How should I approach a univariate timeseries with multiple rows of data per timestamp?
Using stocks as an example, I am attempting to identify a time series model, or suitable alternative, that identifies macro relationships between independent variables and stock prices over time. The issue is that I have multiple rows of data per timestamp.
I do not believe that VARMAX is appropriate, because I am not attempting to identify the effect of one stock on another.
I also do not believe that ARIMAX is appropriate because I have multiple linear equations per timestamp, one for each stock.
I am basically attempting to perform linear regression to obtain coefficients for the independent variables, with a time component added.
I have considered the following three options, but I'm not sure if they are reasonable approaches:
Converting the timestamp into additional features such as "year" and "month", and running non-timeseries models such as xgboost.
Performing regression on each timestamp separately to collapse the matrix of stocks into a single row of coefficients for the timestamp, and then modeling the coefficients and stock prices together in a multivariate VAR model.
Averaging the data across the stocks for an individual timestamp, and then running an ARIMAX model. My concern with this approach being that I'll lose valuable information regarding interactions between features as two features move together or move apart over time.
I want the model to learn information at a macro level regardless of stock, to then apply a prediction at the micro level of a single stock.