tsar is a Python library to model and forecast time series data. It operates on pandas dataframes and uses numba just-in-time compilation to speed up some operations.

The core algorithm used is described in the draft paper Seasonally-Adjusted Auto-Regression of Vector Time Series, E. Busseti, November 2019.

Note: tsar is currently in beta release, its interface might change.

To install, execute in a terminal

pip install tsar

To use it, you need some time series data as a pandas dataframe data with date-time index and constant time spacing (i.e., its data.index.freq attribute must not be None). Only numerical data is supported at this time.

You build a model by

from tsar import tsar
model = tsar(data=data, P=P, F=F)

where P and F are positive integers representing how many points in the past will be used for inference, and how many steps in the future will be predicted, respectively.

If you wish, you can pass to the constructor any hyper-parameter of the model, for example, the rank R and the quadratic regularization parameter quadratic_regularization. If not specified, these will be optimized by greedy grid search (see the paper) on an internally split test set.

The data can have any amount of missing values (np.nan).

Inference is performed with

prediction = model.predict(data=new_data, prediction_time=t)

where new_data is a dataframe with the same column names and time spacing than the one used to build the model, and prediction_time is the timestamp at which to infer. The resulting prediction dataframe has time index spanning from the time t minus P steps in the past, to the time t plus F steps in the future.

Any data that was present (and not equal to np.nan) in the new_data dataframe will be copied to the appropriate position in the prediction dataframe. All other points will be inferred.

The provided data can have any amount of missing values, the prediction has none.