Replicating TRACE: A Practitioner's Guide to Its Threshold and Particle Budget

TL;DR

The study replicates TRACE's synthetic results, revealing that its optimal threshold aligns with the truth margin rather than a fixed value. It shows TRACE primarily recovers immediate causal influences (lag-1), with performance dropping sharply for longer lags due to design choices in the benchmark. F1 scores saturate at just two particles, indicating threshold stability over estimator precision.

Why It Matters

Practitioners must recognize that TRACE's reported performance reflects lag-1 recovery under a skewed benchmark, not general causal discovery. The findings guide better threshold selection and benchmark design, preventing overestimation of algorithmic capability.

Key Findings

• The optimal threshold in TRACE is tied to the truth margin, not a fixed tau, with errors clustering near a delta=0.05 boundary.
• TRACE recovers lag-1 causal edges with high recall (0.97–0.99) but performs poorly on lag-2+ edges, indicating a fundamental limitation in detecting indirect influences.
• F1 scores plateau at N=2 particles, demonstrating that performance depends on threshold margin over noise, not on estimator convergence with more particles.