The Exponentially Weighted Signature
Alexandre Bloch, Samuel N. Cohen, Terry Lyons, et al.
This paper introduces the Exponentially Weighted Signature (EWS), an improved mathematical tool for analyzing sequences and paths that can now give more weight to recent information rather than treating all past equally. Unlike standard signatures, the EWS uses advanced mathematical operators to capture complex memory patterns (like oscillations or growth) while maintaining computational efficiency for machine learning. The authors prove this method generalizes several existing techniques and demonstrate it performs better than standard signatures on prediction tasks.
signaturestime seriesdifferential equationsrepresentation learning