Advanced data analysis in data-warehouse systems involves increasingly sophisticated statistical methods that go well beyond the rollup and drilldown over simple aggregates of traditional BI. In this context, time series forecasting is an important instrument as it is crucial for decision making in many domains. Reasonable forecasts require the specification of a stochastic model that captures the dependency of future on past values.
Currently, a common workflow for forecasting consists of exporting the data to statistical tools, like Matlab or R, and choosing and applying forecast methods externally. However, pushing computation directly to the data has several advantages. First, the knowledge encoded in models by domain experts can be stored and reused. This allows other users, especially those not trained in forecasting, to benefit from forecasts based on these models. Second, the forecast itself can be used inside the DBMS and linked with other source data (e.g., through joins). Third, the separation of conceptional and physical layer in a DBMS opens a wide variety of optimization potential.
The FFQ project (Flash Forward Query Framework) has the goal to integrate model-based forecasting natively into a DBMS. It offers a simple interface to make forecasting usable for any database user. Within this project different topics are addressed that aim to speed up the forecasting approach inside the database while providing a high accuracy, e.g., automatic forecasting, optimization of forecast queries and maintenance of forecast models.