Every time the statistical authority releases new data there is a surge in economic commentary. Not analysis, but commentary. A thoughtful analysis would usually say that a single new data point doesn’t provide enough information to change anything we thought previously. There’s just too much randomness and error in point estimates to be able to tell much from them. Commentary is different because it creates a narrative and fits the data into that narrative.
A good example is the narrative about double and triple-dips in the UK. Commentators made much of the ONS’ revisions to the GDP series that ‘revised away’ the triple-dip, ‘vindicating Osborne’. The revisions may have eliminated a slight dip in GDP but they didn’t change anyone’s understanding of what had happened in the macroeconomy. That data was important for commentary but not for analysis. In fairness to commentators, distinguishing genuine trends from randomness is not easy. Our eyes are drawn to ‘streaks’, whether in football games or economic time series, even when the series is essentially random. Economists are always looking for techniques to separate the streaks from the randomness. The problem we face is that many of the tools are fairly impenetrable to casual observers and hard to explain.
Edward Tufte has suggested using randomised sparklines to visually distinguish genuine trends from deceptive streaks, so I thought I’d give it a go with the last four years of UK unemployment data. Here is the monthly change in the UK unemployment rate since June 2009: . We think that a recovery has begun so the recent years’ falling unemployment looks good. Now let’s try randomising the values and see if the ‘streak’ disappears.
Here are ten sparklines with shuffled values: To my eye there is still a lot of streakiness there, which suggests that the current run might be just randomness. Indeed, a more traditional graphic, such as the partial autocorrelation plot, confirms that the data isn’t streaky:
That doesn’t mean we aren’t seeing a recovery—unemployment tends to lag other indicators and the recovery is recent—but we should be hesitant to causally interpret the latest release. So far, so good, but it’s nothing that any economist couldn’t have told you already. Autocorrelation is already used by most people to gauge ‘streakiness’ far more effectively than the randomised sparklines, so what do they add? When talking to another economist I’m not sure that they are useful. When communicating with people who don’t know their ARs from their MAs they could be extremely helpful in overcoming the narrative bias. The key will be to make them accessible and quick to produce as an explanatory tool since I don’t see them being used in final outputs.