Mixing categorical data and time-series data for regression purpose

Question

I came across a problem and I have been looking on internet how to solve it without finding a solution that fits my need. I am trying to predict investors behavior. To be precise, I would like to predict if an investor will invest and by how much. This is an exploratory project and I'm not sure if it will work but I think it is worth looking. To give you a context, we are a PE firm that offers investment proposition to investors and our goal would be to try to predict investors appetite of our deal. For this purpose, I have retrieve and built two datasets:

1. Mainly categorical and numerical data that describes what my investors are: wealth, where they live, age, family situation, what do they do for living etc...
2. Their investment habits. Namely: how much they have invested on our prior offers, when, on which deals, the time they spent looking at our website, time they spent filling their KYC procedure - Time-series data

Would it be feasible to combine these two datasets with one or multiple models and try to regress a potential investment in our next offer. Ofc, we can back-test this on past data.

Thank you for any help / ideas you may have!

Answer 1

There are multipe approaches you could try. Here are some (without the claim to completeness):

1. You could try to extract meaningful / helpful features from the time-series. A start might be counts, average values, variances and so on. But - using your understanding of the data - you might come up with better features. Combine these with the features from the tabular data and build a model of your choice over it.
2. Stacked embeddings can combine two or more models. Just train the two base models and build another model that takes the base-model-predictions as features.
3. Neural network architectures can deal with multiple data sources. The architecture would be slightly more complex than a simple layer architecture, but you could build recurrent layers for the time-series, feed-forward layers for the tabular data and combine both with more feed forward layers.
4. You could build an embedding over the time series to bring it down to a fixed feature vector. Although this is related to aboves approaches, it is slightly different.
5. You could enrich each time-steps feature vector with the same features from tabular data and build a model over the time-series.