Small multiples for portfolio reporting when eight shows share one page

The portfolio pacing chart had eight lines on it, one per show, each a different colour, all running from week minus 26 to show open. Two of them crossed four times. The legend sat below the chart in two rows of four, so reading any single line meant looking down, matching a colour, looking back up and finding the line again.

Somebody had put real work into it. It answered no question anybody had. Small multiples for portfolios fix this by giving up the one thing that chart was protecting, which is having everything in one frame.

Why does one axis with eight lines fail?

Two reasons, and they compound.

The first is occlusion. Eight registration curves over 26 weeks all rise, all accelerate near the end, and all occupy the same region of the plot. Every line spends part of its length underneath another line, and the reader cannot see where because the crossing points are exactly where the ink is densest.

The second is the legend. A colour legend forces a lookup for every line the reader wants to follow, and the lookup has to be repeated each time attention moves. Tufte named small multiples in The Visual Display of Quantitative Information in 1983 and argued that their economy comes from the opposite property: once a reader has understood the design of one panel, they have immediate access to the data in all the others, and the eye is free to attend to changes in the data instead of changes in the design. A panel titled "Kitchen and Bath, March" needs no lookup at all.

There is a third reason that only shows up in a portfolio. Eight shows in one frame have to share a vertical scale, and a portfolio's shows are not the same size.

What a shared scale costs the small show

Here is the portfolio, closing registrations by show: 42,100, 28,400, 19,700, 14,300, 11,600, 8,900, 5,200 and 3,100.

Draw all eight on a shared axis running from zero to 42,100. The smallest show's curve tops out at 3,100 divided by 42,100, which is 7.4 per cent of the plot height. In a panel 90 pixels tall, the entire twenty-six week campaign of that show is 6.6 pixels of vertical travel. Its curve is a flat line along the bottom, and so are the two above it.

That is not a rendering problem to solve with a bigger chart. At 400 pixels tall the smallest show still only gets 30 pixels, and now the grid does not fit on a page. A shared absolute scale genuinely cannot show the shape of a small show next to a large one, and pretending otherwise is how portfolio pages end up with a chart everybody nods at and nobody uses.

The wrong fix is a free scale per panel, where each panel auto-scales to its own data. The panels then look identical, sit in a tidy grid, and mean completely different things. A reader compares them because the design invites it, and there is no visual cue that the invitation is a trap. If you use free scales, print each panel's maximum in the panel header at the same size as the title, and expect people to miss it anyway.

Free scaling also brings the baseline problem in through the back door. An auto-scaled panel starts its vertical axis just below its own minimum, so a show that moved by 3 per cent across the campaign fills the panel top to bottom and looks like a show that tripled. That is the truncation effect V24 works through, multiplied by eight and hidden behind the reassuring symmetry of a grid. On a panel grid the fix is the same as it is on a single chart, which is to fix the baseline once for the whole grid and let the flat panels be flat.

Indexing, and the reversal it uncovers

The fix that works for pacing is to index each show to its own closing figure, so every panel runs from 0 to 100 and every panel uses its full height.

Work an example. At day minus 60, the smallest show holds 1,860 registrations against a close of 3,100, so its index is 60. The largest show holds 23,155 against a close of 42,100, so its index is 55.

On the shared absolute grid, the small show's day minus 60 point sits at 4.4 per cent of the panel height and the large show's sits at 55 per cent. Any reader glancing at that page concludes the large show is far ahead. On the indexed grid the small show reads 60 and the large one reads 55, and the small show is the one further through its campaign.

Both charts are correct. They answer different questions, and the shared absolute grid answers the question nobody asked, which is which show is bigger. Everyone already knows which show is bigger.

So build two grids and label them. The indexed grid carries the pacing question, with each panel's absolute close printed in its header so the magnitude is available as a number. The absolute grid carries the size question, and it is honestly a bar chart rather than eight curves, because the only value on it that matters is the endpoint.

Indexing to a prior edition's close rather than this edition's projected close is the safer construction, because a projection changes under you and last year's close does not. That choice interacts with how the horizontal axis is built, since comparing shows at the same distance from doors rather than the same calendar week is its own piece of work (O9).

Sorting, and what the reader learns first

Panel order is a free variable and most grids waste it on alphabetical order.

Sort by size descending, reading left to right and top to bottom, and the reader picks up the portfolio's shape for free while looking for something else. Sort by change against prior edition and the largest mover leads, which is what a monthly review usually wants. Either is defensible. Alphabetical is a decision not to have made a decision, and it puts the eight thousand registration show between two forty thousand registration shows for no reason.

Say the sort in the caption. "Eight shows, sorted by registrations at close, matched scale" is nine words and removes the entire class of question that starts with "why is that one there".

Eight panels at four across and two down is a comfortable density on a slide or a screen. Twelve is the point where a reader stops seeing a set and starts seeing a wall, and the visual search for the odd panel out slows down noticeably. A portfolio of thirty shows needs a different page, which is the roll up V38 covers, usually a sorted table with the panels reserved for the shows that moved.

The same logic scales down one more step. A panel shrunk to the width of a table cell becomes a word sized trend inside a row, which is V27's subject and the right answer when you have 24 categories rather than 8 shows.

When is one chart with eight lines better?

Javed, McDonnel and Elmqvist tested this directly and published in IEEE Transactions on Visualization and Computer Graphics in 2010, volume 16, issue 6, pages 927 to 934. They compared techniques that give each series its own chart, including small multiples, against shared space techniques such as a standard line chart, across comparison, slope and discrimination tasks.

Their result splits by visual span. Separate charts are generally more efficient for comparisons across series with a large visual span. Shared space techniques are typically more efficient for comparisons over smaller visual spans.

Translated into a portfolio page: if your eight shows are all between 11,000 and 14,000 registrations and the question is which two are ahead of each other by a few hundred, put them on one axis. The lines will sit close together, the comparison is a short eye movement, and panels would force the reader to hold values in memory across the page. The moment the range widens, as it does in most real portfolios, the panels win.

There is also one question small multiples simply cannot answer, which is when one series overtook another. Crossings live in the shared frame. If that is the finding, draw the two shows that cross on one axis and leave the other six out.

Where this stops

Matched panels make comparison cheap, and cheap comparison of things that are not comparable is a hazard rather than a feature.

Eight shows on a matched pacing grid look like eight instances of one process. They are eight different markets with different lead times, different registration open dates and, frequently, different definitions of what counts as a registration. A grid that renders them identically implies a consistency that the underlying data may not have. Before the grid is worth drawing, the metric behind every panel has to be the same metric, and that is usually more work than the chart.

The second limit is that a small multiples grid shows where something differs and never why. A reader who spots the odd panel has a question, not an answer, and the page needs a route from the panel into the detail behind it or it generates traffic to your inbox.

Take your eight registration curves, index each one to its own close, and draw them as eight panels at matched axes with the absolute close printed in each header. It is an afternoon in whatever tool you already have, and the first thing you will notice is that at least one show is nowhere near where the single-axis chart suggested. Consistency of the axis choice across the rest of your BI and reporting pages can wait until after you have looked at that one.