# Analysis and modeling to forecast in time series: a systematic review

**Authors:** Fatoumata Dama, Christine Sinoquet  
**arXiv:** [2104.00164](https://arxiv.org/abs/2104.00164)  
**Date:** 2021-03-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.00164v1/

## Key points

- The paper provides a comprehensive overview of methods and models for time series analysis and modeling, with a focus on prediction.
- It covers various aspects of the time series analysis process, including decomposition, stationary tests, modeling, and forecasting.
- The authors adopt a unified presentation approach throughout the survey.
- Different frameworks for time series decomposition are presented to separate nonstationary effects from the stochastic component.
- Three popular linear models and five major nonlinear models are introduced.
- Deep learning approaches, particularly artificial neural networks, are discussed in detail for time series forecasting.
- Practical implementation details, including R and Python implementations, are provided.
- The paper aims to provide in-depth knowledge while covering both conventional approaches and recent adaptations of deep learning techniques.
- It consists of 65 pages with bibliographic references, 14 figures, and 6 tables.

## Summary

This paper, titled "Analysis and Modeling to Forecast in Time Series: A Systematic Review," is authored by Fatoumata Dama and Christine Sinoquet. The aim of this study is to provide a comprehensive overview of state-of-the-art methods and models dedicated to time series analysis and modeling, with a specific focus on prediction. The review covers various aspects of the time series analysis process, including time series decomposition, stationary tests, modeling, and forecasting. To facilitate understanding, the authors have adopted a unified presentation approach throughout the survey. They present decomposition frameworks for both linear and nonlinear time series models. The paper begins by exploring the relationship between stationarity and linearity in time series data. It then examines different methods used to test for weak stationarity. The authors present various frameworks for time series decomposition in a unified manner. These decomposition schemes aim to separate nonstationary effects (deterministic components) from the remaining stochastic component. Modeling the stochastic component accurately is crucial for ensuring prediction accuracy. The authors introduce three popular linear models along with two more flexible variants. They also present five major nonlinear models commonly used for time series analysis. Deep learning approaches, particularly artificial neural networks, hold a special place among nonlinear models due to their recent popularity. Therefore, the paper dedicates an entire section to deep learning-based time series forecasting. In addition to providing theoretical insights into different methods and models, the paper includes practical implementation details such as lists of R and Python implementations for the methods discussed throughout the review. Overall, this systematic review aims to provide in-depth knowledge while covering a broad range of well-established conventional approaches as well as more recent adaptations of deep learning techniques for time series forecasting. The paper consists of 65 pages (including 9 pages with bibliographic references) and includes 14 figures and 6 tables; it falls under categories csLG (Machine Learning), csAI (Artificial Intelligence), 68Txx (Artificial Intelligence), and I2:6 (Learning).