AI Trading Algorithms Can Make Markets Less Competitive

Artificial intelligence bots are increasingly acting as market makers, setting prices and matching buyers with sellers in financial markets. But they do not always make those markets more competitive.  

In our new research, we show that these systems can prop prices up instead of pushing them down.

In theory, market makers should drive prices lower by competing for trades, since orders go to whoever offers investors the cheapest price. In practice, they don’t when it comes to AI.

In our simulations, prices slump at first, but then stall. The bots end up sometimes quoting investors’ prices almost twice as high as the “Glosten-Milgrom” rate – the level you’d expect if competition left market makers with no profit.  

And the gap persists even after the bots were trained at length. We ran them through about 1million simulated trades repeatedly, and even then, prices were still too high relative to what competition should deliver. In practice, the AI would need far more data – potentially billions of trades – to learn to be more competitive, but financial markets move too quickly for that to happen.

That said, these systems do not fail completely, since they learn to protect themselves from bad trades. When the risk of selling too cheaply is higher, they hike their prices. But that same learning does not translate into stronger competition. They stop cutting prices too soon and end up charging well above the “Glosten-Milgrom” level.

This matters for how markets are designed.  

When market fixes backfire

Take “tick sizes” – the minimum gap between buy and sell prices set by regulators. Cutting them is meant to boost competition. But it does not always work as intended when AI is involved.

When that gap is smaller, it becomes easier and more profitable for bots to slightly undercut each other, which in theory should push prices down and make markets more competitive. But in the simulations, the opposite can happen. Cutting the minimum tick size can sometimes lead to wider spreads – meaning higher costs for investors.

This is because there are too many options for the bots to consider. So each price gets tested less and that slows down their learning. In effect, they spread themselves too thin – the implication being the rules of the market can backfire when prices are set by algorithms.

US regulators have already started moving in that direction, cutting tick sizes from a penny to half a cent for some stocks to sharpen competition. This points to a broader issue: much of market regulation is still built on models of how humans behave. But when prices are set by machines, those assumptions don’t always hold. One therefore needs a new approach for securities market design in the age of AI-powered trading agents.

Why uncertainty blunts competition

Those assumptions break down in many ways – for example, when demand for an asset or its value becomes more uncertain. In that case, it’s harder for the bots to judge whether undercutting on price is profitable. That further dulls their learning and, in the simulations, leads to weaker competition and wider spreads.

The problem is most pronounced in situations the bots see less often – when they have fewer chances to learn, and prices become even less competitive.

They also react unevenly to changes in market conditions: when their profits are squeezed, they quickly raise prices but when conditions improve, they are slow to cut them again. So the gains are not passed on, but the losses are.

More bots still helps

That said, adding AI in large volumes does help the markets in some ways. In the long run when more bots are added, prices move closer to the competitive “Glosten-Milgrom” benchmark. With more rivals, the gains from undercutting become clearer and less noisy, helping the bots learn faster.  

The result is that spreads fall gradually as more AI enters the market. 

All of this stems from how these systems learn. They rely on trial and error, but the signals they get are noisy, making it hard to tell whether cutting prices is actually profitable. And over time, they stop exploring too soon and settle for prices that seem the most profitable but are not competitive.  

The result is a market where prices adjust, but not fully, and competition is weaker than theory would predict. More automation can ultimately mean less pressure on prices, and leave investors paying more.