Hi all,
I would like to investigate a financial index time series. From the observation, I notice that there might be one or multiple structural breaks embedded in the time series.
Is there a way for Mathematica to detect those structural breaks and the breakpoints? I used to use Eviews for analysis but now want to transform my workflow to Mathematica. I would appreciate it if someone can provide me the hint.
Best,
Tuoyi
A concrete numerical example would be quite useful.
For example, the Bitcoin price data which is easy to obtained. In different periods of time, the series illustrates different behaviors. I can determine at least two breakpoints for the whole time series in Eviews. Can I do the same thing in Mathematica in a straightforward way?
To paraphrase an old quote, There’s nothing like a concrete example. And that was nothing like a concrete example.
A (not too small) list of representative {x,y} values would make for a concrete example.
I do not have access to Eviews on my personal Mac, or else, I’m able show you screenshots of structural break test result by Eviews. However, I just generate some random numbers from normal distribution and combine them together, from which you can obviously notice a jump from graph. Can Mathematica detect that structural break and determine the time of that breakpoint?
https://www.wolframcloud.com/objects/81e17497-e6e3-4b1f-b11f-7576d400ec59
I would expect the result similar to this one Working With Breakpoint Equations with automatic calculation to numbers of breakpoints and places where breakpoints are.
Tuoyi,
If I understand what you are trying to do, I would use Differences to find the jump:
DateListPlot[Differences[ts], PlotRange -> All]
to get the break at the peak
If the noise is a problem you can filter it: (in this case with a 10 day filter)
filt = MeanFilter[ts, 3600*24*10]
DateListPlot[{ts, filt}]
ListLinePlot[Differences[filt], PlotRange -> All]
Now you have a more distinctive break (peak). There are many ways to find the peaks depending on exactly what you are trying to do. I hope this helps.
Regards,
Neil
Hi Neil,
Thank you so much for your answer.
Your solution works well on my sample data, but the case is that my sample data does only contain one innovation on mean. However, for the real world case, such as Bitcoin price data, there might be a combination of innovations on many parameters, and the underlying distribution in each period may vary as well (ARMA + Random Walk + AR, with two structural breaks, for example).
I create a different time series, which simply combines two AR(1) process, but with different parameters. Then the structural break is not that obvious. https://www.wolframcloud.com/objects/7dd2812d-1b12-47bf-a1e8-d48303fe28cd
So, I just wonder, does mathematica have a universal (or partly universal) way/function to detect structural break? Just as Chow test or Bai-Perron test.
Best,
Tuoyi
Hi all,
for anyone who is interested in this topic. You can now refer to this excellent post .
