NNPF: Neural Network Particle Filter for Time Series Data
A novel particle filter based on a neural network for the analysis of volatile time series data.
This paper proposed a novel particle filter to fit structural time series models to volatile time series. Instead of assuming a parametric distribution for the observation equation, a neural network is applied to estimate the density of the observation equation in non-parametric way. This density is used in a particle filter to fit a state space model. With a simulation it is shown that this method outperforms the Kalman filter in the case of volatile time series with heavy tailed distributions for the observation equation. The method is illustrated with a real-life application to road sensor data.
