DEV Community

Academic Quant
Academic Quant

Posted on Originally published at academicquant.co.uk

Will AI trading bots arbitrage away every edge? A spread model with numbers

Someone asked on Hacker News this week whether AI trading bots will eventually take all the edge out of markets. I don't think they can, and there's a short argument for it plus a small model you can check by hand.

The argument is Grossman and Stiglitz's 1980 paper, "On the Impossibility of Informationally Efficient Markets". If prices already reflected everything, nobody would pay to collect information, and then prices couldn't reflect it. So in equilibrium there's just enough mispricing left over to pay the people who go looking for it. Better tools make the looking cheaper, which shrinks what's left, but it doesn't get to zero.

The model is about the other kind of profit people call edge, which is getting paid for taking risk. Say a stock is worth either 0 or 100 tomorrow, each with probability one half, and a market maker has to quote a bid and an ask today. A fraction α of the people who trade with it know the answer and buy only when it's 100 (and sell only when it's 0). Everyone else buys or sells at random.

If the market maker wants to break even, the ask should be what the stock is worth given that someone just bought. A buy comes from an informed trader with probability α when the value is 100, and from the others with probability (1 − α)/2 either way, so the ask works out at 100·(1 + α)/2 and the bid at 100·(1 − α)/2. The spread is 100α.

Informed share α Bid Ask Spread
2% 49 51 2
10% 45 55 10
30% 35 65 30

At α = 10% a random buyer pays 55 for something worth 50 on average and loses 5, and an informed buyer pays 55 for something worth 100 and makes 45. The market maker makes nothing on average. The spread just passes money from the uninformed to the informed, and 10 is exactly what it has to charge to come out even. A real one also wants paying for holding inventory and tying up capital, which is where its profit comes from.

That's the part I'd expect AI to change and not remove. If models make public information cheap to process, more of what traders know is already in the quote, so α (the share of trades carrying news the quote hasn't priced) falls and spreads get tighter (as they did when trading went electronic). But as long as some traders know more than the quote, someone has to be paid to stand there and trade with them.

I've worked through the interview version of this, with how far a single informed trade should move your quote, here: https://academicquant.co.uk/blog-market-making-interview-game?utm_source=devto&utm_medium=post&utm_campaign=devto-ai-arbitrage-edge

Top comments (0)