mulligan: An account predictor for ledger

The motivation behind mulligan was simple. I like to spend as little time as possible on the donkey work on my finances: consolidating, recording, accounting — while still having an excellent monthly overview to make slid, sound financial decisions. Bouncing around the globe so much and having accounts in different countries and currencies makes this harder than it should be, though I’ve been using a ghetto-looking, but amazingly powerful plain-text accounting system (ledger-cli) for years now to make my life easier.

You can read a blog post (which needs an updated 2026 version) on how I handle my setup yourself at Tracking Your Finances with Reckon and Ledger.

No one likes manually entering finance entries except a particularly masochistic breed of retentive accounting types. In my setup, Cantino’s Reckon gem took my useless banks’ csv exports, trained against my past finances, and then used a simple Bayesian predictive model to puzzle out what the accounting category should be for each transaction, writing me out a nice ledger file I could delta onto my main file. I do not exaggerate when I say this saved me on the order of a day’s work per month.

Lately, however, Reckon has been getting a lot of category predictions wrong. I am unsure if it’s the extra slop in a lot of bank records or the fact my data could be a bit confusing: multiple entities of the same name in different countries and currencies crowding the priors, or perhaps even the fact that with all my traveling recency bias should be weighted more.

In particular, I wanted something that could be orchestrated to run unattended against my financial files from my banks each month, return me a ledger formatted file, and did not need to rely on sending my financial data to one of the major LLM providers. As well as flag which predictions it was less sure about so I can take a look at them.

It’s definitely an incremental improvement on the excellent pedigree of Reckon, but it does perform slightly better, at least with my data (it’s also faster, since in Go versus Ruby and forking off the Go code from into-ledger (which both expects you to feed back against it and uses Claude to check predictions in the default config).

I’ve made is as “CLI-compatible” with Reckon as possible, building on the excellent work of Cantino and there is an automated build pipeline to install this via homebrew or via your favourite package manager as a release (ie. brew install wakatara/tap/mulligan).

An initial release, but so far, it has addressed all the mis-classifications I was experiencing with Reckon (mostly things being called Taxi, Restaurant, or Software when they were clearly not and despite being oft-corrected in the master file.). I’m releasing it because it was useful for me, and if you have similar issues, I imagine it will be for you. It needs more testing with other data than my specific finances and curious how it performs on other peoples' data, so please kick the tires and give feedback on quirks, strangeness, or outright problems in the github repo.

I hope you enjoyed the post. If this post was useful to you, or you have something to add or a counterargument drop me a line and let me know, via mail or elephant below. If you have an alternative, , or think you have additions that might improve the approach, please holler. Feel free to mention or ping me on @awws on mastodon or email me at hola@wakatara.com.