I have a dataset with sales numbers for ~500 different markets (assume different cities or regions) and need monthly forecasts for each market. Instead of building 500 different models, I'm interested in just training a single model with a single set of params to prevent me from having to manually feature engineer each market.

Can someone point me to some resources for this problem? I went through the similar Stack questions and they weren't as helpful for this particularly.

Specifically, I know that LSTMs are sometimes used here, but would classical approaches be effective as well?

  • $\begingroup$ I have asked a similar question before, the answer i got was LSTM is capable of doing that in its own. give it a try and plz update us with the results you have got. $\endgroup$
    – asmgx
    Aug 12, 2019 at 23:13

2 Answers 2


You can treat the markets as a categorical feature in a tree based (using decision tree as a weak leaner) ensemble model such as random forest or gradient boosting. Some applications are:


Before we go ahead and create a single model for all markets, please check below points once. If below all points are satisfied then we can easily create single model for all markets.

  1. Is the time series data has same amount of information(in terms of years)
  2. Check with stationary of all markets are same?
  3. Check with seasonality of all markets are same?

if you feel there is seasonality of data for each market is changing then try to reduce it using normalisation.

once you are done with normalisation we need to check the seasonality of all markets is same? they start developing model and check the Rsquared or MSE value for atleast 100 countries randomly.

start with small models like ARIMA or Holtwinters instead of LSTM.

if you feel out of 100 countries 80 countries gives same answer then go ahead and check with all countries.

Always try to explore the data before creating a model

  • $\begingroup$ Thank you! How would I structure the data for ARIMA though? If I'm fitting the same model on two time series, I have two observations for each timestep. Would one be the actual y-value and the others be exogenous variables? $\endgroup$
    – halfcup
    Aug 13, 2019 at 17:47
  • $\begingroup$ Assuming the data is available in 2 columns 1 is date and 2nd is Sales. Since don't know the other notable to comment,still trying to understand what could be other columns which has impact. think about important is that column. $\endgroup$ Aug 14, 2019 at 1:56

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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