Scaling of Forecasting Models

Challenge

The development of complex machine learning models often occurs in a sandbox, an isolated development environment. After successful testing, these models are implemented into the existing IT landscape to be executed regularly. Our client, an international logistics company, faced the challenge of needing to forecast demand for more than 10,000 entities, which made daily forecasting seem impossible.

Approach

To handle the daily computational load, we first developed a custom R package for the client, containing the necessary functions and models, centrally managed and maintained. We then programmed an automated process that uses parallelization to execute over 10,000 models simultaneously on a multicore server environment. Multiple models were trained simultaneously, providing demand forecasts for each customer for the next business day. Thanks to the client's powerful infrastructure, up to 20 models could be trained in parallel.

Result

Based on the parallelized implementation of the models, our client can now predict daily demand for all relevant customers and entities. The parallel execution of model estimations reduced computation time by a factor of 15, making it feasible to estimate all entities.