Hi HN-
Sharing a project that initially started to help me scratch my itch for carbon bike parts (mountain bike wheels in my case).
There's lots of different forums for finding used bikes, and even as an experienced cyclist it's hard to wade through all of them and compare bikes with custom parts to evaluate a good deal.
bike.broker helps by letting you get a price prediction for a used bike, which takes into consideration the upgrades/custom parts. This is in comparison to https://www.bicyclebluebook.com/ which is purely model/brand/age based (and has a reselling business which potentially biases their numbers).
I built this by: - collecting historical sold ads from across the web - training+validating a multimodal deep learning model on this dataset (BERT + tabular features). ~18% MAPE on a >80K ad sample test set across all bike types. - exposing as a chat interface for maximum flexibility (yes, a dashboard would could be more effective for this but I wanted to minimize data exposure)
Limitations: - Ads that dont have model year in them are over predicted for value - Doesn't have a metric for condition of the bike that is robust (attempted adding images to modelling but wasn't worth the complexity for side project) - Doesn't currently have an index for new bike stock/market dynamics. I.e post Covid bike surplus and sales wasn't priced in well -> need to compare to new bike prices as sanity check (A deep research like feature to solve this was something I want to build, but would add quite a bit of cost)
Currently supports pinkbike.com as I find it the most consistent source for more than basic bikes.
Hope you find it helpful!
Comments URL: [https://news.ycombinator.com/item?id=49022946](https://news.ycombinator.com/item?id=49022946)
Points: 1