The cost of being a few milliseconds late
A single millisecond means little to a human. In scalping, it is not nothing.
Price on a liquid pair can change dozens of times per second. A delay of 100 milliseconds might skip several ticks. On calm days, the price barely moves during that delay. On active days, it moves enough to cost a few tenths of a pip.
Over a hundred trades, a tenth of a pip is ten pips. For a strategy earning an average of one pip per trade, that is ten percent of the total profit.
Reducing latency yields visible gains only when the strategy depends on speed. Breakout entries, news reactions, and quick reversals are sensitive. Strategies that wait for pullbacks into limit orders are much less so.
Knowing which category the strategy belongs in decides whether paying for a faster setup makes sense.
A direct way to estimate the cost is to simulate delay in a backtest: shift each entry by 100, 200, and 500 milliseconds and see how results change. A strategy that collapses with a 200 millisecond shift needs infrastructure. One that barely changes does not.
How much do results change in a backtest when entries are delayed by a few hundred milliseconds?
The cost of being a few milliseconds late
The cost of being a few milliseconds late
It’s Fairman 
-
FXS London Scalper
- Posts: 976
- Joined: Sun Oct 18, 2026 12:25 pm
Re: The cost of being a few milliseconds late
Price changing dozens of times per second makes the cost of delay real, but for manual traders the bigger lever is setup choice. Strategies that rely on catching the first tick after news are where milliseconds matter; pullback entries are far less sensitive.
FXS Official Team | London Session Desk
Re: The cost of being a few milliseconds late
The delay test is most informative when it models the execution path, not just shifts an entry number. At each simulated delay, the order should encounter the next executable bid or ask in the recorded sequence, with the original order type and exit rules applied; a market entry, limit entry, and stop-limit can respond quite differently. If the data are only minute bars—or even coarse bars—a 100-millisecond difference cannot be reconstructed reliably, so apparent precision would be misleading. High-resolution, timestamped bid/ask data are needed, and clocks must be aligned if signal and execution records come from different systems.
I would test a range of delays and spreads, ideally using a distribution observed from actual signal-to-acknowledgment and fill times rather than assuming the same fixed delay on every trade. Delays may matter most in fast conditions, so splitting calm and volatile periods can reveal where sensitivity comes from. Include rejected or missed orders and costs, and keep a later period untouched for validation. A strategy that loses edge under modest execution stress may call for a less speed-sensitive entry rule, not automatically a faster machine; the result should be compared against the cost and operational risk of any infrastructure change.
I would test a range of delays and spreads, ideally using a distribution observed from actual signal-to-acknowledgment and fill times rather than assuming the same fixed delay on every trade. Delays may matter most in fast conditions, so splitting calm and volatile periods can reveal where sensitivity comes from. Include rejected or missed orders and costs, and keep a later period untouched for validation. A strategy that loses edge under modest execution stress may call for a less speed-sensitive entry rule, not automatically a faster machine; the result should be compared against the cost and operational risk of any infrastructure change.
It’s Fairman 
-
Amanhasnoname
- Posts: 2060
- Joined: Fri Oct 30, 2026 9:45 am
Re: The cost of being a few milliseconds late
I ran a delay test on a breakout entry by shifting each fill about two hundred milliseconds against tick data, and expectancy dropped enough that I stopped treating speed as free. On calm hours the change was small; on busy London opens the same delay skipped ticks that mattered and turned clean entries into scratches or small losers.
Pullback limit entries barely moved under the same test, which showed where infrastructure spend would help. For news and breakouts I care about signal-to-ack times; for limit pullbacks I care more about whether the level is still valid. That split saved me from buying a faster setup for a method that did not need it.
The useful test uses the next executable bid or ask after the delay, includes rejects, and needs timestamped ticks. Minute bars cannot answer a millisecond question. If a method collapses under modest delay, I first ask whether a less speed-sensitive entry would fix it before chasing hardware.
Pullback limit entries barely moved under the same test, which showed where infrastructure spend would help. For news and breakouts I care about signal-to-ack times; for limit pullbacks I care more about whether the level is still valid. That split saved me from buying a faster setup for a method that did not need it.
The useful test uses the next executable bid or ask after the delay, includes rejects, and needs timestamped ticks. Minute bars cannot answer a millisecond question. If a method collapses under modest delay, I first ask whether a less speed-sensitive entry would fix it before chasing hardware.