Candour and control

A forecast drawn as a single line claims more than anyone knows. The Bank of England made the point in February 1996, when it began publishing its inflation forecast as a fan of nested bands, a form chosen to show the spread of likely outcomes without suggesting a spurious precision. Weather services draw the spread of an ensemble forecast as box-and-whisker steps along the time axis.

Research since then favours forms a reader can count. Matthew Kay and colleagues introduced the quantile dotplot, which turns a distribution into twenty or so dots of equal probability. In a later experiment led by Michael Fernandes, such displays improved decisions about when to leave for a bus. Michael Correll and Michael Gleicher found that a bar with a whisker misleads, because values inside the bar look likelier than values just above it. Susan Joslyn and Jared LeClerc found that showing uncertainty improved decisions and reduced the damage done by a forecast that turned out wrong.

None of this licenses a band on a chart that has no model behind it. A dashboard built on plain arithmetic, with no fitted forecast, should show no interval anywhere. Its expected values stay single figures, and a dotplot can take their place on the day a forecast with measured error exists. The figure below has a switch for a forecast whose error is not yet measured: the spread disappears and the figure says why.

Ben Shneiderman’s work on dynamic queries set the standard for interactive data in the 1990s. Controls stay visible, every change is small and reversible, and the display answers at once. Bret Victor’s Explorable Explanations carried the idea into prose, with numbers in a sentence that the reader can drag while the conclusion rewrites itself.

The New York Times calculator Is It Better to Rent or Buy?, built in 2014 by Mike Bostock, Shan Carter and Archie Tse, adds the decisive detail. Each slider sits on a small chart of the outcome across the whole range of that input, so the reader sees what a change would do before making it. The price plot below goes one step further. Its curve is the revenue an illustrative product earns at each price, on an assumed demand. The price is a point on that curve, and it is dragged where it stands, so the control and its consequence are one object.

Changing a forecast by hand deserves the same care as any other control. Robert Fildes and colleagues studied judgmental adjustments in supply-chain companies. Small adjustments tended to reduce accuracy, larger ones tended to improve it, and adjustments leaned optimistic. In an earlier experiment Paul Goodwin found that asking for a reason before an override improved accuracy, so an override should ask for one and keep a record that can be reviewed later.

A control also needs a memory. Jeffrey Heer and Ben Shneiderman sorted the interactions of visual analysis into twelve tasks in three groups: specifying data and views, manipulating views, and keeping a record of the analysis. The third group is the one dashboards forget. A scenario is only useful if what was changed can be seen, named and undone. In the figure below every move leaves a faint ghost of the point, an override asks for its reason, and the reader can step back through the moves one at a time.

In 1948 the British Treasury asked Ernest Gowers for a short book to help civil servants write so that the public could follow them. Plain Words set out to replace official jargon with a short, clear style. A dashboard has the same duty and less room. It should say lost sales where a planning system says unfilled demand, weeks of stock where it says cover, and dates where it says week numbers. It should keep the terms its readers will look for, such as gross margin and conversion rate, and define them.

Plain words are half of it. The other half is saying what kind of number the reader is looking at. Robin Kinross argued in 1985 that even a railway timetable persuades. A calm, exact style claims authority whether or not the figures have earned it. A dashboard should therefore say what it is: which figures are measured and which are estimated, what model lies behind them, and what it leaves out. The figures in this guide do the same. Their numbers are invented and labelled illustrative, and they are computed in the page.

Definitions belong where the numbers are. A reader should be able to ask of any figure what it counts, over what period and in what unit, and find the answer without leaving the screen. The figure below does that for eight common business terms.