Abstract: Historical anaesthetist: An Ignored but Powerful Baseline for Long Sequence time-series Forecasting) is a short article published by Huawei Cloud Database Innovation Lab and Data and Intelligence Lab of University of Electronic Science and Technology of China. This paper proposes a baseline for long time series.

This article was shared from Huawei cloud community “CIKM’21 Historical andodotic analysis”, author: Cloud database innovation Lab.

takeaway

In this paper, Historical Inertia: An Ignored but Powerful Baseline for Long Sequence time-series Forecasting) is a short article published by Huawei Cloud Database Innovation Lab and Data and Intelligence Lab of University of Electronic Science and Technology of China. This paper proposes a baseline for long time series. CIKM is one of the top academic conferences in the field of information retrieval and data mining. A total of 626 essays were submitted for this conference, including 177 accepted papers, with an acceptance rate of about 28%. This paper is one of the key technical achievements of cloud database innovation LAB in temporal analysis.

1, the

Long Sequence time-series Forecasting (LSTF) is becoming more and more popular due to its wide application. Although a large number of complex models have been proposed to improve the effectiveness and efficiency of prediction, one of the most natural and basic characteristics of time series has been ignored or underestimated: historical inertia. In this paper, we propose a new LSTF baseline, Historical Inertia (HI). In this baseline model, we directly take the historical data point closest to the predicted target in the input time series as the predicted value. We evaluated the effect of HI on four open LSTF datasets and two LSTF tasks and found that up to 82% relative improvement was achieved in HI compared to SOTA. We also discuss the possibility of combining HI with existing methods.

2, the HI

HI directly takes the historical data point closest to the predicted target in the input time series as the predicted value.

3, the experimental

3.1 Prediction results of univariate long time series

For univariate long time series prediction tasks, HI is significantly superior to SOTA model on ETTh1 and ETTm1 datasets. Informer and its variants dominate the optimal results of the ETTh2 dataset. For Electricity datasets, HI, Informer and DeepAR all perform well. Overall, HI achieved a relative increase of up to 80% in MSE and 58% in MAE.

3.2 Prediction results of multivariable long time series

For multivariable long time series prediction tasks, HI was significantly better than SOTA model in most of the four data sets, with a relative improvement of up to 82%.

4, discuss

4.1 Why does HI have such a good effect

We consider the reasons why HI can achieve good results from two perspectives:

  • Numerical value: HI guarantees that the predicted sequence has a similar numerical size to the real sequence.

  • Periodicity: For periodicity and short periodicity data sets, HI can achieve the phase similarity between the predicted sequence and the real sequence.

4.2 How to use HI

We propose two possible directions for utilizing HI

  • Hybrid model: HI can be combined with other models, such as simply weighting and averaging the output results as a trick.

  • Automatic machine learning (AutoML) : Complex models may not work well in some cases, so consider reducing/increasing model complexity based on data adaptive model structure.

As for the direction of fusion model, we designed a simple experiment to verify: average the output results of HI and 2-layer MLP model to obtain the final prediction results. The experimental results show that the HI integrated MLP model can achieve more accurate prediction, and this advantage is more significant in the univariate long time series prediction task.

Huawei cloud database innovation lab’s official website: www.huaweicloud.com/lab/clouddb…

Click to follow, the first time to learn about Huawei cloud fresh technology ~