Predicting good news using regulatory disclosure patterns, Python, SEC

I was curious if I could use insider trading patterns to predict whether the sentiment of 8-K filings were positive or negative.

https://github.com/john-friedman/datamule-python/blob/experimental/examples/predicting-good-news-from-insider-trading/exploration.ipynb

Doesn't look like it - not enough data. However, there is an interesting result. Trading behavior after a positive 8-K is very different for insiders. They wait to sell, possibly to let good news trickle out.

https://github.com/john-friedman/datamule-python/blob/experimental/examples/predicting-good-news-from-insider-trading/plots/sell-side.png

I'm considering looking at companies using their entire corpus of filings (e.g. 10-K to IRAN NOTICE) for a year and applying finbert for another try, but it seems expensive in terms of compute.

I've mostly solved the download problem by hosting my own SEC archive - it takes about 45 minutes to download 50,000 8-Ks and 150,000 Form 4s (1 year), but compute seems harder.

I used Loughran McDonald dictionaries to do the sentiment parsing - took about an hour, so I imagine applying finbert on a larger corpus would take weeks on my laptop.

Not sure if there's a better option?