5
$\begingroup$

I watched a presentation from a large quantitative finance firm that spends a lot of effort around predictive modeling. One of the points the presenter emphasized was that they deal with a lot of asynchronous predictive features. So, for example one feature set might revolve around several company quarterly earnings, and it might be used to predict some future daily statistic related to a single company. The data is asynchronous in that (of course) not all data will arrive at the same time, and it is sparse in that the earnings events might occur with a frequency of only four times per year, for example. In this example, let's just assume that the data features matrix is on a daily time scale resolution. He did not, however, discuss exactly how they might aggregate over or impute the missing data.

I could understand that one might just aggregate the data on a quarterly scale, but in the case he described there are other features that occur on a much higher than quarterly frequency, for the same features matrix (e.g. daily time series). My intuition would be that they might just fill down the empty data for the sparse features.

I'm curious to know if anyone builds these kinds of models, and how they might go about cleaning the data set, before continuing to test some kind of model around it. Any literature with pragmatic examples would be great.

$\endgroup$
  • $\begingroup$ What is the name of the firm? Do you have the presentation? $\endgroup$ – fni May 29 '18 at 13:34
  • $\begingroup$ @igorpozdeev 's response below is helpful. A statistically trained analyst might contend that, as described, such a data structure conforms to a finite mixture model integrating features of mixed scale and temporality. While there are lots of possible methods for analyzing data of this type, panel data models are among the most informative and tractable. That said how would cleaning the data in this case differ from any other type of data or model? $\endgroup$ – DJohnson May 29 '18 at 17:18
5
$\begingroup$

There is large literature on MIDAS (mixed-frequency data sampling) models, the leading scholars being Eric Ghysels and Rossen Valkanov — google their research for references. However, the motivation for these models has mostly been to forecast low-frequency stuff with high-frequency variables, updating, say, quarterly GDP predictions as weekly unemployment figures keep appearing.

Recently, reverse-MIDAS models have been introduced as well (link to one such model), but they seem to me like wrappers for good old regressions with lagged values and of limited use.

$\endgroup$
  • $\begingroup$ upvoted. Some good ideas on how to possibly approach. Will wait a bit before accepting. $\endgroup$ – pat May 29 '18 at 22:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.