Training and Storing System1’s Time Series Models

Building and optimizing complex models is becoming relatively easy due to the growing set of tools available to data scientists. These tools can be used together with hyperparameter optimization methods, cross validation, integration tests to perform scoring operations for recommendation systems. This paper describes how we train forecasting models on time series at System1, focusing on reproducibility, data standardization, model validation, logging, storing the actual models, and serving live recommendations.