Anomaly markers on charts that mean more than a busy Tuesday
An anomaly marker is a mark drawn on a point that falls outside a stated expected range. It earns its place only when the baseline accounts for the weekday and edition pattern the reader already knows about, when the tooltip carries the expected value, and when the marker count on a normal edition stays in single figures.
The pre-show registration chart had markers on it. Thirty-nine of them, across ninety days, each a small orange diamond sitting on the line. The analyst who built it was pleased, because the tool had added them automatically and they made the chart look attentive.
By week three nobody in the campaign meeting looked at them. That is the whole failure mode of anomaly markers on charts, and it happens fast. A marker is a promise that a point deserves attention, and a reader who checks four promises and finds nothing behind them stops checking.
Why does a marker on every spike stop working?
Because the markers were describing the calendar rather than the campaign.
Take the ninety day window and a weekday profile that most B2B shows would recognise. Monday averages 175 registrations, Tuesday 240, Wednesday 210, Thursday 195, Friday 150, Saturday 85 and Sunday 65. Those average out to 160 a day across the week, which is also the mean of the whole ninety days.
Now compute the spread of that series. The seven profile values sit 15, 80, 50, 35, minus 10, minus 75 and minus 95 away from the mean of 160. Square those, average them, take the root, and the standard deviation is about 60 registrations, almost all of it produced by the weekday pattern rather than by anything happening in the market.
Put a band at one standard deviation either side of the mean, which is 100 to 220, and watch what fires. Every Tuesday, at 240. Every Saturday, at 85. Every Sunday, at 65. There are roughly thirteen of each in ninety days, so thirty-nine markers, none of which describes an event. Widen the band to two standard deviations and it runs from 40 to 280, at which point nothing fires at all, including the day your registration form was broken for four hours.
The band cannot be fixed by changing its width, because the variation it is measuring is structure rather than noise.
Stephen Few's framing in Information Dashboard Design, published in 2006, is the right one to hold. Building on Tufte's data-ink ratio, Few argues for reducing the non-data pixels and enhancing the data pixels, and he applies it even to elements that serve a real function, noting that things which do not display data should not be given prominence. Thirty-nine markers are non-data pixels wearing a data costume. They occupy the most visually salient position on the chart and carry no information at all.
Setting the baseline from the same weekday
The construction that works is to give every day its own expected value, taken from the same weekday at the same distance from show open in prior editions.
Day minus 60 falling on a Tuesday is compared against day minus 60 in the last three editions, each of which was also a Tuesday if your show moves with the week, or against the nearest equivalent weekday if it does not. The expected value for that day is then 240 rather than 160, and the question becomes whether today's count is unusual for a Tuesday sixty days out.
Once the structure is removed, the residual is small and roughly even across the week. Say it has a standard deviation of 18 registrations. A band of plus or minus two residual standard deviations is now plus or minus 36, which is a real statement about a Tuesday instead of a statement about Tuesdays in general.
How the residual is produced is a modelling question rather than a chart question, and there is more than one defensible route. Seasonal decomposition is C33's method and building the band from prior editions is V32's. What matters for the marker layer is that the expected value exists as a number you can print, because everything in the next two sections depends on having it.
Power BI's own implementation works this way and exposes the vocabulary. Microsoft describes the feature as enriching the chart with anomalies "and the expected range of values", and states that "when a value goes outside this expected boundary, the process marks it as an anomaly" (Microsoft Learn, 2026). The sensitivity control moves the boundary: increase it and "even a slight deviation is marked as an anomaly", decrease it and the algorithm "is more selective".
Three markers on a ninety day curve
Choose the band width by the marker count you want rather than by convention.
If the residuals are roughly normal, a band at two standard deviations leaves about 4.6 per cent of days outside, which on ninety days is four markers. Tighten to about 2.1 standard deviations and roughly 3.3 per cent fall outside, giving three. That is the target: three markers on an edition where nothing happened, which leaves room for a fourth and fifth to mean something.
Here is what the three looked like on the show in question.
Day minus 56, a Tuesday, 402 registrations against an expected 240. The deviation is 162, which is nine residual standard deviations. This was the early bird deadline.
Day minus 29, a Thursday, 96 against an expected 195. A shortfall of 99, or five and a half residual standard deviations. The registration form had failed for four hours that morning and nobody noticed until the next day.
Day minus 12, a Monday, 388 against an expected 175. A deviation of 213, close to twelve residual standard deviations. A media partner had sent a solus email.
Two of those three were already known to the team before the chart drew them. The early bird deadline is on the campaign plan and the partner send was booked in advance. Marking a spike that everybody expected is a smaller version of the same problem: the marker teaches the reader that markers are often uninteresting.
So annotate the known calendar separately. Deadlines, campaign sends, public holidays and price tier changes go on the chart as light vertical reference lines with labels, drawn in a weaker visual weight than the data. The marker layer is then reserved for points that fall outside the band on days with no annotation, and on this show that leaves exactly one marker, on the morning the form broke, which is the one worth a phone call.
What belongs in the tooltip
A marker that says a point is unusual has asked a question. The tooltip is where it answers.
Four things belong in it. The actual value. The expected value. The expected range. The size of the deviation, in whatever unit the reader thinks in, which for registrations is people rather than standard deviations.
Microsoft's anomaly explanations follow this shape, reporting the observed value alongside a stated expected range for the date in question, and then ranking possible contributing dimensions by a strength figure that the documentation defines as the ratio of the deviation from the expected value when filtered by a dimension to the deviation in the total. That ranking is a useful pointer and a poor conclusion, for the same reason a ranked list of factors anywhere is: it is an association on one day's data.
Keep the marker itself plain. A small hollow circle at the data point, in the same colour as the line, is enough, and it does not compete with the series the way a filled orange diamond does. Encode the size of the deviation in the tooltip rather than in the marker's size, because marker area is a magnitude channel readers judge poorly and three markers do not need a scale.
How many markers is too many?
Five, on one chart, is where I would stop and re-examine the band rather than the show.
That is a rule about the marker layer rather than about detection. Genuinely unusual things do cluster, and an edition where the venue changed will produce a run of real anomalies. The point is that a chart carrying nine or more markers has stopped functioning as a signal and become a texture, and the reader's response will be to ignore all of them rather than to triage.
The multiplication problem is worth knowing about before you build. Microsoft documents that anomaly detection "isn't supported with legends, multiple values, or secondary values in a line chart visual", so a per channel view means one chart per channel. Six channels each firing three markers puts eighteen markers on a page, and the page-level count is what the reader experiences. Budget across the page, not the chart.
Two other documented limits shape the design. Anomaly detection "requires at least four data points", which rules out an edition-level series for a show with three editions of history. And "drilling down to go to the next level in the hierarchy isn't supported", so a reader who sees a marker on the total cannot expand it in place and needs a route to the detail built for them.
Deciding who hears about a marker, and how often, is a different discipline again, and V31 covers what happens when everybody gets subscribed to everything.
Where this stops
A marker says a point is statistically unusual. It never says the point matters, and those two come apart constantly in event data.
A Thursday running 40 registrations light is a five sigma event on a quiet show and completely irrelevant to whether the hall fills. A 2 per cent shortfall in the last fortnight, spread evenly across every day, produces no markers at all and is the thing that will actually cost you. Point anomalies and drift are different questions, and a marker layer only answers the first.
The second limit is that the band is built from history, so the first edition after any structural change is undefined. Move the show three weeks later, change the registration open date, or switch registration platforms, and every prior edition's weekday profile is describing a different campaign. In that year the honest choice is to turn the markers off and read the curve, rather than to leave a detector running against a baseline that no longer applies.
Count the markers on your current pre-show chart. If there are more than five, take the campaign calendar, mark the deadlines and the send dates as annotations, and rebuild the band per weekday from your last three editions. It is an afternoon of work, and it converts a decoration into something the campaign meeting will look at. The same test applies to every other page in the BI and reporting set that draws attention to itself automatically.
Questions people ask about anomaly markers on charts
- How many anomaly markers should a chart show?
- Few enough that each one is worth a look, which for a ninety day pre-show curve means about three on an edition where nothing unusual happened. Choose the band width to produce that count rather than choosing a conventional sigma level, and treat a chart with more than five markers as evidence that the band is set wrong.
- What should an anomaly marker tooltip contain?
- The actual value, the expected value, the expected range and the size of the deviation, so a reader can judge the marker without leaving the chart. Power BI's anomaly explanations name the expected range alongside the observed figure, which is the pattern to copy. A marker showing only that a point is unusual asks a question and answers nothing.
- Why do anomaly markers fire on ordinary weekdays?
- Because the baseline was computed across all days rather than per weekday. Registration volume on a Tuesday can be several times a Sunday, so a single band around the overall mean puts every Tuesday near the top and every weekend near the bottom. Comparing each day against the same weekday in prior editions removes the effect.