A dynamic model for short-term prediction of stream attributes

Multiple data streams coming out of a complex system form the observable state of the system. The streams may correspond to various sensors attached with the system or outcome of internal processes. Such stream data may consist of multiple attributes and may differ in terms of their frequency of generation and observation. The streams may have dependency among themselves. One will have to rely on such data streams for monitoring the health of the system or to take any corrective measure. Predicting the value of certain stream data is an important task that can help one to take decision and act accordingly. In this work, a simple but generic visualization of a complex system is presented and thereafter a linear regression-based dynamic model for short-term prediction is proposed. The model is based on the past history of the attributes of multiple streams as suggested by the domain experts. But, it automatically determines the meaningful attributes and reformulates the model. The model is also re-computed if the prediction error exceeds the allowable tolerance. All these make the model dynamic. Experiment is carried out with stock market data streams to predict the close value well in advance. It is observed that in terms of quality of prediction and performance metric, the proposed model is quite effective.