## What’s with errorbars?

As an applied mathematician, I am often asked to provide errorbars with values. The idea is to give the person reading a statistic or a plot some idea of how much the value or values could be expected to vary or be wrongly estimated, or to indicate how much confidence one has in the statistic. It’s a great idea, and it’s always a good exercise to try to provide the level of uncertainty that one is aware of when quoting numbers. The problem is, it’s actually very tricky to get them right or to even know what “right” means.

A really easy way to screw this up is to give the impression that your data is flawless. Here’s a prime example of this.

More recently we’ve seen how much the government growth rate figures can really suffer from lack of error bars- the market reacts to the first estimate but the data can be revised dramatically later on. This is a case where very simple errorbars (say, showing the average size of the difference between first and final estimates of the data) should be provided and could really help us gauge confidence. [By the way, it also brings up another issue which most people think about as a data issue but really is just as much a modeling issue: when you have data that gets revised, it is crucial to save the first estimates, with a date on that datapoint to indicate when it was first known. If we instead just erase the old estimate and pencil in the new, without changing the date (usually leaving the first date), then it gives us a false sense that we knew the “corrected” data way earlier than we did.]

However, even if you don’t make stupid mistakes, you can still be incredibly misleading, or misled, by errorbars. For example, say we are trying to estimate risk on a stock or a portfolio of stocks. Then people typically use “volatility error bars” to estimate the expected range of values of the stock tomorrow, given how it’s been changing in the past. As I explained in this post, the concept of historical volatility depends crucially on your choice of how far back you look, which is given by a kind of half-life, or equivalently the decay constant. Anything that is so not robust should surely be taken with a grain of salt.

But in any case, volatility error bars, which are usually designed to be either one or two lengths of the measured historical volatility, contain only as much information as the data in the lookback window. In particular, you can get extremely confused if you assume that the underlying distribution of returns is normal, which is exactly what most people do in fact assume, even when they don’t realize they do.

To demonstrate this phenomenon of human nature, recall that during the credit crisis you’d hear things like “We were seeing things that were 25-standard deviation moves, several days in a row,” from Goldman Sachs; the implication was that this was an incredibly unlikely event, near probability zero in fact, that nobody could have foreseen. Considering what we’ve been seeing in the market in the past couple of weeks, it would be nice to understand this statement.

There were actually two flawed assumptions exposed here. First, if we have a fat-tailed distribution, then things can seem “quiet” for long stretches of time (longer than any lookback window), during which the sample volatility is a possibly severe underestimate of the standard of deviation. Then when a fat-tailed event occurs, the sample volatility spikes to being an *overestimate* of the standard of deviation for that distribution.

Second, in the markets, there is clustering of volatility- another way of saying this is that volatility itself is rather auto-correlated, so even if we can’t predict the direction of the return, we can still estimate the *size* of the return. So once the market dives 5% in one day, you can expect many more days of large moves.

In other words, the speaker was measuring the probability that we’d see several returns, 25 standard deviations away from the mean, if the distribution is normal, with a fixed standard deviation, and the returns are independent. This is indeed a very unlikely event. But in fact we aren’t dealing with normal distributions nor independent draws.

Another way to work with errorbars is to have *confidence* errorbars, which relies (explicitly or implicitly) on an actual distributional assumption of your underlying data, and which tells the reader how much you could expect the answer to range given the amount of data you have, with a certain confidence. Unfortunately, there are problems here too- the biggest one being that there’s really never any reason to believe your distributional assumptions beyond the fact that it’s probably convenient, and that so far the data looks good. But if it’s coming from real world stuff, a good level of skepticism is healthy.

In another post I’ll talk a bit more about confidence errorbars, otherwise known as confidence intervals, and I’ll compare them to hypothesis testing.

there are problems here too- the biggest one being that there’s really never any reason to believe your distributional assumptions: this is why God, through Her prophet Efron, gave us the bootstrap.Bootstrapping is also problematic- it assumes that the sample you have is representative!

You might find Holden Karnofsky’s recent blog post titled

Why We Can’t Take Expected Value Estimates Literally (Even When They’re Unbiased)

relevant to the subject of this post.