1
$\begingroup$

I have a data set containing hourly electricity prices for since 1.01.19 until September. Since the process turned out to be (weakly) stationary, I applied an ARIMA model in Python in order to predict the prices for the next day.

It turned out that the best prediction was made using the last two days as historic data and the worst was the one using almost 6000 values.

What is a possible explanation for this happening?

$\endgroup$
1
$\begingroup$

Electricity prices are essentially the same as stock prices: best modelled by a random walk, where the best prediction for tomorrow is the price today. Therefore I am not really surprised that you get worse results using more historical data. Some versions of ARIMA will also include regularisation, which will punish your model for including more and more data - inclusion of new data must be justified by contributing to a lower residual error to be "worth" inclusion.

Other models that tend to be a little more robust use features other than the actual target, here the price. For example, trying to predict the volatility of the price might prove to be more accurate. For this there are GARCH models (Generalised AutoRegressive Conditional Heteroskedasticity).

Another thing you might consider is to include external data... for example, electricity consumption is heavily influenced by the weather - if it is cold outside, a lot of people heat their homes using electrical heaters, they also drink more hot drinks etc.

|improve this answer|||||
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.