Pejman Khadivi, Prithwish Chakraborty, Naren Ramakrishnan
Developing a precise understanding of the dynamic behavior of time series is crucial for the success of forecasting techniques. We introduce a novel communication-theoretic framework for modeling and forecasting time series. In particular, the observed time series is modeled as the output of a noisy communication system with the input as the future values of time series. We use a data-driven probabilistic approach to estimate the unknown parameters of the system which in turn is used for forecasting. We also develop an extension of the proposed framework together with a filtering algorithm to account for the noise and heterogeneity in the quality of time series. Experimental results demonstrate the effectiveness of this approach.
- Date of publication:
- September 17, 2015
- IEEE International Workshop on Machine Learning for Signal Processing (MLSP)