| I never run it manually. I plug this tool in my scripts that import transactions from banks (and stripe). The scripts extract transaction details (amount, payee, currency), then they use `ledger-guesser` to create ledger entries and add them to the journal. The generated entries are "uncleared". Then I manually "clear" the entries. Review. Commit. (You can also use a tag for reviewed transactions if you already use ledger's "cleared" indicator for something else). For the best results, I have 1 journal per bank account. So I have different training data for each bank account. There's no confidence threshold. When there is a new type of transaction is encountered, the guesser will chose the account with the highest probability. The guesses are made based on the words found in the payee (and date). Unknown transaction have generally a few known tokens. Example: "INCOMING TRANSFER FROM NEW_CLIENT"... The classifier will probably classify that entry as "Incomes:OurLargestClient". In that case I have just to fix that entry to change the client, all the rest is good, it still saves a good amount of typing. |