Cohort Retention

Classic retention counts an entity as active if it had at least one event in exactly that period; it is what e-commerce and SaaS teams usually mean. Rolling retention counts an entity as active in a period if it had an event then or in any later period; it can only go down, and it is what D7 or D30 retention means in mobile and games. Sequential retention requires activity in both the period and the one before it, which is the right frame for renewal and churn chains.

The visual offers all three and names the active one in its subtitle, not only in the format pane, so a screenshot in a board pack carries its own definition.

Retention is active entities divided by a denominator, and the denominator is where cohort charts quietly disagree. Cohort base is the classic triangle. Previous period gives period-over-period retention. Surviving base follows the sequential rule for renewal chains, and observed base counts only entities that could still have been seen. All four are available in every tier and the chosen one is written on the chart, because both a wrong denominator and an unlabelled one produce a chart that lies.

A cohort acquired four months ago has no month-twelve retention yet. A spreadsheet or a DAX matrix reports that as zero, the average for month twelve collapses, and someone concludes retention has crashed. In this visual, unobservable cells are hatched and excluded from every average and from the colour scale. Periods that have started but not finished are outlined, flagged in the tooltip and left out of averages by default. The most recent cohort, which is still recruiting, is flagged the same way.

Cohorts with too few entities produce meaningless percentages, so a minimum base setting suppresses those cells and shows the base count instead of a rate.

A visual only sees the data the page sends it. Set a date slicer to the last six months and every entity's first event within that window becomes its first event, so the triangle silently becomes something else. No visual can fix this from the inside, so this one makes it visible instead: the header states the derived cohort range, and an optional Cohort Date field lets you bind a model-computed, filter-stable first-event date that always wins.

With a value field mapped, the premium tier turns the triangle into net revenue retention, revenue per active entity and a cumulative lifetime-value curve per cohort. Type in an acquisition cost and the visual marks the period in which each cohort pays it back. Segments compare side by side on one colour scale with significance markers that separate a real gap from noise, a benchmark overlay tints every cell by its distance from your best, median or first cohort, and the age-versus-calendar panel separates a weak cohort from a bad month for everyone.

Beyond the observed window, a shifted-beta-geometric projection, validated against the published Fader and Hardie reference, draws where retention is heading. The whole engine is tested against independent pandas and scipy implementations, and the test record ships with the visual.