Tools for converting low time series data to high frequency, based on the R package tempdisagg
, and espeically the accompanying paper by Sax and Steiner 2013.
tsdisagg
allows the user to convert low frequency time series data (e.g., yearly or quarterly) to a higher frequency (e.g., quarterly or monthly) in a way that preserves desired aggregate statistics in the high frequency data. It should, for example, sum back to the original low-frequency data.
In addition, regression-based methods are also implemented that allow the user to supply "indicator series", allowing variation from correlated high-frequency time series to be imputed into the low frequency data.
If you have any questions or issues, please open a thread. Pull requests to add features or fix bugs are welcome. Please clone the repository locally to have access to the testing suite.
To install, use
pip install tsdisagg
Currently, only conversion between yearly, quarterly, and monthly data is supported. Conversion to lower frequencies is non-trivial due to the calendar math that needs to be added, but this is on my to-do list.
The following interpolation methods have been implemented:
Single series, non-parametric methods:
- Denton
- Denton-Cholette
Multiseries, regression-based methods:
- Chow-Lin
- Litterman
For example usage, please see the examples.ipynb
notebook. tsdisagg
depends heavily on pandas
to handle time reindexing, so the user is advised to read the associated Pandas documentation, especially as it relates to setting frequencies.
- Refactor codebase to use
statsmodels
model and results objects, as well as.fit()
api - Add missing interpolation methods relative to
timedisagg
(Fernandez, min RSS objective functions) - Add support for finer time frequencies (weekly, daily, hourly)