I have developed different kinds of RNNs (such as LSTM,GRU etc.)to predict future values of thermocouple measurements. The residual errors look like they do not follow normal distribution, so I wanted to explore their skewness and kurtosis values. Can I somehow correlate the values of skewness and kurtosis (for example if skewness is positive or negative and if kurtosis is more or less than 3) with the performance of the model and the predictions?

Thank you!

  • 1
    $\begingroup$ Why are you checking for normality in the residuals given that you are not using linear regression? $\endgroup$
    – utobi
    Commented May 26, 2023 at 15:24
  • $\begingroup$ If kurtosis is high, you have an outlier problem, and that can affect the accuracy of any least squares method. Also, if your residuals are non- normal, then the usual $\pm 2$rmse prediction bounds will be inaccurate, depending on both skewness and kurtosis. $\endgroup$ Commented May 26, 2023 at 20:23


Your Answer

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