Adverse drift analysis on live crypto_shortterm data to establish the publish→fill latency budget for the Phase-1 emit path. Derived from crates/analysis/src/bin/latency_budget.rs, commit 2f5fb4d.

Context

The crypto-shortterm-phase0-depthaware analysis confirmed the edge survives realistic depth-capped fills and established the remaining open risk as latency survival: does the quoted top-of-book price at signal time still exist by the time the fill arrives? This analysis quantifies how long you have before adverse drift wipes the 0.5¢ latency buffer.

Candidate Regime

  • τ ∈ [60, 150)s
  • |p_cal − 0.5| ≥ 0.30
  • Same 20-bin calibration as the depthaware run

Data

  • 823k total rows in crypto_shortterm
  • 3,162 outcomes
  • 143,228 consecutive observation pairs analyzed (adjacent rows within the same window)

Key Findings

Ask-Price Movement at 1s Cadence

29.21% of 1s steps already exceed a 0.5¢ absolute move in ask price. Sub-second latency is not a nice-to-have — over a quarter of ticks cross the budget threshold in a single polling interval.

Adverse Drift Distribution (ask rising only, ¢/sec)

PercentileDrift (¢/sec)
median (p50)0.97
p751.95
p903.91
p955.84

Implied Latency Budget (0.5¢ / drift_rate)

Drift percentile usedImplied budget
median (0.97 ¢/s)~513ms
p90 (3.91 ¢/s)~128ms
p95 (5.84 ¢/s)~86ms

Recommendation

Publish→fill target: 128ms (p90 of adverse drift), treated as an upper bound. At p90 drift the 0.5¢ buffer is consumed in 128ms; anything slower is gambling on a favorable tail.

Strongly Recommended

Target <50ms end-to-end. The p90 budget of 128ms leaves no margin for network jitter, NATS publish latency, or pmv2 order routing. The p95 budget is 86ms.

Sensitivity Verdict

HIGH. 29.2% of 1s steps already breach the 0.5¢ threshold. Sub-second latency is mandatory; the system has no slack at 1s poll cadence.

Important Caveat

Poll Cadence Limitation

The observe binary polls at ~1s cadence. Intra-candle drift faster than 1s is not captured — the true sub-second drift distribution is unobservable from this data. The real latency budget may be tighter than these numbers suggest. The 29.2% figure and the drift percentiles are lower bounds on realized adverse movement.

Files

  • crates/analysis/src/bin/latency_budget.rs — analysis binary
  • docs/superpowers/phase1-latency-budget-report.md — full report
  • Commit: 2f5fb4d (feat(analysis): latency budget — publish→fill target from adverse drift in candidate regime)