Purchase and Redemption Prediction Based on Multi-task Gaussian Process and Dimensionality Reduction
Predicting future cash flows based on historical data is an essential and hard problem of financial business. Most of the previous works have attempted to convert cash flows into time series to make prediction. However, real-valued datasets are mostly multidimensional sources with complex curves, large amplitude and high frequency. The handful of research efforts that consider those truths have met with limited success. This paper proposes an algorithm based on Multi-task Gaussian Process model to predict cash flows in funds. Purchase refers to cash inflow, while redemption refers to cash outflow. MTGP can learn the correlation within multiple time series and make regression on each time series simultaneously. Furthermore, motif discovery is used for dimensionality reduction in data before MTGP to improve the accuracy. Experimental results on real-world data demonstrate the advantages of our proposed algorithm.
KeywordsTime series Gaussian process Motif discovery Prediction
This work is supported by Tianjin Municipal Science and Technology Commission under Grant No. 14JCQNJC00200, and the Research Foundation of Ministry of Education and China Mobile under Grant No. MCM20150507.
- 2.Wu, H., Salzberg, B., Zhang, D.: Matching, online event-driven subsequence over financial data streams. In: SIGMOD Conference, pp. 23–34 (2004)Google Scholar
- 3.Wu, H.: Personal email communication (2005)Google Scholar
- 4.Bonilla, E.V., Chai, K.M.A., Williams, C.K.: Multitask gaussian process prediction. In: Advances in Neural Information Processing Systems, 153–160 (2007)Google Scholar
- 5.Purchase_Redemption_Forecasts. https://tianchi.aliyun.com/datalab/index.htm