1
$\begingroup$

I have recently started learning time series forecasting. I have a dataset of the weekly payment history of 10k clients over 1 year, and I want to predict the future 5 payments for a test set of 1k clients.

From what I have tried, I've found that using LSTMs instead of a simple MLP doesn't improve the prediction as much as I anticipated. My understanding is that LSTMs captures the relations between time steps, whereas simple MLPs treat each time step as a separated feature (doesn't take succession into consideration).

So, my question is: why doesn't the LSTM model improve the forecasting significantly? What are the best models for such a task, given that the time series are short (maximum sequence's length = 52)?

$\endgroup$
2
  • $\begingroup$ I would suggest using statistical methods like ARIMA even before machine learning. $\endgroup$ Jul 11 at 16:39
  • $\begingroup$ I have tried VARMA, since here I have multiple Time Series. But, the RMSE was very high. $\endgroup$ Jul 13 at 16:46
1
$\begingroup$

RNNs are known to be superior to MLP in case of sequential data, like yours. But complex models like LSTM and GRU require a lot of data to achieve their potential. I don't know about your data but you can try to validate your architecture, approach and overall setting using a different, known time-series benchmark data. Maybe something is wrong with architecture, loss function, data, etc... So trying a different but known benchmark data can give you an idea about why you are unable to produce superior results with LSTM.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.