Advanced analysis: trend, change points, anomalies
The advanced analyses apply statistics to a property’s daily history. This page explains decomposition, change points and anomalies: how each is computed, how much data it needs and how to read the flags without over-claiming.
✓ Checked in the workspace code · 7 October 2026
What the advanced analyses are
The Advanced group applies statistical methods to the daily history of a property. It includes time-series decomposition, change-point detection, anomaly detection, CTR curve modeling, opportunity scoring, cannibalization entropy, a full on-page check and a link crawl. The hub labels the analyses Beta. This page covers the three that work on daily clicks and impressions: decomposition, change points and anomalies.
These methods point at when something changed. They do not say why; the cause needs a look at releases, Google updates and the pages themselves.
Time-series decomposition
Decomposition splits daily clicks or impressions into three parts so you can see each without the others.
| Part | How it is computed |
|---|---|
| Trend | a 7-day moving average |
| Weekly pattern | the average difference from the trend for each weekday |
| Residual | what is left: value minus trend minus weekly pattern |
The direction label comes from the slope of the trend relative to the average level: growing above 0.5 percent a day, declining below −0.5 percent, otherwise stable. At least 14 days of data are needed.
Change-point detection
The method looks for lasting shifts in the average level, such as a jump after a migration or a drop after an update. It smooths the series with a 3-day average, then compares the mean of a window before each day with the mean of an equal window after it. The window is one eighth of the period, at least 3 days. A day is flagged when the difference is larger than the sensitivity (1.5 by default) times the standard deviation of the series, and flagged days are kept apart by at least one window. Each point reports the raw average before and after and the percentage change.
A higher sensitivity finds fewer, larger shifts. At least 14 days are needed.
Anomaly detection
Anomalies are single days that differ from what that weekday normally looks like. Each day is compared with the median of the same weekday, and the distance is measured in a robust way, with the median absolute deviation instead of the standard deviation, so one extreme day does not distort the baseline. A weekday needs at least four observations; otherwise the overall median is used. A day is flagged when its score passes the sensitivity (2.5 by default). At least 21 days are needed.
Use anomalies to find days to investigate, such as a tracking gap, an outage or a one-day spike, rather than to explain a trend.
Which analysis answers which question
| Question | Use |
|---|---|
| Is traffic growing or falling once weekdays are removed? | Time-series decomposition |
| When did the level change for good? | Change-point detection |
| Which single days were unusual? | Anomaly detection |
How to read the results safely
- Use a long enough window. More days make baselines steadier.
- Treat flags as leads. Check annotations, releases and Google updates for the date.
- Start with default sensitivity, then raise it if you get too many flags.
- Do not use the newest days. They are preliminary and can change.
- Record what you find as an annotation so the next review starts from it.
Advanced analysis, answered
How much data do I need?
At least 14 days for decomposition and change points, and 21 days for anomalies.
Does a change point prove a cause?
No. It shows when the level moved, not why.
What does sensitivity do?
A higher value flags fewer, stronger events.
Continue
Next, see forecasts and experiments and dates and data delay.