Symbolic Regression Time Series Operator Question

[qizzhao]

Lets said I have a dataset which contains N features data in T days. My X is a matrix with dimension (T, N) , and Y is a vector with length T. There are some basic time-series functions below:

ma( X, d ): d is a random generated constant number. Calculating moving average in the past d days of each feature n in X. X[t, n] = mean( X[t-d+1:t, n] )

delay( X, d ): d is a random generated constant number. Replacing X with the past d day value in each feature n. X[t, n] = X[t-d+1, n]

argmax( X, d): d is a random generated constant number. Taking the largest number in the past d day of each feature n. X[t, n] = max( X[t-d+1:t, n] )

I would like to put those functions above into Symbolic Regression model training and input as custom operators.
Is that possible to achieve that? and how? Give a short example please. Thank you.

[qizzhao]

Could anyone help to answer this question above. I know how to create the binary operator and unary operator. But none of them use rolling data.

[qizzhao]

Could anyone help to answer this question above. I know how to create the binary operator and unary operator. But none of them use rolling data.

[MilesCranmer]

Sorry for the late reply. Have you checked the other discussions? There are a couple about implementing the roll operator.

[qizzhao]

Thank you for the reply. I found that this post is similar to what I expected: #426. Going to study it.
But it only show the case that rolling unary operator. What if I need a rolling binary operator which input is the combination of feature value and random generated integer.