# Mlinear: Rethink the Linear Model for Time-series Forecasting

**Authors:** Jianing Chen, Chuhao Chen, Xiangxu Meng  
**arXiv:** [2305.04800](https://arxiv.org/abs/2305.04800)  
**Date:** 2023-05-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.04800v1/

## Key points

- Significant advancements in time-series forecasting research
- Growing emphasis on analyzing inherent characteristics of time-series data
- Proposal of an efficient time series forecasting model called Mlinear
- Mlinear model consists of two core components: integration of single-channel and multi-channel data for joint forecasting, and a novel loss function
- Experiments conducted on benchmark time series datasets to evaluate performance
- Mlinear model outperforms current state-of-the-art models, achieves 10 times speedup, and has fewer parameters compared to the latest SOTA model
- Contribution to the field by providing an efficient and accurate linear model-based approach
- Implications for applications such as finance, weather prediction, and resource planning.

## Summary

In recent years, there have been significant advancements in time-series forecasting research, with a growing emphasis on analyzing the inherent characteristics of time-series data rather than solely focusing on designing forecasting models. In line with this trend, the authors of this paper propose an efficient time series forecasting model based on linear models called Mlinear. The Mlinear model consists of two important core components. Firstly, it integrates different semantics brought by single-channel and multi-channel data for joint forecasting. This means that the model can effectively leverage the information from both types of data to improve forecasting accuracy. Secondly, the authors introduce a novel loss function that replaces the traditional Mean Squared Error (MSE) loss and Mean Absolute Error (MAE) loss commonly used in time-series forecasting. This new loss function aims to achieve higher accuracy in forecasting. To evaluate the performance of their proposed model, the authors conducted experiments on widely-used benchmark time series datasets. The results demonstrate that the Mlinear model not only outperforms the current state-of-the-art (SOTA) models but also achieves a 10 times speedup and has fewer parameters compared to the latest SOTA model. Overall, this paper contributes to the field of time-series forecasting by providing an efficient and accurate linear model-based approach. By integrating different semantics from single-channel and multi-channel data and introducing a novel loss function, the Mlinear model offers improved performance over existing methods. These findings have implications for various applications that rely on accurate time series forecasting such as finance, weather prediction, and resource planning.