What neighbouring fields know

In February 1996 the Bank of England began to publish its inflation forecast as a fan of probability bands. Erik Britton, Paul Fisher and John Whitley explained why in 1998. The earlier chart, a central line inside one shaded band, had often been read as giving upper and lower bounds. It also drew attention to small moves in the central projection. The fan shows the spread of likely outcomes without a spurious precision. The ECMWF meteogram treats an ensemble weather forecast in a related way, with a box-and-whisker plot for each variable at every step.

Experiments since then point one way. Matthew Kay and colleagues introduced the quantile dotplot in 2016: a distribution drawn as twenty or so dots of equal probability. Probability estimates read from it varied less than estimates read from a density curve. In 2018 Michael Fernandes and colleagues measured decisions. People chose when to leave for a bus, and those using a dotplot of fifty dots earned about 97% of the best possible payoff. Dotplots and cumulative curves beat intervals, density curves and text by one to five percentage points. Susan Joslyn and Jared LeClerc found in 2012 that uncertainty in a forecast improved decisions about salting roads. It also raised trust and reduced the harm done by forecast errors.

The evidence is consistent and narrow. The studies are controlled experiments, mostly with members of the public making everyday choices, and none tested buyers deciding stock. The findings also qualify each other. In Fernandes’s experiment fifty dots did better than twenty, and cumulative curves, often thought too technical for the public, came close behind. Jessica Hullman, Paul Resnick and Eytan Adar found in 2015 that animated draws beat error bars and violin plots for judging which of two quantities is larger. For most questions about one quantity, all three gave similar accuracy. Error bars on bars carry a bias of their own. George Newman and Brian Scholl showed in 2012 that values inside a bar look likelier than values outside it, and Michael Correll and Michael Gleicher found the same bias in 2014.

A plan’s arithmetic usually treats demand as one number. In the example, Product C enters the week of 19 October with 180 units, exactly that week’s expected demand. If 180 is the middle of the forecast, the chance of running out is one in two. Sam Savage calls this the flaw of averages: plans based on average assumptions are wrong on average. A planning dashboard should keep a place for the spread of demand. It should fill that place only when a forecast with measured error exists, because a band drawn without a model claims knowledge no one has.

Some of the earliest statistical charts were drawn for trade and accounts. William Playfair’s Commercial and Political Atlas of 1786 drew England’s imports and exports as two lines and tinted the area between them as the balance of trade. Willard Brinton’s Graphic Methods for Presenting Facts of 1914 collected the chart forms of its day for readers in business. Mary Eleanor Spear’s range bar of 1952 is regarded as a forerunner of the box plot. These forms spread through use, and none was tested as the perception studies test encodings.

Business reporting also has a notation standard. The International Business Communication Standards, developed by Rolf Hichert and colleagues and set out with Jürgen Faisst in Solid, Outlined, Hatched, tell scenarios apart by fill. Actual figures are solid dark grey, the previous year solid light grey, plans outlined and forecasts hatched, and hue is left for other jobs. Version 2.0 appeared in June 2026, aligned with ISO 24896, a notation for business reporting. The standard rests on consistency and practice, and has been studied far less than the encodings of the perception literature.

Practitioners disagree about comparators. Stephen Few’s bullet graph, specified in 2006 and revised in 2013, sets a measure against one or two comparators and up to five qualitative ranges in one bar. Nick Desbarats listed eleven downsides in 2022: problems do not stand out, measures where lower is better or a middle range is best need variants, and words such as good and poor mean different things to different readers. He now colours a single dot by four action thresholds. Donald Wheeler’s Understanding Variation of 1993 gives a comparator with a statistical basis, limits at 2.66 average moving ranges either side of the mean. Amazon’s weekly business review, as Cedric Chin describes it from Colin Bryar’s practice, draws most of its 400 to 500 metrics in one form, the trailing six weeks beside the trailing twelve months.

Decision analysts add a way to rank assumptions. A tornado diagram moves one input at a time between a low and a high value, draws the change in the result as a bar, and sorts the bars longest first. Operations researchers usually cite Ronald Howard’s review of 1988 for it. Practitioners take the low and high values from an assessment of each input, often its 10th and 90th percentiles, in place of a fixed percentage. Ted Eschenbach found in 1992 that it complements the spiderplot, which shows fewer variables in more detail. Together these conventions give a planning dashboard three disciplines: one fill per scenario, a named comparator for every figure, and assumptions ranked by what they do to margin.

Merchandise planners work in a weekly grid of sales, stock and intake, from which the budget left to buy follows. Robert Fildes, Shaohui Ma and Stephan Kolassa reviewed retail forecasting research in 2022 and found little evidence for many of the complex methods proposed. The M5 competition asked thousands of teams to forecast 42,840 series of Walmart unit sales, many of them sporadic. Machine-learning methods won. Spyros Makridakis and his co-organisers found the gains mainly at the upper levels of the hierarchy and small for single products in single stores, where stock is decided.

Planners rarely accept a forecast as it comes. Fildes, Paul Goodwin, Michael Lawrence and Konstantinos Nikolopoulos examined more than 60,000 forecasts from four supply-chain companies in 2009. In three of them, adjustments improved accuracy on average. Larger adjustments tended to help and small ones often did harm. Upward adjustments helped less often than downward ones and more often went the wrong way, a sign of optimism. In an earlier experiment Goodwin found two controls that improved accuracy: making the statistical forecast the default, and asking for a reason before changing it. Michael Gilliland’s forecast value added measures each step of a process against a naive forecast, so a review that makes the forecast worse shows as a loss.

Other findings bear on what to buy. Marshall Fisher and colleagues described in 1994 how Sport Obermeyer, a skiwear maker, read the uncertainty of each style from the spread of its buyers’ independent forecasts, made predictable styles first, and halved the cost of stockouts and of price cuts on unsold stock. For a single buy before a season, the newsvendor model stocks to the quantile of demand given by the critical ratio: the cost of a lost sale over that cost plus the cost of an unsold unit. Aris Syntetos, John Boylan and John Croston sorted demand into four classes in 2005, with cut-offs at 1.32 for the average interval between sales and 0.49 for the variability of their sizes; Andrey Kostenko and Rob Hyndman replied in 2006 that only two classes matter. Field experiments at Zara and at Rue La La found that model-based clearance and pricing raised revenue, by about 6% and 10%.

This evidence comes from real companies, which makes it relevant and narrow. Its lessons for a planning screen are concrete. A forecast override should start from the statistical figure, ask for a reason and keep a record, so that its value can be measured later. A line’s demand class belongs beside its forecast. A stock target can come from the cost of a lost sale and the cost of an unsold unit, in place of a flat number of weeks.

Visualization tools usually treat the table as a preparatory step, a way of getting data ready for a chart. Lyn Bartram, Michael Correll and Melanie Tory interviewed twelve people who work with data in their jobs. They reported in 2022 that these people want to see and handle the table at every stage of an analysis. They reshape it, mark it up and build alternatives inside it. The authors argued that the interactive table is a visualization idiom in its own right. Their evidence is qualitative and their sample small.

The rules for setting a table have changed little in a century. Andrew Ehrenberg’s Rudiments of Numeracy, of 1977, asks for figures rounded to two effective digits, rows and columns ordered by size, and numbers to be compared placed in a column. Richard Rutter’s guide of 2017 quotes Jan Tschichold’s advice to do without rules where possible, and asks for numbers aligned right in tabular figures. Jonathan Schwabish’s ten guidelines of 2020 restate the same points for current software. These rules come from the practice of statisticians and typographers more than from experiments, and they rarely contradict one another.

A table can also become a graphic without losing its figures. Jacques Bertin’s reorderable matrix, set out in his book of 1977 on graphics and information processing, draws each cell as a mark, then moves rows and columns until groups appear. Charles Perin, Pierre Dragicevic and Jean-Daniel Fekete rebuilt the method as the web tool Bertifier in 2014. Ramana Rao and Stuart Card’s Table Lens of 1994 drew bars inside cells and opened a row to its figures on focus. LineUp, by Samuel Gratzl and colleagues in 2013, builds rankings from weighted bars in a table. Polaris, by Chris Stolte, Diane Tang and Pat Hanrahan in 2002, extended the pivot table into an algebra of small charts and became the research behind Tableau.

For series over time the evidence is experimental, and it depends on the task. Waqas Javed, Bryan McDonnel and Niklas Elmqvist found in 2010 that separate charts, such as small multiples, are generally more efficient for comparisons that span many series. One shared chart is more efficient for comparisons close together. Jeffrey Heer, Nicholas Kong and Maneesh Agrawala found in 2009 that mirroring negative values costs no accuracy. Below 24 pixels of height a two-band horizon graph read more accurately than a line chart, and four or more bands slowed readers down. On a planning screen the grid therefore stays a table, ordered by what matters and rounded to the digits that vary. Each product line gets its own small chart when lines are compared across the season.

In 1984 William Cleveland and Robert McGill ranked the elementary tasks a reader performs to decode a quantity from a chart. Position along a common scale came first, then position along separate scales, then length, direction and angle, then area, then volume and curvature, and shading and colour saturation last. Their experiments tested the upper part of the ranking. Jeffrey Heer and Michael Bostock repeated them in 2010 with workers recruited online and found the same order for position and length. They added angle and area. Both were read less accurately than position, area less accurately than angle, rectangles as poorly as circles, and squares worst of all.

The ranking has known limits. Theory predicts that angle is read less accurately than length, and neither study found it. The task in both is a judgement of proportion between two marked elements, which is a small part of reading a dashboard. Steven Franconeri and colleagues’ review of 2021 adds that a viewer extracts broad statistics from a display quickly while comparing subsets of values is slow. The order is reliable for decoding single values and says less about what a reader notices first.

Two older findings govern surfaces. Stephen Palmer showed in 1992 that elements inside a common region are seen as a group, strongly enough to overcome nearness and similarity. Josef Albers showed in Interaction of Color, 1963, that a colour is read against its surround, so one grey looks different on two parents. Colour also depends on size. Maureen Stone argued in 2012 that small marks need larger colour differences, and Danielle Albers Szafir measured the effect in 2018 for points, bars and lines. Cartographers reached the same restraint through practice. Eduard Imhof’s rules keep strong colours for small areas on muted grounds. Hand-drawn Swiss relief shading turns its light locally to bring out each landform, and in 2021 Bernhard Jenny and colleagues trained a neural network to imitate it.

On a planning dashboard the consequences are plain. Quantities go on aligned positions, bars from a common baseline or points on a shared scale, and the number is printed wherever the amount matters. Surface tone marks grouping and hierarchy, and because a tint is read against its parent it never carries a value from one surface to another. Saturated colour goes on few, small marks, with differences large enough for the size of the mark.

Process plants are run from screens that hold many values and must make a deviation obvious at once. Bill Hollifield, Dana Oliver, Ian Nimmo and Eddie Habibi set out the high-performance approach in 2008: grey backgrounds, analog indicators with the normal range marked, and colour kept for abnormal conditions, so that colour itself becomes the signal. The ISA-101 standard of 2015 makes a written HMI philosophy and style guide part of each display’s life cycle, and leaves choices such as colour to them. In 2013 Alf Ove Braseth and Trond Are Øritsland set out principles for large overview screens, from the Information Rich Design concept developed in Halden.

Engineers also built a theory of such displays. Kim Vicente and Jens Rasmussen’s ecological interface design, published in 1992, asks a display to make the constraints of the work domain visible. Kevin Bennett and John Flach describe the result as a configural display: variables drawn so that their relation forms a shape, which a fault distorts. David Woods named visual momentum in 1984, the ease with which a reader carries what was just seen into the next view, and listed its aids, among them overviews, landmarks and fixed positions. Medicine has the same form at a small scale. Seth Powsner and Edward Tufte’s graphical summary of patient status, in 1994, put a patient’s results, vital signs and treatments on one page as a matrix of small graphs. Each shows its latest value and normal limits, on a time axis that gives today, the past week and the past month similar widths.

The evidence is uneven. Vicente’s own review of 2002 found that ecological interfaces improved problem solving in experiments, and that none had yet been evaluated with professional operators on a full-fidelity simulator. The high-performance guidance was written by practitioners from industrial experience. Practitioners also disagree about how much to show. Stephen Few’s bullet graph draws every measure with its qualitative ranges, while the control-room approach, like Nick Desbarats’s action dots, leaves normal states almost silent and spends emphasis only on exceptions.

For a planning dashboard the lesson is to make normal quiet. A measure inside its normal range can be a grey band and a pointer, with colour only when it leaves the range, so the eye goes where the problem is. Positions stay fixed from week to week, so a reader learns where each measure lives. A relation such as a cash balance or the budget left to buy can be drawn so that its terms form one figure, as the guide’s rule on drawing the identity asks.

Researchers have only recently studied dashboards as a genre. In 2019 Alper Sarikaya, Michael Correll, Lyn Bartram, Melanie Tory and Danyel Fisher coded 83 dashboards against fifteen factors, grouped under purpose, audience, visual and interactive features, and the meaning of the data. The dashboards fell into seven clusters, from strategic and operational decision-making to static displays, personal tracking and public communication. The word covers very different objects, and a design that suits one kind can fail another.

Benjamin Bach and colleagues extended the corpus to 144 dashboards and in 2023 published 41 design patterns in eight groups; an earlier preprint listed 42. Three groups, with 19 patterns, concern content: data information, meta information and visual representation. Five groups, with 22, concern composition: page layout, screen space, structure, interaction and colour. The authors also proposed a model of trade-offs. Abstraction, screen space, the number of pages and interaction pull against one another, so reducing one raises another.

Older work explains the parts. Michelle Wang Baldonado, Allison Woodruff and Allan Kuchinsky gave eight rules for multiple views in 2000, among them parsimony: add a view only when it earns its cost. Michael Gleicher and colleagues showed in 2011 that designs for comparison are assembled from three operations: juxtaposition, superposition and an explicit encoding of the difference. Nicholas Kong and Maneesh Agrawala described five kinds of overlay in 2012, from gridlines to annotations. In 2023 Chase Stokes and colleagues asked 302 people to rank line charts carrying different amounts of text. The most heavily annotated charts were preferred, both to charts with less text and to text alone.

Most of this work is descriptive, built by coding examples. The few experiments cut against strict minimalism. Scott Bateman and colleagues found in 2010 that embellished charts were described as accurately as plain ones and remembered better two to three weeks later. Ohad Inbar, Noam Tractinsky and Joachim Meyer found in 2007 that students preferred a standard bar chart to its most reduced version. Robin Kinross had argued in 1985 that even a railway timetable persuades, so a plain style is a choice with effects of its own. For a planning dashboard this leaves three decisions to make on purpose: which kind of dashboard it is, which comparison design each view uses, and what text sits at the marks.

Dense figures have had their own typefaces for decades. Matthew Carter drew Bell Centennial in the late 1970s for AT&T’s telephone directories, set small and printed fast on cheap paper. Tobias Frere-Jones drew Retina in 1999 for the stock listings of The Wall Street Journal, at seven points and below. Adrian Frutiger planned Univers in 1957 as a family of twenty-one variants named by two-digit numbers, the first for weight and the second for width and slope, so that a page could use several and keep one design.

On a screen the requirements are few. Figures must be tabular, so that every digit has one width and columns align; OpenType calls the feature tnum. They should also keep that width when the weight changes, so that a selected row can turn bold without its columns moving. Variable fonts offer other routes. Roboto Flex has a grade axis, which changes how heavy letters look without changing their width. Instrument Sans has a width axis from 75 to 100, which gives narrower labels from the same family.

Seventeen open typefaces were measured for this guide in a browser, and seven are reported here. Instrument Sans, IBM Plex Sans and Hanken Grotesk keep every digit’s width between regular and bold. Their minus sign is as wide as a digit, so signed columns align, and Hanken Grotesk also keeps the width of its decimal point. Geist and Archivo widen their digits in bold. A recheck on 3 October 2026 qualified two results. In Red Hat Text the zero keeps its width, but the bold nine is wider than the other digits. Roboto Flex has an optical-size axis, so its result depends on size. At 1,000 pixels its bold zero is about a quarter wider; at 15 pixels the difference is under 0.5%.

The historical faces are evidence of practice. The width test is a measurement anyone can repeat, though its result depends on the font build and on the size at which it is made. On a planning dashboard, choose a face whose digits and minus sign keep one width in every weight the screen uses. Test it in the browser at those sizes, and use a true minus sign. Where a face fails, emphasis can come from grade or colour in place of weight.