Glossary of economics research

Results of search for cross-validation follow:

cross-validation: A way of choosing the window width for a kernel estimation. The method is to select, from a set of possible window widths, one that minimizes the sum of errors made in predicting each data point by using kernel regression on the others.

Formally, let J be the number of data points, j an index to each one, from one to J, yj the dependent variable for each j, Xj the independent variables for that j, Yj the dependent variable for that j, and {hi} for i=1 to n the set of candidate window widths. The hi's might be a set of equally spaced values on a grid. The algorithm for choosing one of the hi's is:

For each candidate window width hi
{
..For each j from 1 to J
..{
....Drop the data point (Xj, Yj) from the sample temporarily
....Run a kernel regression to estimate Yj using the remaining X's and Y's
....Keep track of the square of the error made in that prediction
..}
..Sum the squares of the errors for every j to get a score for candidate window width hi
..Record that in a list as the score for hi
}
Select as the outcome h of this algorithm the hi with the lowest score

The grid approach is necessary because the problem is not concave. Otherwise one might try a simpler maximization e.g., with the first order conditions.
Note however that a complete execution of the cross-validation method can be very slow because it requires as many kernel regressions as there are data points. E.g. in this author's experience, the cross-validation computation for one window width on 500 data points on a Pentium-90 in Gauss took about five seconds, 1000 data points took circa seventeen seconds, but for 15000 data points it took an hour. (Then it takes another hour to check another window width; so even the very simplest choice, between two window widths, takes two hours.)

Source: Hardle, 1990
Contexts: nonparametrics; estimation; econometrics; statistics


Back to top

Type 'help' for a list of relevant categories. Use * as a general wild card or ? for one character.

Feedback to econterms at econterms.com