MarketGraphsThe Lens · MethodologyAnomaly (peer-percentile)
← all engines

Anomaly (peer-percentile)

What's not behaving like its peers?
ExperimentalVerdict · EXPERIMENTALRecall · 0.41
Alternative view · behind a stricter gate · anchor metric: peer-percentile divergence
What it measures

Flags instruments that have broken from their cohort — ranks each by where it sits in the percentile distribution of its peer group. The peer-percentile axis resurrected the anomaly read from a near-dead 0.05 to 0.41 (experimental). It replaced the original macro-residual anomaly engine, which was retired after its OLS basket leaked the very divergences it was meant to catch.

\text{for each instrument } i \text{ in peer group } G:
\quad \text{pct}_i = \operatorname{percentile-rank}(\text{metric}_i \text{ within } G)
\text{flag if } \text{pct}_i \text{ in tails } (\ll 0.10 \text{ or } \gg 0.90)
\text{rank by } |\text{pct}_i - 0.50|

A reading in its cohort's far tail means it's the outlier — moving against the group, not against itself. Distinct from Stretched (own history) and Breakout (own regime): this is relative-to-peers.

Limitations
  • Experimental gate — composer treats it as ALTERNATIVE_VIEW behind a stricter gate until it clears 0.50 sustained.
  • Cohort definition sensitivity — the read depends on how the peer group is drawn.
In the product

Primary read of the Anomalies chapter ("What's not behaving like its peers?"). The hybrid variant runs alongside as a second method.

Grading & stability
Gate rate
loading…
Stability / churn
not emitted
Mean confidence
loading…
fetching /api/panel/about/lens_engine.anomaly_peer_pct/W/all …

Graded on the Tier-1 regime-test harness — curated macro events × 18-yr backfill × recall@K. Verdict EXPERIMENTAL (gate > 0.50 to ship).

Charts it reads
Relative rotation graph
rrg_scatter
Sector scorecard
sector_scorecard
Equal vs cap weight
equal_vs_cap_weight
Anomaly (peer-percentile) · The Lens methodology← all engines