Yes that makes sense and thanks for getting back to me. I run some tests and using differentiated prices as feature does slightly improve the results. I’ll try run diff on some other non stationary features and see how my model behaves then. Get Outlook for iOS<https://aka.ms/o0ukef>

…

________________________________ From: simaki ***@***.***> Sent: Saturday, May 21, 2022 4:08:32 AM To: fracdiff/fracdiff ***@***.***> Cc: aigarspetresevics ***@***.***>; Author ***@***.***> Subject: Re: [fracdiff/fracdiff] -d does not return back the original series (Issue #127) I personally don't take the slogan that fracdiff keeps the memory of time series at the face value. fracdiff does lose some memory in exchange for taming nonstationarity - the advantage is just that this exchange is relatively less than diff. the approximation is quite rough - you would need a quite wide window to keep, say, 90% of the memory. https://github.com/fracdiff/fracdiff/blob/main/fracdiff/sklearn/tol.py#L56-L59 also, fracdiff erases the additive scale of time series - fracdiffs of X and X+c are the same. thus it wouldn't always be feasible to "convert back to the original scale after prediction". i would use fracdiff as just a feature/signal — Reply to this email directly, view it on GitHub<#127 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ASW7CL4NIKJLJEHBC4GVX5DVLAZRBANCNFSM5V6D5OSQ>. You are receiving this because you authored the thread.Message ID: ***@***.***>

-d does not return back the original series #127

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions