I have a model that uses autoregressive multidimensional time series to make a prediction.

I use a smoothed X’X and a smoothed X’Y (of the recent past) to do an in-sample regression: B = (X’X)^-1 X’Y. When I create a prediction F = X B, I then determine R2 in the recent past by looking at the smoothed F’F the smoothed F’Y_test. In the single-dimensional setting R2 = (2F’Y_test - F’F) / Y_test’Y_test. And the optimal gain = (F’F)^-1 F’Y_test. It occurred to me that I could use this R2 matrix to update the gains on my model: B = (X’X)^-1 X’Y (F’F)^-1 F’Y_test. And I could do it even when the data is multidimensional.

Is this a common technique? Does it have a name? You could even apply this technique recursively. It feels like a Kalman filter or like back-prop, but I don’t fully understand the relations.



Your Answer

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

Browse other questions tagged or ask your own question.