Whenever I make a large trade I always develop a statistical model for it. I find that gives me a framework to trade around. Now, models are not always statistically good, and there may be reasons I don't trust them anyway. But they are a first step. Here's how I went about the oil model.

My simplest framework is this. First I deflate the price. then I form a moving average of the deflated price. I then estimate a model using the price ten years ago and time as independent variables. So it looks like this:

log(defPrice) = a + b * log(MovAvg(defPrice[-10]) + c * time

The -10 term is for mean reversion. Time is there because most commodities have secular trends in price. For most the trend is downward; commodities are getting cheaper relative to services etc. For a few the trend is upward, probably because of increasing difficulty of extraction.

Oil is one that has had an increasing trend. In the model, I manually stopped the trend five years ago. This is due to a new technology, fracking, that has somewhat changed the game.

Using monthly data, the oil model is surprisingly accurate. It's R-squared is about 85%. All the other diagnostics (about 20 of them) are also good. For example, here's a plot of the residuals versus pure normality:

This is fairly good. The residuals are pretty much normally distributed, although they miss on the upper tail. That doesn't bother me. There's a lot of evidence that liquid asset prices have fat tails, and in commodities, the "fat" is usually on the upside.

Forecast for spot Brent...
Assuming inflation at 2% and the $US unchanged,
the price will be 64% higher in ten years.

Now that may not seem like a big move. I have found, however, that a commodity rarely takes the full ten years to get to equilibrium. More like two or three.

For shorter term (1 - 3 years) forecasts, you need to input data on supply/demand/capacity. That's for another post.