Regression modelling

[IanQS]

Skeleton for early work. In the future, we'll split off all the models from here so that we don't have to PR to get new linear_models into the system

[IanQS]

Plan

1) we'll try a bunch of linear models whatever we can get our hands on. We should exhaust sklearn's provided methods 2) combinations through a controller which is some ensemble model 3) hierarchical learning 4) portfolio management methods

Resources

• make sure licenses of any packages you use are permissive licenses

Time series analysis tutorial

Time series analysis library

Don't calculate absolute values, calculate logs or pct

[IanQS]

The base example will be done in the next hour or so. After which I'll explain add more commentary.

Tasks: 1) @bryantpq @wrhm populate a bunch of different models

• (whether from statsmodels or wherever else), but I specifically want lightboost, xgboost, and catboost.
• Always pull and then push, and communicate with each other about who is doing what models)
• if the model isn't "SOTA (xgvoost, catboost, lightgbm), I'd prefer if it was an "online" model, but it's fine if it's not as the coordinator will hopefully decide when it's optimal to retrain
• bear in mind that the data is time-series data so you may need to pre-process it for whatever your application is. Keep communication open to discuss any utility functions you may need

LMK when you're done or if you ever need help rebasin

After we've got like 10 models or so lemme know and we'll talk about

2) @ianqs write a coordinator that takes a bunch of models, and depending on the state of the market decides: i) what models to retrain based on some cost function ii) how to weight the other models (based on some notion of momentum in the market?)

Validation Scheme

The validation scheme will be across a different pair from what the thing was trained on.

[wrhm]

Bookkeeping questions/ideas for tracking the models we implement:

• A spreadsheet may be useful to track: who's working on which models, which models have existing open-source implementations, and possibly a priority for which models to implement next.
  • Is there a way for us to maintain something like a spreadsheet within GitHub? Or should create and link a Google Sheet or something?
• I think it makes the most sense to create a new issue for each new model being worked on, to make code review as efficient as possible. Thoughts?

[IanQS]

Is there a way for us to maintain something like a spreadsheet within GitHub? Or should create and link a Google Sheet or something?

Can be done on github projects board

I think it makes the most sense to create a new issue for each new model being worked on, to make code review as efficient as possible. Thoughts?

I'd rather not - a lot of things like utils will be common among models and rebasing constantly will be a PITA

[IanQS]

Coordination