[l72]

Has anyone tried training their own personal machine learning model to take export from their Bank+receipts and auto categorize everything? It seems like it'd be a fairly simple classification model that wouldn't require too much training...

[phito]

I've been tinkering with that the last few days. My main issue is that most of my bank statement descriptions are really bad. It's often impossible to know what the transaction was about just from the description. Local LLMs (9B parameters, I don't have the required hardware for bigger models) don't help at all as they don't know the context around the transaction.

I use Firefly III for my finances management, and in the end it is just much better to setup "dumb" rules that look for keywords in the description/IBAN numbers etc. I have about 30 of them and they cover 90% of my transactions.

[tecoholic]

I have tried and given up. As others noted, my bank statements are just random short codes and numbers most of the time. So it requires a fair amount of manual review. There is an article comparing Fuzzy searching using Elasticsearch and a ML model training, that concluded a fuzzy search is better. I have tried with Miellisearch and found it okay. But still ended up doing everything manually in an Excel sheet, because bank descriptions just suck.

[AstroNoise58]

There are tools to do just that, see for example banks2ledger: https://git.hq.sig7.se/banks2ledger.git

[koevet]

I did, I use beancount plus some ML for automatic transaction categorization. Working on a new version right now that can also integrate with an LLM