# Recurrent Neural Networks for Time Series Forecasting: Current Status and Future Directions

**Authors:** Hansika Hewamalage, Christoph Bergmeir, Kasun Bandara  
**arXiv:** [1909.00590](https://arxiv.org/abs/1909.00590)  
**Date:** 2019-09-02  
**Summary page:** https://www.summarizepaper.com/en/paper/1909.00590v5/

## Key points

- Recurrent Neural Networks (RNN) are competitive in time series forecasting, as shown by recent success in the M4 competition.
- Traditional statistical models like ETS and ARIMA are still popular due to their high accuracy and user-friendly nature.
- RNNs have room for improvement in terms of robustness, efficiency, and automation.
- An extensive empirical study and open-source software framework have been presented to explore the capabilities of RNNs in forecasting.
- RNNs can effectively model seasonality directly if the dataset exhibits consistent seasonal patterns; otherwise, a deseasonalization step is recommended.
- Comparative analyses show that (semi-)automatic RNN models offer competitive alternatives in many forecasting scenarios but are not a one-size-fits-all solution.
- Understanding the strengths and limitations of different forecasting methods is crucial for selecting an appropriate model for a given dataset.

## Summary

In their paper titled "Recurrent Neural Networks for Time Series Forecasting: Current Status and Future Directions," authors Hansika Hewamalage, Christoph Bergmeir, and Kasun Bandara discuss the competitive nature of Recurrent Neural Networks (RNN) in forecasting. The recent success of RNNs in the M4 competition highlights their potential as a powerful tool for time series forecasting. However, traditional statistical models like ETS and ARIMA remain popular due to their high accuracy and user-friendly nature. Despite this, RNNs still have room for improvement in terms of robustness, efficiency, and automation. To address these concerns and further explore the capabilities of RNNs in forecasting, the authors present an extensive empirical study along with an open-source software framework that focuses on existing RNN architectures. Through this research, they aim to establish guidelines and best practices for utilizing RNNs effectively. One key finding is that RNNs can effectively model seasonality directly if the dataset exhibits consistent seasonal patterns. However, in cases where the seasonal patterns are not uniform, a deseasonalization step is recommended. Comparative analyses against ETS and ARIMA reveal that while (semi-)automatic RNN models are not a one-size-fits-all solution, they do offer competitive alternatives in many forecasting scenarios. This study emphasizes the importance of understanding the strengths and limitations of different forecasting methods to make informed decisions when selecting an appropriate model for a given dataset. Overall, this comprehensive examination sheds light on the current state of RNNs in time series forecasting and provides insights into potential future directions for further advancements in this field.