---
title: "How Thresholds Work"
description: "How LENS sets and dynamically adjusts the thresholds that decide when an anomaly is worth an alert."
canonical_url: "https://docs.sealmetrics.com/lens/anomaly-detection/thresholds"
lang: "en"
date_generated: "2026-08-11T17:34:37.681Z"
source_hash: "321fc47f67fcbe8427dc2cc718aa6762b801c2f856b964d7c4cc9b1d597dae07"
content_type: "documentation"
owner: "docs"
llm_priority: "useful"
source_file: "lens/anomaly-detection/thresholds.mdx"
publisher: "Sealmetrics"
---

# How Thresholds Work

Canonical page: https://docs.sealmetrics.com/lens/anomaly-detection/thresholds

**Caution:**
Automated anomaly detection is **built but not active**. No detection rule currently runs on customer accounts and no automated alert is generated. This page documents the intended behaviour; the thresholds below describe how rules will decide once enabled. For what LENS does today, see [LENS AI](/lens).

A threshold is the boundary between a "normal" fluctuation and an anomaly worth surfacing. In LENS, thresholds are **not** a single static percentage you pick from a menu. Each rule starts from a base threshold and then **adjusts it automatically** based on how much data is available and on seasonality.

This page explains that mechanism so you understand why LENS alerts when it does.

## The base threshold

Every rule defines a base threshold (a percentage change). For most rules the default is around 20%, but each rule sets its own value appropriate to the metric it watches. LENS resolves the effective base threshold in this order:

1. A custom value configured for the rule (if any)
2. A global setting for that threshold key (if defined)
3. The rule's built-in default

The resolved base threshold is then adjusted by the two mechanisms below before any alert is raised.

## Adjustment 1: Confidence-based dynamic thresholds

Smaller samples are noisier, so LENS requires a **larger** change before alerting when there's less data. It computes a **confidence level** from the sample size and the number of days of data, then multiplies the base threshold:

| Confidence | When it applies | Threshold multiplier |
|------------|-----------------|----------------------|
| **High** | Large sample, enough history | ×1.0 (no change) |
| **Medium** | Moderate sample/history | ×1.25 (25% more conservative) |
| **Low** | Small sample/short history | ×1.5 (50% more conservative) |

The confidence level itself comes from these data thresholds (the more conservative of the sample-based and days-based result wins):

| Signal | Medium confidence | High confidence |
|--------|-------------------|-----------------|
| Days of data | 7+ days | 14+ days |
| Traffic sample (events) | 100+ | 1,000+ |
| Conversions | 20+ | 100+ |

**Example:** a rule with a 20% base threshold, evaluated on a small sample (low confidence), effectively requires a ≈30% change (20% × 1.5) before it flags an anomaly. The same rule on a large, mature dataset uses the full 20%.

Every insight carries its confidence level (`high`, `medium`, or `low`) so you can weigh it accordingly.

## Adjustment 2: Seasonality

LENS can adapt to recurring, expected variation through seasonality patterns. A pattern can influence detection in three ways:

- **Threshold relaxation** — during a known seasonal period, the base threshold is multiplied by a configurable factor (typically making it more lenient) so ordinary seasonal swings don't trigger alerts.
- **Year-over-year comparison** — when a pattern requests it and the account has the historical data, the baseline shifts to the same period last year instead of the recent past. This compares like-for-like (e.g. this Black Friday vs. last Black Friday).
- **Expected-change suppression** — a pattern can declare an expected change for traffic, conversion, or revenue. If the observed change falls within tolerance of that expectation, LENS treats it as normal and does not alert.

A seasonality pattern can be global or specific to an account, and defines the time window it applies to plus any of the adjustments above.

## How it fits together

For each rule evaluation:

```
base threshold
   → resolve (custom → setting → default)
   → apply seasonality multiplier (if in a seasonal period)
   → apply confidence multiplier (from sample size + days of data)
   = effective threshold

observed change vs. effective threshold
   → if within an "expected for the season" range → suppressed
   → else if beyond the effective threshold → anomaly
```

## Statistical rigor for rate metrics

For proportion-based metrics (such as bounce rate or conversion rate), LENS uses **Wilson score confidence intervals** at 95% confidence rather than naive point comparisons. This avoids over-reacting to rate swings that are just small-sample noise.

## What you control

- **Enabling notifications** — make sure email notifications are on in your account settings; email is the only delivery channel.
- **Seasonality patterns** — configure patterns so recurring events (sales, holidays, campaigns) adjust the baseline and relax thresholds automatically.

The confidence multipliers and data-volume thresholds described above are managed by LENS and apply consistently across accounts; you don't need to tune them manually.

## Next steps

- [How anomaly detection works →](/lens/anomaly-detection)
- [View the planned rule library →](/lens/anomaly-detection/rule-types)