Algorithmic trading may congest markets. One fifth of all transactions on Nasdaq are clustered in sub-millisecond “microbursts” of activity (Menkveld, 2018). This is not surprising: Machines can react to trading signals much faster than humans, which means their orders can essentially reach the exchange simultaneously.
On the face of it, this is a technology issue. To tackle surges in trading activity, exchanges invest in extra capacity, faster computer chips, and bigger data buffers. At the start of 2007, the New York Stock Exchange was able to process 17,000 messages per second. Four years later, once high-frequency trading took off in 2011, NYSE’s capacity jumped to 250,000 messages per second.
Building up excess capacity is costly, as idle computer power can be directed to other, more productive tasks. How big is the buffer? Menkveld (2018) documents that only 10% of the microseconds in a trading day will experience an activity surge. Within a microburst, exchange messaging activity can increase tenfold. Therefore, 90% of exchange processing capacity remains idle for 90% of the time.
Not only is the idle capacity costly in itself, but the market quality benefits of maintaining a sizeable buffer are also unclear. Exchanges build up capacity to speed up trading, enabling high-frequency races between algorithmic traders who rush to trade on the same opportunity. A faster market, however, may not be a better market: such high-frequency races are usually zero-sum games. Shkilko and Sokolov (2019) find that when rain disrupts microwave networks, slowing down trading, market liquidity and stability improves. To counteract the negative impact of speed, some exchanges explicitly seek to slow down trading, either via order delays (speed bumps as implemented on Toronto’s TSX Alpha) or frequent batch auctions instead of continuous trading.
In a new research paper (link) joint with Michael Brolley, we propose a FinTech-oriented, market-based solution to the high-frequency arms’ race and, in particular, to the excess exchange infrastructure buffers.
Our proposal relies on a recent development in financial technology, decentralized exchanges (DEX). Decentralized exchanges are currently being implemented in the crypto-asset space: for example, platforms such as Binance DEX, EtherDelta, or IDEX. As opposed to exchange server rooms hosting the infrastructure, on a DEX the limit order book data and the trade matching software are distributed as smart contracts in a peer-to-peer network. Each participant in the network (miner) has a copy of the exchange itself and may “rent out” spare computer power (CPU) to process incoming orders, for a fee which depends on the supply and demand of trading infrastructure.
A DEX would allow for flexible dynamic pricing of trading infrastructure. In some sense, trading speed itself becomes liquid and trade-able. The mechanism is very much like the way Uber implements surge pricing when the demand for taxi rides spikes: When a trading surge occurs, excess demand pushes up the price of computing power, and miners allocate more computer power (CPU) resources to the exchange. In normal market times, miners can re-route idle CPU power towards other, more productive goals.
Our main finding is that a decentralized exchange can improve the allocation of computer infrastructure. In a centralized exchange, high frequency traders (HFTs) have to purchase technology subscriptions such as colocation on a continuous basis, poised to act on short-term opportunities when and if they emerge. By contrast, in decentralized exchanges, HFTs are able to rent technology on-demand: Profitable trading opportunities simultaneously generate “micro-bursts” in trading activity as well as surges in the price of processing power.
A decentralized exchange might actually speed up price discovery. If trading at a centralized market resembles a marathon where runners need to pace themselves, trading at DEX mirrors a sprint where trading speed can surge higher over short intervals. Consequently, the expected time from receiving an informative trading signal to a price update is lower in decentralized markets.
Intense HFT competition for speed during microbursts triggers a surge in the price of computer power. On-demand pricing of low-latency infrastructure achieves two objectives at the same time. High-frequency traders earn lower rents from low-latency trading and, at the same time, the overall resource consumption is lower.
Existing proposals to tackle the costs of low-latency trading focus primarily on removing traders’ incentives to be fast (i.e., frequent batch auctions), or increasing trading latency directly (speed bumps). We propose a market-based solution, recognizing idle exchange infrastructure as a negative externality of fast trading. Decentralized markets do not eliminate high-frequency races, but rather concentrate them within microbursts. The outcome is a shorter idle time for exchange infrastructure, while at the same time limiting excessive rents from low-latency trading.
Link to research paper is here: [link].