For multi-location operations + analytics leadership
Outlier alerts that compare each location to its real peers
Not the company-wide average that buries the individual signals. Per-location two-sigma outlier flagging against the peer cohort, seasonality-adjusted, multi-metric, and wired into the anomaly-detection pipeline downstream.
What this gets you
- Per-location outlier detection — two-sigma against the peer cohort, not the company-wide average.
- Cohort-aware statistical signal — gym Sun-Belt peer set treated separately from gym Northeast peer set. Same brand, different baselines.
- Seasonality-adjusted detection — Saturday vs Tuesday, peak vs off-peak, Thanksgiving vs any other Thursday all handled per cohort.
- Multi-metric outlier scoring — revenue, foot-traffic, conversion, NPS, and review-rate combined into a single per-location anomaly score.
- Wired into the anomaly-detection pipeline — Flag-Outlier signals connect to Correlate, Route, Suppress, Subscribe, and Dedupe downstream so the alert stream is actionable, not noisy.
The brand-wide average buries 49 location signals
Most multi-location operators run outlier detection at the brand level. “Revenue this week is down 6% versus last week” surfaces when the company-wide aggregate moves. It misses the 50 individual location signals that compose the aggregate. A single location dropping 30% gets buried in the 49 holding steady. The brand average drifts a few points, the dashboard stays green, and the location keeps bleeding until the brand-level signal finally builds three or four weeks later.
Per-location outlier flagging surfaces every individual location anomaly — but only when the comparison is to that location’s peer cohort, not the company-wide average. A new gym in Phoenix compared to six other new gyms in Sun-Belt markets with comparable demographics produces a meaningful outlier signal. The same gym compared to the company average is noise in either direction.
Two-sigma flagging at the peer-cohort level produces a different alert stream. Every location is compared to its own peer cohort. Seasonality is adjusted per cohort. Multi-metric scoring combines revenue plus foot-traffic plus conversion plus NPS plus review-rate into a single per-location anomaly signal. The output is a feed of location-level outliers that connects to the downstream anomaly-detection pipeline — correlate, route, suppress, subscribe, dedupe.
For a 50-location operator with weekly cohort comparison, per-location two-sigma flagging surfaces three-to-five times more actionable signals than brand-level flagging. Real performance variance lives at the location level. The brand average is where it disappears.
What is in market — and what each category leaves to you
The statistical primitives and the alert-routing infrastructure are mature. The per-location peer-cohort definition and the multi-metric business-signal layer are operator-side wiring.
Business anomaly detection — Anodot, Outlier.ai
Built to detect anomalies in business metrics rather than IT operations. Strong on revenue, conversion, and funnel-level signal. The per-location peer-cohort layer and the multi-location seasonality model are typically assembled on top.
AIOps platforms — Datadog AIOps, Splunk, New Relic, Dynatrace, AppDynamics, PagerDuty
Designed for IT operations metrics — CPU, memory, latency, error rates. The statistical methods transfer cleanly to business metrics, but the per-location and per-cohort framing is operator-side. These tools fire when CPU spikes; they do not natively fire when one location underperforms its cohort.
BI-embedded anomaly — ThoughtSpot SpotIQ, Sisense
Anomaly detection built into the BI tool. Helpful as a diagnostic surface during dashboard review; not designed as a per-location alert stream.
Manual exception reports
Most multi-location operators sit here: an analyst runs a weekly exception query, surfaces the top-five underperforming locations, and emails the operations director. Useful, manual, fragile, and always running one cycle behind the actual signal.
The pipeline, end to end
- Peer-cohort computation. Per-location peer set defined by region, vertical, scale, demographics, market age, and brand banner. Cohorts refresh on a defined cadence so they stay aligned with the operating reality.
- Per-cohort baseline distribution. Mean, variance, and seasonality model per metric per cohort — computed against a rolling historical window. Each cohort gets its own baseline, not a copy of the brand baseline.
- Two-sigma threshold per metric per cohort. Threshold sits at the cohort’s two-sigma boundary and adapts when cohort variance shifts. Fixed thresholds fail on cohorts with naturally high variance.
- Multi-metric outlier scoring. Revenue, foot-traffic, conversion, NPS, and review-rate combined per location into a single anomaly score. A location that is two-sigma low on three of five metrics is a louder signal than one that is two-sigma low on one of five.
- Seasonality-adjusted detection.Saturday vs Tuesday, peak vs shoulder vs off-peak, holiday calendar per market — all handled inside the cohort’s seasonality model. Detection runs against the seasonally-adjusted baseline, not the raw mean.
- Point vs contextual anomaly classifier. Point anomalies are observations outside the baseline. Contextual anomalies look normal in isolation but are anomalous given context. Both run; both feed the alert stream.
- Per-location alert stream. Outliers grouped per location (not per metric) so the recipient sees the location story, not five disconnected metric stories. Each alert carries the cohort context, the baseline reference, and the failing metrics.
- Connection to anomaly-detection pipeline. Flag-Outlier signals feed Detect, Correlate, Route, Suppress, Subscribe, and Dedupe downstream. The output is a curated alert stream, not raw outliers.
- False-positive suppression. Outliers that are repeatedly noise — a location whose two-sigma trips without operating cause — get suppressed via the same suppression layer used by the anomaly-detection agent.
- Minimum cohort-size enforcement. Cohorts smaller than the minimum yield unreliable variance estimates. The pipeline blocks two-sigma detection on under-sized cohorts and falls back to absolute thresholds while the cohort grows.
- Cohort drift detection.Alerts when a location’s peer cohort composition shifts — new locations join, old ones close, demographics change. Cohort drift invalidates the baseline; the alert prompts a refresh.
- Operator dashboard. Per-location outlier feed, cohort-level outlier rates, seasonality-adjustment quality, false-positive rate, suppression-rule status — one view across the pipeline.
Frequently asked
What is anomaly detection software?
Anomaly detection software identifies data points that deviate from an expected pattern. Most commercial anomaly-detection tools (Datadog AIOps, Anodot, Outlier.ai, Splunk, New Relic, Dynatrace, AppDynamics) were built for IT operations metrics — CPU, memory, latency. The same statistical primitives apply to multi-location business metrics, but the per-location and per-cohort wiring is rarely shipped in the product.
How does two-sigma outlier detection work?
Two-sigma outlier detection flags any observation that falls more than two standard deviations from the mean of a baseline distribution. Roughly 95% of normal observations fall inside the two-sigma band; observations outside it are statistically unusual. The math is straightforward; the work is choosing the right baseline — and at multi-location scale, the right baseline is the per-location peer cohort, not the company-wide average.
Why does per-location detection beat brand-level detection?
Brand-level detection asks "is the company-wide aggregate moving?" Per-location detection asks "is any individual location moving against its peers?" At a 50-location operator, one location dropping 30% gets buried in the 49 holding steady — the brand aggregate barely moves. Per-location flagging surfaces the individual signal in week one instead of week three or four when the brand average finally drifts.
How is this different from Anodot, Outlier.ai, Datadog AIOps, Splunk, or Dynatrace?
Those platforms own the statistical primitives, the streaming infrastructure, and the alert routing. They are excellent at IT-operations anomaly detection. The per-location peer-cohort definition, the seasonality model per cohort, the multi-metric scoring across revenue plus foot-traffic plus conversion plus NPS plus review-rate, and the integration with the downstream anomaly-detection pipeline are operator-side wiring.
What is a peer cohort?
A peer cohort is the set of similar locations that a target location is compared against — similar by region, vertical, scale, demographics, age, and market type. A new gym in Phoenix compared to its peer cohort (six other new gyms in Sun Belt markets with comparable demographics) produces a meaningful outlier signal. The same gym compared to the company-wide average is noise either way.
What is the difference between point anomaly and contextual anomaly?
A point anomaly is a single observation outside the baseline distribution — Friday revenue at one location is 3 sigma below cohort. A contextual anomaly is an observation that looks normal in isolation but is anomalous given context — Thanksgiving Friday at the same location is 3 sigma above its Black-Friday baseline despite being well within the year-round revenue band. Multi-location operators need both detectors running, not just point.
Hire the agent that runs the flagging
The location-benchmarking agent owns peer-cohort computation, per-cohort baseline modeling, two-sigma flagging, multi-metric scoring, and the wire into the downstream anomaly-detection pipeline.
We scope on the call and send a private checkout link after.
Related reading: Multi-location reporting · Hyper-local search trends