Observation-driven hierarchical density models for missing data imputation

We propose an observation-driven dynamic common factor model for missing value imputation in high-dimensional panel data. The model exploits both serial and cross-sectional information in the data and can easily cope with time-variation in conditional means and variances, as well as with either isolated or long patches of missing values. The approach not only provides point forecasts, but also density forecasts for the missing data, thus allowing the researcher to quantify imputation uncertainty. The model's static parameters can be estimated by standard maximum likelihood methods due to the model's observation-driven structure. We apply the model to impute densities and risk quantiles in high-dimensional panels of daily global credit curves with large gaps of missings.