Should I concat multiple stock timeseries datasets into one?

External question — Data Science Stack Exchange Author: Ubler Original post: https://datascience.stackexchange.com/questions/86002 License: CC BY-SA 4.0 — https://creativecommons.org/licenses/by-sa/4.0/ Adaptation: HTML converted to plain text; contact email addresses removed. I have several timeseries datasets of stock data, with fundamental indicators. I would like to build a model that selects stocks for buy and hold. I understand that to perform this task I have two options: Train a model for each stock: This way, I understand that it is the most practical, however, the amount of data for each model will be very reduced (Each dataset has less than 1000 lines). Putting all the data together in a single dataset: I didn't find anything on the internet to support this idea, however, I understand that the model would be more robust and would have a much larger amount of data to be trained. So, what would be the correct way to perform this type of analysis? Any of you would suggest another way? Thannk you in advance!

External answer — Data Science Stack Exchange Author: jottbe Original post: https://datascience.stackexchange.com/a/86093 License: CC BY-SA 4.0 — https://creativecommons.org/licenses/by-sa/4.0/ Adaptation: HTML converted to plain text; contact email addresses removed. I would suggest option two. Because this way your model would have the chance to learn something for one stock, which it can apply for other stocks as well. If you provide the type of stock as an input feature, it should be able to distinguish between the specialities which only occure within one stock and the common things. So it is kind of able to transfer knowledge from one stock to the other. But at the same time, I would suggest to try out both and choose whatever performs the best. This way you can also justify your choice in the end.

External answer — Data Science Stack Exchange Author: leviva Original post: https://datascience.stackexchange.com/a/112946 License: CC BY-SA 4.0 — https://creativecommons.org/licenses/by-sa/4.0/ Adaptation: HTML converted to plain text; contact email addresses removed. The two methods suggest two different assumptions. The first method implies that the same features on two different stocks lead to different outcomes, and therefore learning from both is counterproductive. The second method implies that if you learn something on one stock it will still be true for the other stock. In other words, a good model for one stock is good for all stocks. How you choose to model the world is left up to you. I would suggest when using the second method to check that your features don't leak information. For example concatenating two time-series with different value ranges would allow your model to "know" on what series it's working on.