csot '25/low-latency systems track

Same algorithm.
Different code.

The strategy is fixed. You compete on how fast your code executes it. The leaderboard ranks p50, p99, and p999 latency of a single function call, on the same dataset, on the same machine.

Sign in with DevClubWatch the leaderboard20 participants · 7 days remaining
Reference implementation, today
median on_tick
612ns

the typical call, when cache and branch predictor are warm

p99 on_tick
3,104ns

the slow 1% — where most rankings are decided

ticks replayed
16.2M/s

end-to-end throughput on the held-out dataset

How it works

Five steps from sign-in to leaderboard.

  1. 01

    Get the cohort kit

    Download the starter pack and the strategy spec. The strategy is fixed; you just implement it.

  2. 02

    Implement the spec, exactly

    Per-symbol rolling window of 64 mid-prices, z-score |z| ≥ 2.0 entry, |z| ≤ 0.5 exit, one unit of position per symbol. The cohort owns the algorithm. You own the speed.

  3. 03

    Submit your build

    ./submit.sh strategy.so from your terminal, or drag and drop on the dashboard. One submission per participant per ten minutes.

  4. 04

    Correctness, then latency

    Your code is replayed against a held-out 10 M tick dataset. If a single emitted order disagrees with the reference, you do not rank. If you pass, every on_tick(...) call is timed.

  5. 05

    The leaderboard updates

    p50 is 40% of the score. p99 is 40%. Throughput is 15%. Compliance is 5%. Trading P&L is identical across correct submissions, so it is never part of rank.

Top 3 · hidden phase

Currently leading

Full leaderboard →
01
page-fault-tolerant
p50 643 ns·p99 3,352 ns·16.20 M ticks/s
118.2
02
rdtsc-and-chill
p50 659 ns·p99 3,467 ns·16.14 M ticks/s
114.1
03
2024MT60685
p50 696 ns·p99 3,724 ns·16.01 M ticks/s
110.0