How to handle Holidays in Time-Series Datasets?

Question

Im currently analyzing a Dataset of the German Stock market. While Holidays like Christmas or New Year aren't a problem for Return Calculation or Portfolio Performance, im testing some regressions and don't know how to handle these Dates.

I'm regressing the Return of my Portfolio, on the Market Returns of the last ten days. Then im adding the betas up, so i can plot the time varying betas of my sample for every point in time.

the right side of my regression looks like this:

Do these days have to be cancelled? I don't think the guys of the paper im replicating cancelled out each holiday plus the ten days before. However the regression results would be biased if Non-Trading-days are in the sample.

Answer 1

Preliminary:

I assume from your previous post, that you are using Thomson Reuters Datastream.

---

There are several additional parameters available on your Datatypes. Let's for example look at the Datatype MV, which is the market value of a company:

1. If you are using MV, than you obtain a time-series, where the last available value is repeated, if a stock is (a) not traded, (b) is suspended or (c) on exchange holidays. The last value is also repeated up to your requested date, if the company went dead!

2. If you are using MV#S, the description from Datastream states:

   The #S qualifier unpads values where the underlying data point is stored as a null value - so displays N/A for null values rather than pad the last real value.

You may use MV#S to set (incorrect and repeated) values on exchange holidays to NA within your request. However, if you are dealing with meanwhile delisted stocks, you may additionally request MV#T to set repeated values after delisting to NA, or manually search on Data item TIME (or Worlscope Item WC07015), which

represents the day on which a company became privately held, merged, liquidated, or otherwise became inactive

and (manually) set values newer than this inactive date to NA.

Answer 2

When you handle data of any type you might have the issue of missing elements.

You can generally handle it by global ommission or by imputation.

Imputation can be feed forward, feed backward, some averaging or interpolation scheme.

One such scheme might be to use your existing data to build a machine learning algorithm that predicts market movements conditioned on other values, as a Bayesian problem.

For example FTSE rallies 100 points, S&P rallies 50 points and your model predicts the DAX rallies X points. You then use X as the imputated value for your data. This approach may strengthen your results otherwise.