Finding structural breaks in heteroskedastic time series

Question

I'm trying to identify structural breaks in the movement of reserve currencies. I'm not yet all that versed in the finer details of time series, but I've been reading up on ARCH and GARCH estimators. However, in my model I don't really want to look at any external factors, like government interventions, gdp, macroeconomic drivers, etc. Instead I want to establish a certain confidence interval (on returns or variance?) and identify abnormal movement outside that bandwidth.

The classical structural break identifiers seem to me to be very "exclusive", meaning they are good at identifying massive underlying changes; however for me it wouldn't be a problem to have some false positives, because I want to test how breaks over a certain magnitude occur.

I was thinking of employing some sort of rolling vintage, or segmenting the data at different rates so that changes can be identified, but my mind is all over the place in this regard.

What I'm asking is how you would go about detecting abnormal variance or behavior in a univariate timeseries with tons of datapoints. Do I have to go and set up an ARCH/GARCH model or is there a simpler, cruder way for an algorithm that I can tune to give me points in which behavior of the time series changes?

Answer 1

There are at least a couple of R packages that you might find useful. They are discussed on R-bloggers, but obviously you will have to go deeper into the respective documentation.

You are especially referring to heteroscedasticity. So, if for example you will try segmented, make sure you run the fit with glm (generalized linear model, which does not assume constant variance).

I started using segmented for my own purposes, and saw after some time that the results were not very stable. You might be in a data rich setting compared to mine, so this might work better for you.

What I'm doing now, and I would recommend you to think about it, is to use non-parametric methods. Use some smoothing method and make sure to use cross-validation to choose the best bandwidth (see Shalizi, Advanced Data Analysis chap. 4). A quick google search showed me several results that might be relevant to you.