I left my MMT manuscript to rest. I think I will do another pass of reading it within the next week or so, and if I am happy, send off to my editor. It is easier to spot problems in text that you have written after not looking at it after awhile, as otherwise, you tend to read what you meant to write (and not what you actually read).
In the meantime, I have been doing research on mainstream macro, aimed at filling in some chapters in Recessions: Volume II. (Other parts of the book had been written, but the discussion of mainstream macro has been deferred.)
I wanted to write up the r* estimation algorithms, with an eye on giving a qualitative explanation as to what the algorithms are doing. However, I have been somewhat surprised by the behaviour. It is finally dawning on me that the Kalman filter algorithm has not been behaving in the manner that I expected.
The Kalman Filter is an ancient bit of control theory. It was no longer being used in the industrial applications I was concerned with, so I never paid much attention to it. From what I have seen of the control literature, there have long been concerns with its robustness. Unfortunately, it looks like I am going down the rabbit hole. Given its importance in empirical mainstream work, this is not insignificant.
Since that research is getting messier than I expected, I might be quiet on the writing front for awhile (unless I see something about fiscal policy that I want to rant about).
(c) Brian Romanchuk 2020
Hi Brian, I've worked with Kalman Filters in the aerospace industry for a long time, and the KF is still used in industry in estimation and control problems.
ReplyDeleteI attended a lecture by Prof Kalman at Heriot-Watt University in Edinburgh, who gave him an honorary doctorate in 1990. He was a very inspiring lecturer.
Seems a pity to give up on your application, and I'm surprised you found a problem with robustness. I'd say such problems are not the fault of the KF in a wide range of estimation problems, so I'd be interested to hear a bit more about how you are using it.
Not really my application, a I’m puzzling through what some economists are doing with it. The issue might be in the model parameter estimation technique, and I don’t want to dig through the three layers involved.
DeleteGood to hear about it being used. Wasn’t a preferred technique in what I looked at, but I left control systems after on a few years in academia.
Is there a paper in the field of r* estimation you could point me to, where the KF was used or where some other estimation technique was preferred?
Delete