Rent Comp Models · Griffin Trygg

Why I built it

After months of weekly surveys I had a real dataset sitting there. I wanted to know two things. Could a model predict a floorplan's rent well enough to sanity-check the survey? And at a few thousand rows, does a neural net actually beat a linear regression? I also wanted an excuse to learn PyTorch properly.

What it does

Plan-level dataset exported from raw survey returns, with a flag for carried-forward figures.
Effective rent derived from free-text concession language.
Cross-validated comparison of log-rent linear regression, LightGBM, and a PyTorch model with property and market embeddings.
Readable weekly HTML pages per market.
External context from Census ACS tracts, Zillow rent index and HUD fair market rents, tested as features.

How it works

Two test splits. Cold means properties the model has never seen. Warm means new weeks for properties it knows. LightGBM got to 9.3% error cold and 7.2% warm.

The PyTorch model with property and market embeddings landed at 13%, which loses to plain linear regression at this data size. Census tract features didn't help either. Both of those are in the write-up as results.

Calls I made

I wrote down the loss

The deep model was the reason I started, and the honest answer is that it lost. The repo says so in the first paragraph.

Notes after every session

Every concept I worked through went into a NOTES.md, so the repo doubles as my study log.