The non-parametric regression model to be estimated looks like the following

x_t = b(x_t-1) + epsilon_t

Forfinding the optimal bandwith h in the kernel regression a cross-validation method (leave a point out and estimate it with the rest of the sample) is used and evaluated via the expected prediction error (EPE). This takes a while since the kernel has to be calculated each time for a range of potential h-values and every sample without x_i.

Therefore the use of FFT (Paper: link and similar post: link) can speed up the calculations. In my understanding the following is the way to go:

  1. Discretize the data (x_t-1 values) by using linear binning
  2. Take the Fourier-transform of the binned data
  3. Multiply the transformed binned data with the Fourier-Transform of the Gaussian-kernel
  4. Inverse-transform the product back to get the kernel estimation

But I fail to understand how implement this since once the data is binned (I assume here x_t-1) the connenction to the y-values of the kernel regression (here x_t) is broken?


1 Answer 1


In case anyone else is also watching:

A nice explanation is given by Ichimura and Todd (2007) in their handbook chapter, "Implementing Nonparametric and Semiparametric Estimators," chapter 8.

Basically, when you "bin" your data, you are chopping the x range up into intervals, $[t_k; t_{k+1}]$. For each such interval, you construct one new y value which averages among the observations that fall in the bin. But there's a particular weighting scheme which needs to be followed, including the distance of the bins, which I believe has to be uniform (so that $t_{k+1} - t_k = \delta$ for all $k$). So I encourage you to check out the paper.


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.