The no-free-lunch theorem (Wolpert and McReady, 1992) states that no machine learning (or "artificial intelligence" if you want to sound trendy) method can be claimed to be universally better than all others. Therefore, you cannot expect a simple answer to your question.

In addition, both Recurrent neural networks (RNNs) and Multilayer Perceptrons are subsets of Artificial neural networks (ANNs). Therefore, the question does not make much sense. I suppose that you actually mean the following: for time series prediction, should one use Feedforward neural networks (FFNNs), of which Multilayer Perceptrons (MLPs) are a subset, or Recurrent neural networks (of which LSTMs are a subset)? This is an old question.

The following results were proven mathematically more than twenty years ago: if the noise present in the data is output noise, then recurrent networks are optimal (meaning that FFNNs will fail in that situation, irrespective of the training algorithm, the feature selection algorithm, etc.); by contrast, if the noise present in the data is state noise, then feedforward neural networks are optimal (meaning that RNNs will fail in that situation, irrespective of the training algorithm, the feature selection algorithm, etc.). Again, note that these are mathematically proven results, not results of a couple of more or less poorly designed benchmarks ; they can be easily verified on numerical examples. For more details on neural networks for forecasting and process modeling, and especially on your question, see for instance « Neural networks, methodology and applications », pp. 149-174 (Springer, 2005), and references therein.

According to previous works, LSTM works better for time series prediction.

The link below is a good tutorial to understand the difference between MLP and RNNs.

http://colah.github.io/posts/2015-08-Understanding-LSTMs/

Go for LSTM, well-suited for forecasting!

Try NARX . I think it's results are better for time series

You can try temporal convolutions over LSTM as it has been shown to be effective at processing temporal and sequential data. Sometimes more effective than plain LSTM.

More details can be found at this paper: Article WaveNet: A Generative Model for Raw Audio

LSTM will work for time series prediction

RNN can be used for prediction

It depends on your task. Lately, it has been shown that purely convolutional models (such as wavenet or TCNs) can outperform recurrent models such as LTSMs and RNNs. This paper has info, but might be too dense for beginners: https://arxiv.org/abs/1803.01271

The no-free-lunch theorem (Wolpert and McReady, 1992) states that no machine learning (or "artificial intelligence" if you want to sound trendy) method can be claimed to be universally better than all others. Therefore, you cannot expect a simple answer to your question.

In addition, both Recurrent neural networks (RNNs) and Multilayer Perceptrons are subsets of Artificial neural networks (ANNs). Therefore, the question does not make much sense. I suppose that you actually mean the following: for time series prediction, should one use Feedforward neural networks (FFNNs), of which Multilayer Perceptrons (MLPs) are a subset, or Recurrent neural networks (of which LSTMs are a subset)? This is an old question.

The following results were proven mathematically more than twenty years ago: if the noise present in the data is output noise, then recurrent networks are optimal (meaning that FFNNs will fail in that situation, irrespective of the training algorithm, the feature selection algorithm, etc.); by contrast, if the noise present in the data is state noise, then feedforward neural networks are optimal (meaning that RNNs will fail in that situation, irrespective of the training algorithm, the feature selection algorithm, etc.). Again, note that these are mathematically proven results, not results of a couple of more or less poorly designed benchmarks ; they can be easily verified on numerical examples. For more details on neural networks for forecasting and process modeling, and especially on your question, see for instance « Neural networks, methodology and applications », pp. 149-174 (Springer, 2005), and references therein.

One specific type of ANNs you can use for Time Series Analysis and Forecasting are the Multi-Layered Nets of Active Neurons (for more info see "A Framework for Developing Multi-Layered Networks of Active Neurons for Simulation Experiments and Model-Based Business Games Using Self-Organizing Data Mining with the Group Method of Data Handling" - in Simulation Gaming. Applications for Sustainable Cities and Smart Infrastructures, pp. 191-199, Springer 2018). To apply this approach you can use one of the leading software platforms in this area at KnowledgeMiner.eu (available for 30 days free trial ) - empirically proven to be one of the best, for many years!

One specific type of ANNs you can use for Time Series Analysis and Forecasting are the Multi-Layered Nets of Active Neurons (for more info see A Framework for Developing Multi-Layered Networks of Active Neurons for Simulation Experiments and Model-Based Business Games Using Self-Organizing Data Mining with the Group Method of Data Handling - in Simulation Gaming. Applications for Sustainable Cities and Smart Infrastructures, pp. 191-199, Springer 2018). To apply this approach you can use one of the leading software platforms in this area KnowledgeMiner (available for 30 days free trial ) - empirically proven to be one of the best, for many years!

RNN is a great technique for time series applications. But RNN suffers from long term dependencies (vanishing and exploding gradients).

LSTM is a great medication type for time series applications. It was originally designed to tackle long term dependencies.

Thanks everyone. All your contributions are well appreciated.