The world's financial exchanges no longer resemble the chaotic trading floors of popular imagination — shouting traders, hand signals, paper tickets. Today, the vast majority of equity trading in the United States and Europe is executed by algorithms: computer programmes that read market data, make decisions, and submit orders in microseconds. Artificial intelligence has made these systems more adaptive, more predictive, and more profitable than their rule-based predecessors.
A Brief History of Algorithmic Trading
Electronic trading began in earnest in the 1970s with the introduction of the NASDAQ exchange. Programme trading — using computers to execute a basket of stocks simultaneously — emerged in the 1980s and was associated with the 1987 market crash, when cascading sell programmes amplified the decline. Regulatory changes in the 1990s and 2000s, particularly the SEC's Regulation NMS (National Market System) in 2005, fragmented markets across multiple exchanges and created the conditions for high-frequency trading (HFT) to flourish.
By the early 2010s, HFT firms were responsible for 50–70% of US equity trading volume. These firms competed on latency — the time between receiving market data and transmitting an order. Co-location services (hosting servers in the same data centre as the exchange) and proprietary fibre-optic and microwave networks reduced latency to microseconds and nanoseconds respectively.
Types of Algorithmic Strategies
Execution algorithms are designed to minimise market impact when trading large orders. Strategies such as VWAP (Volume-Weighted Average Price) and TWAP (Time-Weighted Average Price) slice large orders into smaller pieces and distribute them over time according to historical volume patterns. More sophisticated algorithms predict short-term price movements and accelerate or slow execution accordingly — adaptive algorithms that respond to real-time conditions.
Market-making algorithms continuously post bid and ask quotes, profiting from the spread between the two while managing inventory risk. AI-driven market makers use predictive models to adjust quotes in response to order flow toxicity — the probability that an incoming order comes from an informed trader who knows something the market maker does not.
Statistical arbitrage strategies exploit small, temporary price discrepancies between related instruments. Pairs trading — taking opposing positions in two historically correlated stocks when their spread diverges — is a classic example. Modern implementations use ML models to identify arbitrage relationships across hundreds of instruments simultaneously.
Machine Learning in Execution Optimisation
Reinforcement learning has emerged as a powerful framework for execution optimisation. Unlike supervised learning approaches that predict a target variable, RL trains an agent to make sequential decisions that maximise cumulative reward — in this case, minimising implementation shortfall (the cost of executing a large order relative to the mid-price at the time the order was placed).
An RL execution agent observes the current market state — order book depth, recent price trend, remaining inventory, time remaining — and decides how aggressively to trade. By training across historical market data, the agent learns nuanced policies that outperform static VWAP benchmarks, particularly in volatile or illiquid conditions.
Alternative Data and Signal Generation
Beyond traditional price and volume data, algorithmic traders consume an ever-expanding universe of alternative data sources. Satellite imagery is used to track oil tanker movements, estimate crop yields, and count cars in retail car parks. Credit card transaction data provides real-time estimates of consumer spending. Web scraping of job postings reveals companies that are expanding or contracting. Natural language processing of news, regulatory filings, and social media enables near-instantaneous parsing of text for market-moving information.
Processing these signals and converting them into actionable trading ideas requires sophisticated ML pipelines. Feature engineering, signal decay analysis, and portfolio construction optimisation are all areas where modern machine learning adds substantial value over classical statistical methods.
Risk Management and AI
Speed and automation introduce new forms of risk. "Flash crashes" — sudden, violent price moves caused by algorithmic feedback loops — have occurred multiple times, most notably in May 2010 when the Dow Jones Industrial Average fell nearly 1,000 points in minutes before recovering. Modern risk systems use ML anomaly detection to identify unusual algorithmic behaviour and automatic circuit breakers to pause trading when conditions deteriorate.
Regulatory oversight has intensified as algorithmic trading has grown. Regulators in both the US and EU now require firms to maintain detailed audit trails of algorithmic decisions and to demonstrate robust pre-trade and post-trade risk controls. The challenge for compliance teams is applying ML at the same speed as trading systems to ensure oversight does not introduce unacceptable latency.