What's up with tornado plots?

There's been attention recently on a peculiar kind of graph. I don't know if it has a name, but let's call it a "tornado plot." An example is this COVID-19 chart. You can see other examples in the book Slowdown.

A tornado plot is a way to chart a time series. Unlike a conventional graph, which shows time on the x-axis and a value on the y-axis, a tornado plot has a twist. The value is still on the y-axis, but the x-axis shows the rate of change at each moment in time: that is, how much the value is increasing or decreasing.

To understand how to read this kind of plot, I made the interactive tool below. You can draw your own time series in the first chart (in the form of a normal graph) and then see how a tornado plot shows the same data. Try it out!

If you draw graphs with ups and downs, you'll see they become loops in the tornado plot. The bigger the variations, the wilder the loops.

When I first saw a tornado plot, I tried to interpret these loops as some kind of cyclic behavior. For instance, in the COVID-19 chart, I wondered about day-of-week effects. And it's true: a time series with a cyclic pattern will produce loops.

However, as Anton Geraschenko pointed out to me, any smooth graph with ups and downs will produce a loopy result! That's a natural consequence of the fact that, by construction, the curve in the tornado plot has to go downward when it's on the left side, and upward when it's on the right side.

In other words, a tornado plot is a machine for generating loops. It's possible that for near-periodic time series, this might be a useful feature, making deviations from a periodic baseline more obvious. And it certainly generates engaging shapes that make you look twice, so could be useful for storytelling.

But for workaday data analysis, or straightforward data communication, I think it's easier to read an old-fashioned x-y graph.