We want to predict Y based on some function of X, i.e., Y_hat = f(X). How can we show that the conditional expectation f*(X) = E(Y|X) is the mean-square optimal predictor, i.e., the function f* solves the minimization problem below?

min E{[Y - f(X)]^2}

Here the unrestricted mean-square optimal predictor is the conditional mean.

  • $\begingroup$ It does not matter whether your conditional distribution is conditional on time series data, or on other predictors, or on nothing at all. The (conditional) expectation minimizes the expected square error for any (conditional) distribution, and the textbook proof works straightforwardly. $\endgroup$ May 9, 2021 at 16:52
  • $\begingroup$ @StephanKolassa Could you please refer me to some textbook where I could have a look at the proof? Would really appreciate. $\endgroup$
    – user321152
    May 9, 2021 at 17:00
  • $\begingroup$ This thread gives one possible proof, you just need to translate the symbols to your case. $\endgroup$ May 9, 2021 at 17:05
  • $\begingroup$ Actually, this thread is much more informative. $\endgroup$ May 10, 2021 at 19:20


Browse other questions tagged or ask your own question.