The forecast card sits at the top of every Marvin dashboard. It tells you one number: how much money you'll have on your next pay day. People ask us all the time how it works. We won't show you the recipe — but we will tell you what it is, and isn't.
It's a forecast, not a fortune teller
Most apps that promise to "predict" your money lean hard on machine-learning language. Pattern detection. Anomaly scores. AI-powered. The marketing reads better than the output usually does.
Marvin's forecast is the opposite of that. It's grounded in facts you can see — your balance, the bills you've told us about, the income you've told us about, and what your spending has actually looked like recently. There's no black box doing something clever you can't trace. If the number changes, it's because something you can see also changed.
What it uses
Four kinds of inputs, all visible to you in the app:
- The balance you last shared with us. Marvin doesn't fetch it from your bank — we don't ask for that. You update it, we use it. The forecast shows you when it was last updated, so you always know how fresh the number is.
- Income you've set up. Salary, side income, whatever you've told Marvin to expect. We don't guess at money that might land in your account — guessing income is how forecasts lie to you.
- Bills you've set up. Rent, subscriptions, the stuff that auto-debits. Marvin places each occurrence on a timeline and counts only the ones falling before your next pay day.
- Your recent spending pattern. Not a personality model. Not "AI learns who you are." Just a sober look at what you've actually been spending recently, used to estimate what's likely to continue.
Why it's not "AI-powered"
We have AI. Marvin's chat uses a large language model — that's how it can answer "can I afford this?" in plain English. But the forecast number itself isn't generated by the model. The model asks the forecast, the way you do.
We made that split on purpose, for three reasons:
- Auditability. When you ask "why is my forecast this number?", we can point at the inputs that produced it. With ML you'd shrug.
- Honesty. An ML forecast feels confident even when it shouldn't. A grounded forecast knows what it doesn't know.
- Repair. If the number looks wrong, you can fix it: update your balance, add a missed bill, log a forgotten purchase. The forecast obeys you. ML models don't.
What it deliberately doesn't do
The forecast does not:
- Try to predict bills you haven't entered yet.
- Account for one-off events — birthdays, festivals, travel — unless you've put them on the timeline.
- Adjust for inflation, currency drift, or credit-card interest.
- Pretend to see further into the future than it can.
That last one matters most. A forecast is only honest while its assumptions are reasonable. Stretch the horizon and the assumptions stack up and the number stops being trustworthy. Marvin makes a deliberate choice about how far it'll project, and stops there.
The honest version
The forecast isn't smart. It's transparent. Every input is something you put in, can see, and can change. Nothing mysterious is happening behind the curtain — and that's the point.
We could dress this up with neural-net language and call it "AI-powered." We don't. The moment we did, we'd lose the thing that makes the number useful: you can see why it is what it is, and you can do something about it.
More on what we mean by "honest" software — read what "honest" means in a money app.