Building a Model for Automatic Time Series Prediction. Application for Bank Customer's Lifetime Value

Time series forecasting algorithms are essential in a banking system. Currently, there are many different approaches for time series forecasting, such as autoregressive models of various types, linear models, recurrent neural networks and gradient boosting based solutions. However, in the case of forecasting multiple time series, building an individual model for every specific time series is time-consuming and expensive. Therefore, the automatic algorithm is needed. One the one hand such algorithm should incorporate information from many time series, on the other hand, high data variability could be harmful to forecasting accuracy. Thus, we propose a pipeline which as a first step splits data into multiple clusters (time series inside each cluster are very similar) and secondly trains a separate linear model for each cluster. The algorithmic pipeline was implemented in Python and tested on various banking data.