Stage Drop-Off Waterfall¶
Visualization artifact — instantiates Funnel Attrition Localization
Renders the population cascading from the initial cohort down to final yield one stage at a time, so the size and exact location of every loss is read off a single denominator-preserving chart.
A Stage Drop-Off Waterfall takes an ordered process and draws it as a descending staircase: the full entering population sits at the top, each stage lops off a labelled block equal to the count lost there, and what remains at the bottom is the final yield. Its defining move — the thing that separates it from a row of per-stage bar charts — is that every block begins where the previous one ended, so a single denominator is carried down the whole cascade and each loss is expressed in the same currency: share of the original population. You see not merely that a stage converts poorly but how much of the final shortfall it accounts for, because the blocks literally stack down to the yield you actually got.
Example¶
A platform offering a free multi-week online course wants to understand why so few learners finish. The endpoint is stark — only a small fraction complete — but that lone number says nothing about where people leave. The team lays the process out in order: enrolled → started lesson 1 → reached the midpoint → submitted the final project → completed. Then they draw the waterfall. The top block is everyone who enrolled (100%). The first step-down is enrolled-but-never-started, a large silent block. The next is started-but-quit-before-midpoint. A smaller block falls away before the final project, and a last sliver between submission and completion.
Read as a waterfall, the picture reframes the problem. The largest single block is not in the "hard" middle of the course but right at the top — people who enroll and never open lesson one — and, shown as a share of the original cohort, it is bigger than all the later drops combined. That block is invisible in a completion-rate headline. The chart does not say why they never start and does not rank the fix, but it shows, in one denominator-preserving picture, exactly where the population is going.
How it works¶
- Fix the stage order and the top-of-funnel population, then compute, at each transition, the count that fails to advance.
- Draw each loss as a block whose height is its share of the original population, stacked so each bar continues from the previous remainder — survivors plus all losses so far always sum to the starting cohort.
- Label every block in two readings: as a loss expressed against the starting cohort, and (on annotation) as the local stage conversion rate. The two answer different questions, and holding both against a single baseline is the chart's whole value.
What it deliberately does not do: rank blocks by fixability, split them by segment, or judge whether a loss is normal. It is a faithful picture, not a verdict.
Tuning parameters¶
- Stage granularity — few coarse stages or many fine ones. Finer stages localize a loss more precisely but can slice one real drop-off across several bars and add noise; coarser stages hide where inside a big step the loss actually sits.
- Absolute vs. relative height — blocks scaled to raw count or to share-of-cohort. Absolute makes early, high-denominator losses dominate; relative surfaces a catastrophic late-stage rate that touches few people. Showing both guards against misreading.
- Loss vs. survival framing — draw the shrinking survivor bar, the accumulating lost blocks, or both. The survivor view flatters; the loss view confronts.
- Baseline overlay — whether an expected or prior-period cascade is drawn behind the current one. Without it the chart shows where losses are; with it, where they changed.
- Ordering fidelity — how strictly the stages are treated as a single line. A truly branching process forced into one staircase misplaces losses — a caution the branch-path model, not this chart, resolves.
When it helps, and when it misleads¶
Its strength is turning a lone endpoint into an accountable chain you can point at: the denominator-preserving stack makes hidden early losses visible and forces every drop into comparable units, which is exactly what a raw conversion rate hides.
Its characteristic failure is reading the biggest bar as the biggest opportunity. A tall block early in the funnel is a lot of lost people, but a short block late in the funnel can hide the near-total collapse of a small, high-intent population — absolute height and local conversion rate diverge, and treating the visually dominant block as the priority is a base-rate error.[n1] The chart is also easy to draw backwards — pick the stage boundaries or the denominator that make a favoured culprit look like the villain — and it silently assumes the stages really are a clean, single-file order. The discipline is to read every block in both units (share-of-cohort and local rate), fix the stage definitions before looking at the data, and hand the ranking question to a mechanism built for it rather than eyeballing the tallest bar.
How it implements the components¶
Stage Drop-Off Waterfall realizes the representation side of the archetype — the components that turn an endpoint into a legible, denominator-honest picture:
ordered_stage_model— the chart is the ordered stages, drawn top-to-bottom as a fixed sequence of transitions.denominator_transition_frame— its signature property: each bar continues from the previous remainder, so one denominator (the starting cohort) is carried the whole way down and every loss is expressed against it.per_stage_conversion_and_loss_metrics— each block quantifies the loss at its transition, in both share-of-cohort and local-rate terms.
It does not compile or rank a loss register or estimate recoverable yield (that is Loss Pareto Review), split the cascade by segment (that is Segment Funnel Comparison), watch it against a baseline over time (that is Stage Conversion Anomaly Alert), or check that the population it draws is the true one (that is Survivorship Bias Audit).
Related¶
- Instantiates: Funnel Attrition Localization — the waterfall is the archetype's baseline picture, the shared view the other mechanisms annotate, rank, slice, or audit.
- Sibling mechanisms: Loss Pareto Review · Segment Funnel Comparison · Stage Conversion Anomaly Alert · Survivorship Bias Audit · Cohort Transition Table · Conversion Funnel Dashboard · Denominator Reconciliation Checklist · Event Instrumentation Specification · Event Trace Process Mining · Funnel Experiment Backlog
Editorial Notes¶
Form Classification¶
Form family: Interface, Display & Cue
Rationale: Stage Drop Off Waterfall operates by visually encodes each stage loss as a proportional stacked block to make attrition salient. That concrete deployed or enacted form is Interface, Display & Cue under the frozen taxonomy.
Nearest alternative: Representation, Specification & Plan — Although Representation, Specification & Plan can support this mechanism, the frozen evidence makes its operative form the act that visually encodes each stage loss as a proportional stacked block to make attrition salient; the alternative is therefore secondary rather than defining.
Review outcome: Adjudicated after independent review; medium confidence.
Origin Attribution¶
Primary origin: Data Science & Analytics
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: A denominator-preserving cascade chart is funnel analytics and visualization.
Related originating lineages:
- Human-Computer Interaction — Waterfall form makes loss location legible.
- Statistics & Experimental Design — Stage-specific attrition rates require valid denominators.
Review resolution: The blind reviewers agree that data_science is the primary origin and differ only on alternate origin disagreement, origin mode disagreement, domain reach disagreement, encyclopedia synthesis disagreement. I preserve every independently explained alternate from both records rather than imposing a numeric cap. I retain convergent because the combined evidence shows independent disciplinary development. The broader reach of multi_domain records portability separately from historical provenance; encyclopedia_synthesis=true preserves the affirmative synthesis judgment where either reviewer identified one.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Base-rate neglect — judging by a salient conditional (here, the visual size of a block) while ignoring the underlying population it is drawn from. A 90% drop at a stage only ≈3% of the cohort reaches loses less absolute yield than a 15% drop at the top; the waterfall's absolute bars can invert that intuition, which is why each block should be read against both its denominator and its local rate. ↩