A State Space Filtering-Based Approach for Price Prediction

Abstract

We present a method of the forecasting and the data filtering of a linear dynamic system based on the dimension reduction of the space of unobservable states. The method relies on the singular value decomposition of the Hankel matrix. The decomposition is used to calculate unknown parameters of the model. The elements of the singular value decomposition are separated into blocks enabling to estimate the initial state and the system matrices and predict the system dynamics and the data filtering by identifying exponential trends and periods of seasonal fluctuations.

To illustrate the quality of fitting and the determined periods of an oscillatory system with trends and the white noise, we conducted numerical simulations of such systems. The parameter estimates were obtained with high precision. Then, daily electricity price data from the NordPool system from 2016 to 2020 were used to generate in-sample and out-of-sample forecasts.

The advantages of the proposed method include the ability to handle ill-conditioned matrices and to determine the periods of oscillatory systems. This is significant due to the presence of seasonality in many economic indicators. In the analyzed daily electricity price data, the method identified the presence of biweekly and monthly seasonality.