Price Forecasting of Japan Electric Power Exchange using Time-varying AR Model

In this article, we built a state space model to analyze the price time series in Japan electric power exchange(JEPX) spot market. In building the model, we aimed to achieve the following two goals that the model was able to a) forecast prices with reasonable accuracy, and b) understand the underlying market dynamics by decomposing the price time series into a reasonable set of contributing factors. To capture the time-variability of the contributing factors to price, self-AR(autoregressive) process was introduced to allow continuous change in the magnitude of influence from each explanatory variable. To estimate the model, Kalman Filter algorithm was applied for stepwise recursive estimation. After optimizing the model under the maximum likelihood method(MLM) coupled with minimum AIC(Akaike information criteria) conditions, the model was able to decompose the 15:00-15:30 JEPX spot electricity strip price into a couple of the most contributing factors with significant time-dependencies. Our model also yielded as good a forecasting accuracy with conventional AR econometric model estimated with ordinary least square method(OLS), with a squared error of about 1.12 [yen/kWh] per forecasting period.