Dynamic Factor Model(DFM) doubt with mixed frequency data

[Vikas Kumar]

@chadfulton,

Hi Chad,

I would like to know if the DFM class could handle mix freq data.in the example, you have shown for DFM, all the regressors are of the same frequency. I want to know if I can use mix frequency regressors. Basically I want to apply the DFM approach explained in the Forni et al (2001) paper for end of sample unbalanced data set.

Also if you have any examples of DFM with mixed freq data, that would be great.

Thanks

[Chad Fulton]

Hello,

First, in a technical sense, yes the DFM class can handle mixed frequency data, in the sense that it can handle missing data in observations. So for example if you had monthly and quarterly data, you could have your model would be at the monthly frequency and the quarterly series would have NaNs for most months.

Second, in a practical sense, no the DFM class cannot handle mixed frequency data, because for many types of variables (e.g. stock variables), the quarterly observation is not equal to the monthly observation even when it is observed (for example, first quarter GDP is the sum of the first three months of the year, whereas the corresponding monthly observations would be GDP only for a given month). To properly account for this, you need to augment the state space with cumulator equations; see for example Banbura et al. (2013) https://www.econstor.eu/bitstream/10419/153709/1/ecbwp1275.pdf.

Third, as far as I know, the paper you reference does not use the state space form to compute its generalized principal components, but maybe you had something in mind like the Banbura et al. paper instead?

Fourth, if you were interested in the Banbura et al. state space model, then it is possible to implement it using the Statsmodels framework, but we do not currently have a model written for that case.

Finally, if you're interested in writing code to estimate that model, we would certainly appreciate it as a contribution to Statsmodels, and I would be happy to help along the way.

Best,

Chad

[Brock Mendel]

I've got an interest in this too, would pitch in if there's interest in making a push.  I think at some point I saw a comment from Chad about MIDAS models, which falls in a similar category.