There's also the heterogeneity of transactions. Looking at the csv export I've got from my bank, there are three "reference" columns and they play different roles depending on how I'm paying:

    - debit card: merchant name + date, card number, blank
    - alternative card-based payment system: constant (account number?), merchant, constant
    - alternative app-based payment system: constant, merchant, user custom reference

And a data sanitation issue: one of the merchant names has a comma in it, and it's not escaped or quoted in the csv so that record is corrupted.