Clusters & Key Driver Analysis | ATX AnalyticsTable Handbook

Clusters, key drivers and insight questionsWhich kinds of customers do we have, what drives the number, and what changed — answered by statistics that run on your own machine, never by a guess. Most dashboards answer "how much". Three questions come straight after it, and they are the ones a board actually asks: what kinds of customers, products or counterparties do we have; what drives the number up or down; and is anything unusual, where is it heading, and how did we do against last month or last year. ATX AnalyticsTable answers them with statistics rather than with a model's impression. A cluster chart groups the rows that resemble each other across several numeric columns — spend and visits, P&L and volume, exposure and turnover — using k-means, chooses the number of clusters by how well separated they are, and profiles each one. A key-driver chart takes a target — revenue, margin, delay or a yes/no outcome such as churn — and up to twelve candidate columns, numbers or categories alike, and ranks them by how much of what they explain together each one contributes, with its direction, its significance and the segments that stand out. Both are ordinary chart types: they sit on a sheet or a dashboard, cross-filter the board when clicked, export as PNG, SVG and CSV and print in PDF reports. Ask Greg uses the same algorithms — so the segments it describes are exactly the ones its proposed chart draws — and adds three tools of its own for anomalies, forecasts and period comparison. Every one of them is deterministic, runs on your computer or server through the same row and column rules as any query, and sends a model only the aggregates the data policy allows.

What it does

Clusters as a chart type — Reports → Charts → Clusters on a sheet, Dashboard → Add Widget → Clusters on a board, or Clusters as the type of any existing chart, with the same sources, formatting, palette, exports and PDF printing as every chart
One to twelve numeric features, an optional label column saying what one point is, and "Combine rows per label" to segment customers by their average basket rather than by individual visits — the combining done in the database
Number of clusters chosen automatically (every k from 2 to 8, keeping the best-separated by silhouette and preferring the smaller k on a near tie), fixed, or read from a project parameter that can point at a cell
Scaling by z-score (default), robust median and IQR, min–max or none, so a column in millions does not outweigh one in percent
A scatter on the two principal components — the axis titles say how much variation each carries — or on two chosen features, with cluster centres and far-off members drawn as rings; and a profile with each cluster's size, share, feature means coloured against the overall mean, and the features it stands out by
Quality printed under every cluster chart: k and how it was chosen, silhouette with a strong / reasonable / weak grade, points, rows left out, outliers, Davies-Bouldin and the seed
Key drivers as a chart type — Reports → Charts → Key drivers or Dashboard → Add Widget → Key drivers — with a target and up to twelve drivers, numbers and categories alike; selecting a table first binds an outcome column such as revenue or churn as the target
Yes/no targets recognised automatically and analysed as a rate, with the event value typed by hand when needed
Each driver ranked by its Shapley share of R² — exact over every combination, so shares add up to 100 % of what the drivers explain, do not depend on the order they were listed in, and correlated drivers share credit — coloured by direction, faded when not significant at 5 %
Segments that stand out: category levels and the quartiles of every numeric driver, with rows and their distance from the overall average or rate; segments under 2 % of rows (and never fewer than five) are not shown
Clicking a cluster or a segment filters every other widget on the dashboard to it; slicers and cross-filters apply before the analysis
Up to 20,000 rows read by default and 100,000 on request, with plain SQL that runs unchanged on DuckDB, PostgreSQL, SQL Server, Oracle, BigQuery, Snowflake, Databricks, MySQL and Redshift
CSV exports of every point with its cluster, distance and outlier flag, one row per cluster profile, and the driver ranking with every statistic and the segments
Ask Greg answers "what kinds of customers do we have?" and "what drives revenue?" with the same algorithms, and proposes the matching chart
Ask Greg's insight tools: anomalies against the series' own trend and season, forecasts from the same model as the forecast chart, and period-over-period change with each group's share of the total change

How it works

Segmenting customers: select the customer table and click Reports → Charts → Clusters. Its numeric columns become the features and a column of mostly distinct names becomes the label. Leave Number of clusters on Automatic and Scaling on Standardise, read the quality line, read the profile, and only then name the clusters — Loyal, Occasional, At risk. Clusters are numbered by size, largest first.
Reproducibility is built in: the random start is seeded (42 unless changed) and the rows are put in one fixed order before fitting, so the same data and settings give exactly the same clusters on every machine, in every PDF and on every engine. Restarts runs k-means several times and keeps the tightest fit. Changing k, scaling, the seed, the view or the names never re-runs the query.
Large tables stay fast: automatic k is chosen on a fixed sample of 3,000 rows and the silhouette measured on 1,000, but the final clusters always use every row read; at most 3,000 points are drawn, every outlier kept. A stray text cell in a feature — n/a, a dash — becomes a missing value on every engine instead of failing the chart or counting as zero.
Finding what drives revenue: select the sales table and click Reports → Charts → Key drivers. The left panel ranks the drivers — green when revenue rises with the driver, red when it falls, the palette colour for a category whose effect differs by level; hover a bar for its share, direction, p-value and any warning that two drivers move too closely to split reliably. The right panel lists the segments that stand out.
The method is one linear model of the target on every driver, a category entering as a group of indicator columns, with each driver's Shapley value computed exactly; significance is the partial F-test of a driver given all the others. The quality line shows R² and adjusted R², the rows read, the target's average or rate — and, always, that this is association, not causation. Below R² 0.15 a chip says the model is weak.
Asking Greg: "anything odd in sales?" checks each period against the series' own trend and season and reports the expected value, the deviation and whether it is a spike or a dip, with a lasting step reported as a level shift. "Forecast sales for the next six months" runs the forecast chart's own model and quotes its accuracy and 80 % and 95 % intervals. "This month against last year, which region drove the fall?" gives each group's change and, for a sum or a count, its share of the total change. The period still in progress is left out by default, and the answer says so.
Ask Greg's cluster and driver answers come with a proposed chart showing exactly what it described. A cluster chart is certified by the number of points and clusters predicted, a key-driver chart by the top driver and the row count — and on a sheet, by analysing the cells again without SQL. Neither is certified as meaningful, which no check can prove.

Practices that keep it reliable

Choose two to six features that describe behaviour, not identity — an id or a postcode makes clusters that mean nothing
Read the silhouette before the scatter: below 0.25 the clusters overlap, so try other features, robust scaling or another k before naming anything
Use Combine rows per label when a row is an event (a visit, a trade) and the question is about the customer or the trader
Treat key drivers as a ranked list of places to look, never as proof of cause — a driver that moves with the target may share a cause with it
Leave identifiers out of the drivers and watch the collinearity warning; two drivers that always move together split their credit unpredictably
Turn a recurring insight question into an insight watch or an alert rule, so the next spike is reported without anyone asking (chapter 27)

In a regulated environment

Deterministic by design: the same rows and settings give the same clusters and the same driver ranking on every machine, engine and printed report — the property an auditor needs to reproduce a figure that went into a board pack.
The analysis runs on your computer or server, through the same row- and column-level security as every query. By default only aggregates reach a model — cluster sizes and means, the driver ranking and segments — every segment label is masked under its own column's rules, and a blocked table or column is refused.
Small-group protection where the AI policy allows aggregates only: a cluster with fewer than five members has its means withheld, segments are never smaller than five rows, and in Ask Greg's period and anomaly answers a figure resting on fewer than five rows is folded into Other rather than shown.
Honest by construction: answers built on these tools carry the Unverified badge, because the figures are computed from a table rather than from a certified metric, and the key-driver chart states on every rendering that it shows association, not causation.

Glossary terms on this page

KPI dashboard: definition

ATX Institute on LinkedIn

This site needs JavaScript for the full experience. Every page above is reachable without it.