I have a regression task (to predict price for finanical market.)

I build a mlp to do the regression.

I found mlp will stop at giving a constant prediction, which i think it's useless.

Does this mean, I can't do a better prediction than constant value? but I tried catboost, it give a meaningful prediction.

So, if the mlp give the constant value, is there any common reason? and how to improve that?

  • $\begingroup$ If Catboost works then the problem is not (all) in the data. It's probably a mistake you made in buiding your Neural Network. Could you provide information on how your dataset and the kind of MLP and its hyperparameters you built? $\endgroup$
    – Leevo
    May 10, 2021 at 7:18
  • $\begingroup$ Can you please elaborate a bit on the problem you are trying to solve, what type of data and network architecture you re using so that we can help you more? $\endgroup$
    – hH1sG0n3
    Nov 4, 2021 at 12:26

1 Answer 1


This means that your features do ot have predictive power to estimayevthe target. Hence the MLP has learnt to predict the mean of the target distribution so as to minimize the loss.

Try checking if the constant value is indeed the mean of the target variable.

  • $\begingroup$ no, not mean value. and some of features have predictive power, because i tried them in catboost. but i do mix some noisy features in. it's odd that they mess up the whole model. $\endgroup$
    – nick
    May 10, 2021 at 6:03

Your Answer

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

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