sounds really interesting!
Regarding your time series forecasting, I would recommend to consider:
- simple ARIMA models, see also this thread: Deep learning for sequence pattern recognition - #2 by Christian_Simonis
- a naive or linear prediction at least as a fair benchmark, as illustrated in this repo.
As a hint: besides the model resp. the (chained) algorithms, please consider that you will need many more components in your software system to realise an effective and satisfying operations such as pipeline monitoring, analytics, data cleaning, preprocessing, config, maintenance and other components to satisfy non-functional requirements as well as your specific MLOps concept etc., see also: Overview of ML Pipelines | Machine Learning | Google for Developers
Good thing is: with modern serverless infrastructure you can quickly build and validate your business hypothesis in a quite lean and scalable way.
I would be interested in your project. Feel free to keep us posted here if you like and please do not hesitate to ask for any advice or help.
Best regards
Christian