Time Series Forecasting Engine

Challenge

The prediction of time series is one of the most common use cases in the data science field. Many data sets exhibit a temporal order that can be leveraged for forecasting through appropriate models. Despite the heterogeneity of time series data, the process of analysis and forecasting is usually quite similar—and relatively labor-intensive. Our client, an international automobile manufacturer, faced the problem of needing to forecast many different time series.

Approach

Correctly forecasting time series is methodologically challenging and requires various sequential steps: preparation of raw data (e.g., removing outliers, detecting structural breaks), creation of a selection of informative influencing factors (also known as feature engineering and feature selection), training and selection of various time series or machine learning models (such as ARIMA(X), linear or tree-based methods), and validation and interpretation of the resulting models. We addressed all these steps for our client by developing a standardized forecasting engine. This engine can now be applied automatically and scalably to an unlimited number of forecasting challenges.

Result

By fully automating the modeling process, the time required to develop an initial model (time-to-model) was reduced to an absolute minimum. This enabled our client to implement a series of demanding forecasting use cases in a very short time.