I've built an ARIMA model for an electricity pricing forecast that gives a 24-hour prediction using 17 training days. The model automatically picks its parameters based on a minimal AIC score.
I've tested it for a whole year and generally, the average error over a month is less than 7 dollars. (which was my benchmark), oftentimes it's even close to 2. But on certain isolated days, whether it's end of April or mid summer or the cold season, it jumps to almost 200$. In the data containing the real prices nothing unexpected happens and the prices gravitate around the same range of values. Why is that?