How to use the latency advantage calculator
This is a hypothetical model for thinking about speed, not a measurement of any market. Fill in four groups of inputs:
- The opportunity. How many races a day you contest, the average edge in basis points if you win and the trade size. The edge in dollars is trade size × edge ÷ 10,000.
- Costs per race. Fees and hedging costs on a winning trade, and the cost of a lost race. A canceled or rejected order costs almost nothing; being filled after the price has moved, or having to hedge a leg that did fill, costs more.
- Latencies. Your average latency and its jitter, and the mean and standard deviation of the fastest competitor’s latency, in microseconds. Measure them the same way: from the triggering market event to the order reaching the matching engine.
- Fixed costs. Monthly infrastructure (colocation, links, data, hardware) and the number of trading days a year.
The formulas
Net P&L per day = N × P&L per race − Infrastructure × 12 ÷ DE is the edge in dollars, C the cost of a winning trade, K the cost of a lost race, N races per day and D trading days per year.
Break-even latency = μc − Φ−1(P*) × √(σc2 + σy2)P* is the win rate at which net P&L is zero. Value of 1 μs = N × (E − C + K) × φ(z) ÷ σ × D, the slope of annual P&L at your current latency, where φ is the normal density.
A worked example
A strategy contests 2,000 races a day. Each is worth 1 bp on a $100,000 trade, $10.00, and costs $3 to execute; a lost race costs $0.50. Its latency averages 50 μs with 3 μs of jitter, against a rival averaging 55 μs with a standard deviation of 8 μs.
The 5 μs head start gives a win probability of 72.1%, about 1,442 races a day. Each race is worth $4.91 on average, so trading earns $9,812 a day before infrastructure. After the $40,000 monthly bill, $1,905 a day, the expected net is $7,907 a day, or $1,992,606 a year.
The setup stays profitable as long as the win rate is above 19.4%, which the model reaches at a break-even latency of 62.4 μs. Each microsecond shaved off at the current speed is worth roughly $148,722 a year.
Speed only pays relative to the competition
Latency races are winner-take-all, so the absolute number matters less than the gap to the next-fastest trader. Moving the example’s latency while everything else stays fixed:
| Your latency | Win probability | Net per year | Value of 1 μs |
|---|---|---|---|
| 40 μs | 96.0% | $2,898,399 | $37,796 |
| 50 μs | 72.1% | $1,992,606 | $148,722 |
| 55 μs (equal) | 50.0% | $1,158,000 | $176,498 |
| 60 μs | 27.9% | $323,394 | $148,722 |
Two lessons follow. First, a few microseconds swing results from comfortably profitable to marginal, because the win probability moves along the steep middle of an S-curve. Second, once you are clearly ahead, more speed is worth little: at 40 μs you already win almost every race, and the next microsecond buys $37,796 a year instead of $176,498 for a trader who is level with the rival.
Jitter matters as much as the mean. If both sides become far more consistent, with a rival standard deviation of 2 μs and your own jitter at 1 μs, the same 5 μs advantage wins 98.7% of races. Predictable latency turns a small average edge into near-certain wins, which is why firms spend as much effort on removing variation as on reducing the average.
The arms race and the costs behind it
Economists Eric Budish, Peter Cramton and John Shim described this dynamic in 2015 as an arms race: in a continuous order book, stale quotes create mechanical arbitrage opportunities, firms invest heavily to be first, and faster technology does not make the opportunities go away; it only raises the speed needed to win them. They proposed frequent batch auctions as an alternative market design. Whatever the policy view, the economics in this model reflect the same point: the edge per trade is fixed by the market, so spending on speed only pays if it wins races that would otherwise be lost.
Costs scale differently from revenue. Doubling the infrastructure bill to $80,000 a month cuts annual P&L from $1,992,606 to $1,512,606, while halving the edge to 0.5 basis points cuts it to $176,202. Strategies with thin edges are fragile to both, and a rival’s upgrade can turn a profitable setup into a loss without any change on your side.
What this model does not include
- Several competitors. The model races you against a single fastest rival. With many independent rivals, your win probability is lower than shown; treat the rival’s distribution as the fastest of the group.
- Correlated latencies. Market-wide bursts slow everyone down together, which the independence assumption ignores.
- Non-normal tails. Real latency distributions have long right tails from queuing and packet loss.
- Changing opportunity sets. The number and size of opportunities are fixed; in practice they vary with volatility and fall as more firms compete.
- Queue position and partial fills. Winning is all-or-nothing here. The HFT transaction cost calculator covers fees, rebates and adverse selection in more detail.