Michael Brolley

Liquid speed:

a micro-burst fee for low-latency exchanges

Marius Zoican

The cost of latency arbitrage

U.K.-based evidence (Aquilina, Budish, O'Neill, 2020) shows that latency arbitrage races are:

• Frequent: 537 races/symbol-day in FTSE100 stocks.
• Fast: winner speed margin of 5-10 microseconds.
• Concentrated: top 6 firms win 87% of races.
• Costly: A daily tax of 0.42 bps of total trading volume.
• Harmful for markets: Reduce liquidity for investors by 17%.

Existing market design solutions

• Focus on eliminating HFT races and improving liquidity.
• Not concerned with exchange revenues & incentive compatibility
• Sometimes at odds with current fragmented environment.

Micro-burst fees

Our solution: A micro-burst (congestion) fee

• Main idea: Surge pricing of speed (à la Uber).
• How? Implement a micro-burst fee on liquidity taking orders.
• Fee per message within sub-millisecond clusters.
• Fee level     # liquidity taking orders, surge during micro-bursts.
• Market price of speed: latency arbitrage is priced in.
• Mechanism produces revenue for exchanges.

A world with co-location fees

A world with no latency arbitrage

(e.g., idealized batch auctions)

Our proposal: A world with micro-burst fees

Micro-burst fee design:

Two parameters:

• A fee scale parameter
• The number of messages in the cluster

• Cluster size of 100 microseconds to 10 milliseconds).
• A fee of zero if a single marketable order in a cluster.
• Fees are computed intraday, but charged at low frequencies.
• Unlikely to accidentally pay fees: Daily race duration is 0.05s.

Implementation: Inbound firewall timestamps

Implementation: Inbound firewall timestamps

Implementation: Criteria for a microburst

• Multiple (2+) aggressive orders on the same security, side, and price level(s).
  • Require unique UserID (different trading desks)
  • LO or IOC is aggressive if buy at or above best ask;
  • ------------------------------------ sell at or below best bid;
• Some aggressive orders succeed and some fail.
  • IOC fails if matching engine sends "expired" message;
  • LO fails if obtains zero quantity at price p.
• Orders arrive at the same time.

What does "at the same time" mean?

M1 and M2 arrive at same time if:

M2 Inbound < M1 Inbound + AEL + MRT

• Values can be re-calibrated
• A cap for (AEL+MRT) can be set, e.g., 500 micros.

Game-theoretic equilibrium

Two important outcomes:

• Liquidity (    ):  the outstanding half-spread
• Race intensity (    ): probability of an HFT to race on a short-lived signal.

Impact of the micro-burst/congestion fee

Exchange revenues

Exchange revenues from micro-burst fees

How does the fee magnitude impact revenues?

• Direct effect:          implies higher revenues for each trade;
• Indirect effect:          reduces HFT incentives to snipe         less intense races.

How to optimally set     ?

Co-location fees lead to capped revenues

• Co-location fees do not scale with trading activity and

• they need to be paid in advance on all exchanges.

Micro-burst fee revenue > Co-location fee revenue

Precise fee shape is irrelevant for revenue

Competition and regulation

A fragmented exchange environment

• HFTs pay co-location fees on all exchanges:
  Profits are not competed away.
• With micro-burst fees, profits may be competed away since each exchange tries to capture the entire revenue.
• Solution: a (lax) regulatory cap on fees.

What is the incentive compatible fee cap?

Calibration

We calibrate caps on a constant fee structure:

• The revenue-maximizing fee is $0.34 per message.
• The regulatory cap is $3.98 per message.
• Any fee above 7.8 bps per share ($0.078/message):
  • generates higher liquidity than co-location fees
  • provides higher revenues than co-location fees.

Closest implementation:

IEX Crumbling Quote Remove Fee

• A fee on liquidity taking order when CQI is on
  (i.e., IEX predicts an HFT race).
   
• Micro-burst fees do not depend on a signal.
  • Cannot be "gamed" by exploiting signal flaws.
  • Are market-driven and more precise than a third party's assessment of races.

HFT entry and micro-burst fee caps

How do micro-burst fee compare?