LSTM versus CNN for time series data

Hi mentors.

  1. I have a general question WHY one might use a CNN rather than LSTM for time series sequence data? I have seem some CNN implementations for financial time series.

  2. Is LSTM always preferred as it’s more flexible than CNN for handling time series data? Or do CNN have additional benefits in capturing other dimensional features?

  3. Also, am I correct to say the use of skip connections on CNN is similar to using Cell memory LSTM process. But maybe LSTM has a more flexible updating process.

Thanks in Advance, Jason

Hi Jasonwtli,

CNNs can extract features of the time-series data, whereas LSTM layers can identify short-term and long-term dependencies. So they can also be combined, as discussed in this paper. Skip connections can be used in LSTMs to further tackle long-term dependencies, as discussed in this article.


Thanks very much Reinoud and pointing to relevant articles. They should give me ideas to consider in implementing LSTM architectures. I guess we just to need to experiment with different architectures like transformers that incorporate sequence model + CNN style features.