Submitted by **AttentionImaginary54** t3_102mf6v
in **MachineLearning**

I recently came across this paper Are Transformers Effective for Time Series Forecasting? and it seems to cast doubt on the recent trend of using transformers for time series forecasting, suggesting a simple model can out perform complex transformers.

Personally, in many of my experiments using transformers on temporal data besides the quite commonly tested benchmarks (ETH, exchange, etc) they perform poorly compared to other simple(r) models like GRUs or DA-RNN. Yet we are still seeing an explosion of papers about them in the research community. Are there other recent deep learning based alternatives?

eamonnkeoght1_j2vpjfb wroteWhy assume that ANY deep learning is useful for time series forecasting ?

​

Do We Really Need Deep Learning Models for Time Series Forecasting? https://arxiv.org/abs/2101.02118

Deep Learning for Road Traffic Forecasting: Does it Make a Difference?

https://arxiv.org/abs/2012.02260