Skip to content

Latest commit

 

History

History
25 lines (19 loc) · 1.47 KB

README.md

File metadata and controls

25 lines (19 loc) · 1.47 KB

TimeSeriesES-Cell

This repository contains code for time series analysis, developed in the paper:
Time Series Using Exponential Smoothing Cells, by Avner Abrami, Aleksandr Aravkin, and Younghun Kim.

Overview

Exponential smoothing (ES) techniques such as the Holt-Winters model, break down in challenging situations, including

  • high level of noise
  • large or frequent outliers
  • significant portions of missing data
  • nonstationary features.

We developed a global approach for fitting time series, formulated as a convex optimization problem. The approach is built using the notion of linked ES cells, equipped with robust losses for outlier detection and denoising. The links enforce time series structure but allow non-stationary signals.

Experiments

The code reproduces two experiments from the paper:

Optimization

The problem is currently built and solved using cvx.py, which does not exploit special problem structure. A custom solver that does is under development.