There are still people on the New York Stock Exchange floor. They hold phones, stand at trade posts in blue jackets, and sometimes glance at screens that show data that change more quickly than the human eye can comprehend. Even while the real trading that matters takes place in data centers in Mahwah, New Jersey, in co-located server racks running algorithms that react to market movements in microseconds, the ambiance still has a hint of what the exchange floor used to be: chaotic, noisy, and tactile. To a certain extent, the floor has always been theatrical. More and more, the market is the machine.
The foundation of contemporary stock exchange liquidity is algorithmic market making. An automated system that was already present on both sides of the transaction—offering to buy at one price and sell at a slightly higher price, continuously, across thousands of securities simultaneously, adjusting both quotes in real time based on live market conditions—made it possible for a retail investor to purchase 50 shares of a company thru an online brokerage and the trade to execute almost instantly at a price very close to the last quoted price. A computer running a machine learning model in a climate-controlled room without windows has largely replaced the human market maker of the 1990s, shouting, gesturing, and using a pen to manage a physical order book.
These models are far more sophisticated than just quoting prices. Models trained on massive historical datasets of order book behavior across numerous exchanges at the same time are needed for predictive order routing, which is the process of predicting where liquidity is most likely to arise in the next few hundred milliseconds and routing orders to that place.
The type of execution intelligence required for a large institutional order that must purchase two million shares of a stock without significantly altering the price is one that divides the order into segments, directs each segment to the venue where it will have the least impact on the market, and modifies the execution pace based on real-time signals about what other participants are doing. A human trader cannot oversee this procedure in real time. A model oversees this procedure.
The AI component is most consistently active in the dynamic spread adjustment function. The gap between what a market maker will pay to buy and what it will sell for is known as the bid-ask spread, and it is not fixed. They compress during quiet times when the cost of supplying liquidity is reduced, and they expand during times of high volatility when the risk of keeping inventory in a rapidly moving market rises.
These spreads are being adjusted in real time by deep learning models operating across Citadel Securities, Virtu Financial, Jane Street, and a few other companies that collectively supply a significant amount of global equity market liquidity. These models are based on signals that no human could process at the necessary speed, such as order flow imbalances, options market signals, news sentiment feeds, cross-asset correlations, and volatility surface changes.
When this structure works as intended, there are actual market benefits. Over the past 20 years, spreads in major equity markets have significantly tightened, primarily due to algorithmic competition. When a retail investor purchases a liquid US stock now, the transaction costs are far lower than they would have been in 2000. The quality of order execution has increased. The markets clear more quickly. Systems that are always there, always quoting have essentially eliminated the friction of locating a counterparty for a trade, which was once a real constraint that added time and cost to every transaction.
The system’s limitations are revealed in the risk management section. Inventory risk is limited for algorithmic market makers. The systems retreat when there is an abrupt rise in volatility, such as a flash crash, an unexpected news release, or a big order that comes and goes from the order book in a fashion that the models aren’t trained to manage. They delete quotes completely or drastically widen spreads, which eliminates liquidity just when market participants most need it. This trend was illustrated in real time during the 2010 Flash Crash: as automated market makers retreated, the void they created caused prices to decline in ways unrelated to any underlying fundamental shift.
