After "how much", the next question a board asks is "why": what drives revenue, why customers churn, what moves the margin. Most BI tools answer with a correlation per column, which double-counts drivers that move together and changes with the order columns are listed in. ATX AnalyticsTable ranks the candidates by their Shapley share of R² instead. The shares are computed exactly, add up to what the drivers explain together, and split credit fairly between correlated drivers. The chart also shows the direction, the significance and the segments that stand out, and says on every rendering that this is association, not causation.
It needs a target, the column to explain, and up to twelve candidate drivers, numbers and categories alike. A target with two values or with text is treated as a yes/no outcome and analysed as a rate, so churn, default or a late delivery works as well as revenue.
The target is modelled on all the drivers at once, a category entering as one group of indicator columns. Each driver's share is its Shapley value: its average contribution to R² over every order in which the drivers could enter the model. This decomposition is also known as LMG or dominance analysis, and it is computed over all 2^p subsets rather than sampled.
So the shares add up to 100 % of what the drivers explain together, they do not depend on the order the columns were listed in, and two drivers that move together share the credit instead of both claiming it. Significance is the partial F-test of each driver given all the others. Each driver also carries its correlation with the target and a variance-inflation factor that warns when its share is unstable.
Exact Shapley shares of R² for every driver, computed inside the product.
The quality line under the chart gives R² and adjusted R², the rows read and the target's average or rate. Below an R² of 0.15 a chip says the model is weak: the ranking is real, but most of what moves the target lies elsewhere. Every rendering says that this is association, not causation, because a driver that moves with the target may share a cause with it rather than be one.
Up to 20,000 rows are read by default and up to 100,000 if you raise the limit, and a chip says when the limit was reached. Twelve drivers on 20,000 rows take well under a second. The query asks for the columns as they are, with no function applied, so it runs unchanged on DuckDB, PostgreSQL, SQL Server, Oracle, BigQuery, Snowflake, Databricks, MySQL and Redshift.
A way to rank the factors that explain an outcome, such as revenue, satisfaction or churn, by how much of the outcome's variation each one accounts for when all of them are considered together.
Correlations count shared influence twice, and regression coefficients change with scale and with which other drivers are included. A Shapley share averages each driver's contribution to R² over every order in which drivers could be added, so the shares add up to the total and correlated drivers split the credit.
Yes. A target with two values, or text, is analysed as a rate. The chart recognises yes, true and 1 as the event, or you name the value that counts as the event.
Not necessarily. The analysis measures association. A driver that moves with the target may share a cause with it, and the chart says so every time.
This site needs JavaScript for the full experience. Every page above is reachable without it.