Large Financial Organization
CIB Group of a Financial Institution was looking for a way to process an increasing number of time series analyses.
The company performs complex analysis of regular time series: forecasting, outlier detection and removal, gap filling, trend analysis, pattern detection, etc. The initial objective is to handle 10-100-1000-time series per quarter (potentially 1M time series per quarter). Increasing data volume leaves no deadline guarantees. Manual analysis and decision making was based on expert judgment and not formalized. Selection of methods and algorithms, as well as setting of their parameters, was poorly justified.
We have proposed a mixed workflow, which includes a manual Task analysis phase and highly automated “Main Run”/” Self-Adoption”/” Knowledge Base update” phases.
An analyst is responsible for criterion definition and adjustment of the methods and algorithms used. Application and evaluation of methods/algorithms is fully automated and is performed by the system. Upon processing every batch, the system updates its own Knowledge Base, which is later used for selection of best algorithms based on their previous behavior with regards to detected properties of time series.
- Analysis performance (gap filling, outlier detection, forecasting) was increased by 120%.
- A comparable increase in the number of processed time series and reduction of the responsible department (~1.5 times both).
- Lower count of reports rejected by internal audit due to clearer explanations of decisions (~20% decrease in the number of rejected reports and improvement actions raised).
- Quality does not entirely depend on expert judgment, which leads to easier resource reallocation between subject areas of the business
Technology Set: WEKA, R, MatLab, Java Classification, Perceptron, Polynomial NN Cross validation, Time Series Analysis