E-piphany

In an excellent (and free!) daily email I receive, the Daily Shot, I ran across a chart that touched off my quant BS alert.

This chart is from here, and is obviously a few years out-of-date, but that isn’t the problem. The problem is that the chart suggests that gold prices rise 5.5% every year. If you buy gold in January, at an index value of 100, and hold it through the flat part of January-June, then you reap the 5% rally in the second half of the year.

No wonder people love gold! You can get a 10% annual return simply by buying in July and selling in December!

The problem is that this is not the way you should do a seasonal chart. It has not be detrended. We detrend data because that way, we can express the expected return for any given day as (the normal expected return) plus (the seasonal component). This is valuable because, as analysts, we might have a general forecast for gold but we will want to adjust that forecast to a holding period return based on a seasonal pattern. This is very important, for example, with TIPS yields and breakevens, because inflation itself is highly seasonal.

Now, the seasonal chart done correctly still suggests that the best time to own gold is in the second half of the year, but it no longer suggests that owning gold is an automatic winner. (It is a separate argument whether we can reject the null hypothesis of zero seasonality altogether, but that’s not my point here).

If I was doing this chart, I would also include only full calendar years, so if I move the start date back to January 1, 1982 and the end date to December 31, 2014 here is what I get:

Frankly, I would also use real prices rather than nominal prices, since it is much easier to make a statement about the expected real return to gold (roughly zero over time, although it may be more or less than that based on current valuation metrics) than it is to make a statement about the expected nominal return to gold, since the latter includes an embedded assumption about the inflation rate, which I would prefer to strip out. And I would also include data from the 1970s.