output. Not named in UK rules, but effectively required in practice: a firm subject to the Consumer Duty must evidence fair outcomes, which is hard if the model cannot be interrogated. More →

Feature

An input variable supplied to a model. Feature construction is where most look-ahead bias enters, because a feature computed with information unavailable at the decision moment leaks the future. More →

Foundation model

A large general-purpose model adapted to specific tasks rather than trained for one. In the Bank of England and FCA 2024 survey, 17% of respondent firms reported using them.

Large language model

A model trained on very large text corpora to predict and generate language. Useful for code, explanation and document work in a trading context; not a validated forecasting system, and its training corpus creates a leakage channel when backtesting periods it has read about. More →

LSTM

Long short-term memory, a recurrent neural network architecture designed for sequential data. Widely used in published price-prediction research; Sirignano and Cont reported approximately 65–75% directional accuracy for deep learning against 57–67% for linear models on a high-frequency task across 500 NASDAQ stocks. More →

Machine learning

Methods that fit a model to data rather than encoding rules directly, so that behaviour is learned from examples. Covers regression, tree ensembles, neural networks and much else.

Model drift

Deterioration in performance as the relationships a model learned change. Distinct from overfitting: a drifted model once worked, an overfitted model never did.

Natural language processing

Extracting structure or meaning from text — news, filings, transcripts, social media. Used for sentiment features and document summarisation. Language is ambiguous and can be deliberately manipulated, which limits reliability.

Reinforcement learning

Training an agent to choose actions that maximise a reward signal through interaction with an environment. Applied to execution and to strategy design; published backtest results should be read with care, since very high reported Sharpe ratios usually indicate a methodology problem. More →

Transformer

A neural network architecture based on attention rather than recurrence, dominant in language modelling and increasingly applied to financial sequences. A 2025 study reported about 69.1% directional accuracy for a transformer approach. More →

Testing and validation

The vocabulary of establishing whether a result means anything.

Backtest

Simulation of a strategy against historical data. A backtest without realistic costs, slippage, delisted instruments and out-of-sample validation is not evidence of tradability. More →

Baseline

The trivial predictor a model must beat, usually the majority class. If a test period had upward moves 56% of the time, a model always predicting up scores 56% — so a reported 58% is two points of information, not eight. More →

Data leakage

Any unintended flow of information from outside the training period into the model. Look-ahead bias is the time-ordered case; leakage also occurs through a target variable encoded in a feature. More →

Deflated Sharpe Ratio

A statistic from Bailey and López de Prado (2014) adjusting an observed Sharpe ratio for the number of trials conducted, the length of the track record and the non-normality of returns. Requires an honest trial count, which is the hardest input to supply. More →

Directional accuracy

The proportion of times a model correctly predicted the sign of the next move over a defined horizon. Careful peer-reviewed work clusters at roughly 55–65%. Says nothing about magnitude, cost or frequency. More →

Look-ahead bias

Using information in a model that was not available at the moment of the simulated decision. Common forms: shuffling time-ordered data before splitting, scaling features using whole-dataset statistics, and using fundamentals dated before publication. More →

Out-of-sample

Data held back from model development and used to evaluate it. Single-use by nature: each time you look at it and adjust in response, it becomes training data. More →

Overfitting

Fitting the noise in a historical sample rather than a durable relationship, usually through repeated adjustment guided by the result. Characteristic symptom: live performance that never resembles the backtest, rather than gradually decaying from it. More →

Paper trading

Running a system live without capital. The value is measuring divergence between live signals and backtest expectations, which exposes feed, timing and fill assumptions that simulation hides. More →

Point-in-time data

Data recorded as it stood at each historical moment, including publication dates and subsequent revisions. Necessary to avoid acting on information the market did not yet have. More →

Probability of Backtest Overfitting

A measure, also from Bailey and López de Prado, of how likely the in-sample best configuration is to underperform the median configuration out of sample. It evaluates the selection procedure rather than the strategy. More →

Selection bias

Reporting the best of many tested variants as though it were a single test. At a 5% significance threshold, twenty worthless strategies give roughly a 64% chance that at least one appears significant. More →

Survivorship bias

Testing on instruments that still exist today, which excludes every company that failed, was delisted or was acquired. The excluded set is precisely the set of bad outcomes. More →

Trial count

The total number of configurations examined during a search, including those abandoned quickly. It is the input that makes statistical correction possible and the one most often understated. More →

Walk-forward testing

Repeatedly training on an earlier window and testing on the next unseen window, rolling both forward. Approximates operating a strategy through changing conditions and prevents tuning to one test period. More →

Performance measures

How results are reported, and what each measure omits.

Alpha

Return in excess of an appropriate benchmark after adjusting for risk. Claimed frequently and demonstrated rarely. More →

Benchmark

The alternative a strategy must beat to be worth running — usually buy-and-hold on a relevant index, measured with the same costs. More →

CAGR

Compound annual growth rate — the constant annual rate that would produce the observed total return over a period.

Capacity

The amount of capital a strategy can deploy before its own market impact erodes the edge. A strategy that works small can stop working at size. More →

Drawdown

Decline from a peak in portfolio value to a subsequent trough. Maximum drawdown is the largest such decline over a period.

Sharpe ratio

Return relative to volatility. Requires care over the risk-free rate and sampling frequency. In a retail backtest a value above 3 is usually a warning about methodology rather than a discovery. More →

Turnover

How frequently a strategy trades. Costs scale with turnover while the edge per trade does not, so high-frequency signals must clear a much higher bar. More →

Execution and microstructure

How an order reaches a market and what it costs to get there.

Co-location

Placing trading servers physically close to an exchange's matching engine to reduce latency. One of the characteristics identifying high-frequency algorithmic trading technique.

Direct electronic access

An arrangement where a client uses a member firm's infrastructure to send orders directly to a venue. Providers of DEA fall within the UK algorithmic trading regime. More →

Iceberg order

An order displaying only part of its size, with the remainder hidden and replenished as the visible portion fills. Reduces information leakage at the cost of queue position. More →

Implementation shortfall

Both a measure — the gap between the price when the decision was made and the final realised outcome — and a family of algorithms that trade off market impact against the risk of the price moving while waiting. More →

Latency

Delay between an event and a system's response to it. Central to high-frequency strategies and largely irrelevant to strategies holding positions for days.

Market impact

Price movement caused by your own order. Grows with order size relative to available liquidity and is the mechanism behind capacity limits.

Order book

The set of resting buy and sell orders at each price for an instrument. Its shape is the raw material for microstructure prediction.

Percentage of volume

An execution algorithm that participates at a set proportion of market volume, so the order accelerates when the market is active and slows when it is quiet. More →

Slippage

The difference between the price at the moment of decision and the price actually obtained. Omitting it from a backtest fills orders at prices that were not available.

Smart order routing

Deciding which venue or venues an order goes to, and in what sequence, across a fragmented market. Distinct from deciding what the order should be. More →

Spread

The difference between the best bid and best offer. Paid on every round trip and typically the largest single cost at high turnover.

TWAP

Time-weighted average price — an execution algorithm slicing an order into equal parts at regular intervals, ignoring volume. Simpler and more predictable than VWAP, and easier for others to anticipate. More →

VWAP

Volume-weighted average price — both a benchmark and an execution algorithm that splits an order across the day in proportion to expected volume, aiming to match the day's VWAP. More →

Risk controls

The mechanisms that limit damage when something goes wrong.

Crowding

Many participants holding similar positions from similar signals. The Bank of England noted in April 2025 that correlated positions arising from similar models and data can amplify shocks. More →

Kill switch

A mechanism halting trading and cancelling open orders immediately. Under UK-onshored RTS 6 article 12, firms engaged in algorithmic trading must be able to cancel unexecuted orders immediately; ASIC's CP 386 proposes an equivalent control. It must work independently of the strategy code. More →

Leverage

Trading a position larger than the capital backing it. At 30:1 a 3.3% adverse move eliminates the deposit. ASIC caps retail CFD leverage at 30:1 for major currency pairs down to 2:1 for crypto-asset CFDs. More →

Margin close-out

A rule requiring a provider to close a retail client's positions when account equity falls to a set proportion of initial margin — 50% under ASIC's product intervention order. It caps how far a losing position runs but does not guarantee the exit price in a gapping market. More →

Negative balance protection

A requirement that a retail client cannot lose more than the money in their account. Mandatory for retail CFD clients in Australia and the UK; does not apply to wholesale or professional clients. More →

Position sizing

Determining how much capital to allocate to each position. Often matters more to outcomes than signal quality.

Regime change

A structural shift in market behaviour that invalidates relationships a model learned. Distinct from ordinary volatility: the relationship stops existing rather than becoming noisier.

Tail risk

Exposure to rare, large adverse moves. Stop-losses can fail during gaps or illiquid conditions, so a modelled worst case is usually optimistic.

Regulatory terms

Defined terms from the rulebooks of the three markets covered.

AFSL

Australian Financial Services Licence, required to provide financial services to others. Section 912A of the Corporations Act 2001 requires services to be provided efficiently, honestly and fairly. More →

Consumer Duty

The FCA requirement to deliver and evidence good outcomes for retail customers across products, price and value, understanding and support. Technology-neutral, so it applies to AI-driven decisions unchanged. More →

DIMS

Discretionary Investment Management Service — deciding and executing investments for a client without seeking approval for each transaction. Licensed separately in New Zealand and the regime most likely to be triggered unintentionally by an automated arrangement. More →

FAP licence

Financial Advice Provider licence, the FMA authorisation required to give regulated financial advice to retail clients in New Zealand, alongside FSPR registration and dispute scheme membership. More →

MAR 7A

The FCA Handbook chapter governing algorithmic trading. Applies to authorised firms engaged in algorithmic trading, providers of direct electronic access and general clearing members. More →

Product intervention order

An ASIC order restricting or banning a product for retail clients. The retail CFD order took effect 29 March 2021 and runs to 23 May 2027; binary options were banned for retail clients from 3 May 2021. More →

RTS 6

The UK-onshored MiFID II technical standard for algorithmic trading systems. Article 6 covers conformance testing, article 9 annual self-assessment, article 10 stress testing and article 12 kill functionality. More →

SM&CR

The Senior Managers and Certification Regime, under which a named individual is accountable for their area of responsibility. There is no AI-specific function; whoever owns the business area owns the model. More →

Trading Algorithm

A term proposed in ASIC CP 386 (27 August 2025) for a computer algorithm that automatically determines, with limited or no human intervention, one or more parameters of an order. A consultation proposal, not a rule in force. More →

Wholesale client

An Australian classification removing most retail protections, including leverage caps and much of the disclosure regime. A reduction in protection rather than a status upgrade. More →

Products and tools

Instruments and software a reader will encounter.

API key

A credential pair letting software act on an exchange or broker account without a password. Permissions are set at creation and define the ceiling on any later compromise; a trading bot never needs withdrawal permission. More →

Binary option

A derivative paying a fixed amount if a condition is met and nothing otherwise. Banned for retail clients in Australia from 3 May 2021 and prohibited in the UK. More →

CFD

Contract for difference, a leveraged derivative settling the change in an instrument's price. ASIC found 68% of Australian retail CFD investors lost money in FY2024; the FCA states approximately 80% of customers lose money on CFDs. More →

DCA bot

A bot buying at fixed intervals or on defined triggers to average the entry price. Averaging into a falling asset increases exposure to it. More →

Expert Advisor

An automated strategy running inside the MetaTrader platform, historically the most common retail automation route in the UK, Australia and New Zealand.

Grid bot

A bot placing buy and sell orders at intervals across a price range, profiting from oscillation within it. When price leaves the range below the lower bound, purchases stop while the asset continues falling. More →

Index fund

A fund tracking a market index rather than selecting holdings. The benchmark against which active approaches are measured, and the underlying holding of most robo-advisor portfolios. More →

Sources

• FCA Handbook, MAR 7A; UK-onshored MiFID II RTS 6, articles 6, 9, 10 and 12; definitions of algorithmic trading and high-frequency algorithmic trading technique.
• ASIC — CP 386, 27 August 2025, proposed Trading Algorithm definition; product intervention order leverage caps and margin close-out; REP 828, 20 January 2026.
• FMA and the Financial Markets Conduct Act 2013 — FAP licensing and DIMS.
• Bank of England and FCA — "Artificial intelligence in UK financial services – 2024", 21 November 2024; Bank of England, Financial Stability in Focus, 9 April 2025.
• Bailey and López de Prado — Deflated Sharpe Ratio and Probability of Backtest Overfitting; Harvey, Liu and Zhu, Review of Financial Studies 29(1), January 2016.
• Sirignano and Cont, Quantitative Finance, 2019; ScienceDirect transformer study, 2025.
• IOSCO — IOSCOPD788, 12 March 2025; IOSCOPD823 (FR/02/2026), May 2026.

Informational research only. Nothing on this page is financial, legal, tax or investment advice. Regulatory definitions are paraphrased and simplified.