1
$\begingroup$

I am doing time series forecasting with neural network (feedforward for now, but I will test also RNNs) and my problem is that, even though the network learned general patterns, it doesn't forecast well sudden peaks in the data. Example: enter image description here

Blue is real data and orange is forecasted. Data was standard scaled and RMSE was used.

From my experiments I noticed that MAE is worse than RMSE, probably due to RMSE giving extra high loss for the wrongly predicted high values due to squaring them.

My guess is that the network might be 'shy' to forecast big peaks, because if it will forecast high peak wrongly it will get very high penalty.

Is there any technique that might help here or a better metric for this use case to help forecast high peaks? Could giving higher weights to units with high values help without hurting overall performance?

Improve this question
$\endgroup$
1
  • $\begingroup$ The target has probably skewed distribution. RNN predicts the expected value of the distribution. Possible solutions is to transform the target to make it more symmetric, or add more informative variables to explain more variation and reduce the variance of the residuals distribution. I don't think you can do much just by changing the model structure. $\endgroup$ Commented Feb 6, 2023 at 20:48

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.