I have a problem where I have transaction data for many banking accounts. The task is to train a model on historical debit/expense transactions and then forecast expense transactions for the next n days.

The data for each account is aggregated daily - I have total expense values for each account per day. I have done some research and found that the ConvLSTM model (proposed here) is powerful for these multistep ahead forecasts. A example of this model to forecast the next 7 days of power consumption of a household is shown at the bottom of this tutorial.

I am struggling to understand how the convolutional unit is used within the LSTM unit to make these forecasts. Specifically I am having trouble understanding the shape that my input should be in to feed it into TensorFlow's ConvLSTMCell (I think I will need to use the Conv1DLSTMCell sub-class?) and also what format the output/predictions will be returned as.

  • $\begingroup$ You should have seen that this method is developed for spatiotemporal data... What is your spatial information? I'm wondering if you should go for regular LSTM... $\endgroup$ – ignatius Dec 20 '18 at 11:20
  • $\begingroup$ @ignatius That's what I thought, but in the tutorial I provided, he uses it just for regular time series data. I also asked this question yesterday datascience.stackexchange.com/questions/42872/… where another user recommended that i use ConvLSTM $\endgroup$ – KOB Dec 20 '18 at 11:22
  • $\begingroup$ Please be sure to update the question with your progress, I am very interested in this topic and would be curious to see your progress, for this questions and your linked one. $\endgroup$ – Aesir Dec 20 '18 at 11:29
  • $\begingroup$ Can you write a table/plot of your data, to see if it is 2D? $\endgroup$ – ignatius Dec 20 '18 at 12:23

Your Answer

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

Browse other questions tagged or ask your own question.