Every calorie app gives you a number from a formula. The formula is a guess, and for a lot of people it's off by a few hundred calories a day. Marble treats it as the opening guess. After two weeks of weigh-ins, it compares what your weight was supposed to do with what it did, and corrects the target.
The opening guess
Your resting burn comes from the Mifflin–St Jeor equation, the one with the smallest average error in validation studies. That's multiplied by an activity level you pick, from 1.2 (mostly sitting) to 1.9 (physically demanding work). Then your goal sets the gap: a deficit for losing fat (about 0.7% of bodyweight a week by default, never more than 1%, never deeper than 500 calories a day), a small one for recomp, maintenance for eating better, a 10% surplus for building muscle.
The equation itself is usually within about 10% of resting burn. The weak part is the activity level: it's a self-assessment over a wide range, and people don't know which band they're in. No equation can fix that. Your scale can.
The correction
Marble smooths your weigh-ins into a trend, because one morning's weight swings by a kilogram or more from water, salt and food. From the trend it reads how fast you're actually losing or gaining.
The rule, in full
- Wait for two weeks of weigh-ins. Less than that is mostly water.
- Move only if the gap is clear. The gap between planned and measured rate has to be at least 1.25 times what your own scale's day-to-day noise could explain, measured from your last eight weeks of weigh-ins. A calm scale and daily weigh-ins earn a correction sooner than a jumpy scale weighed once a week.
- Correct half the gap. The trend lags reality, so a full correction aims at where your body was. Half converges without swinging back and forth.
- Small steps. At most 250 calories a step, or 500 when the gap is at least twice the noise. Under 60 calories it doesn't bother.
- Two weeks between steps, so it doesn't react to its own last change.
- Never below the floor of 1,200 calories for women or 1,500 for men, and it only reaches the floor on that same twice-the-noise evidence.
You see each change, with the numbers behind it, before it applies.
Why it doesn't estimate your metabolism from your log
The obvious design is to add up what you logged, look at what your weight did, and work out your true daily burn. That sounds more scientific and is worse. Calorie logs run low for most people, photo estimates included, and a burn worked out from a low log is wrong in a way that looks perfectly consistent, so nobody ever catches it.
Correcting the target instead cancels the bias. If you log 15% under, the target settles 15% lower in your own units, and the rate you actually lose at comes out right. The number stops being a claim about your physiology and becomes a setting that works for how you log.
How we checked it
Testing a correction on real people takes months and a lot of people we don't have yet. So we test it on simulated people whose bodies follow the NIH's dynamic model of weight change (Hall et al., Lancet 2011): fat, lean mass, glycogen, water, and the way metabolism adapts to a deficit. Each person gets a true daily burn the formula doesn't know, a noisy scale, a weighing habit (daily, sparse, messy), and a logging bias. The app's own Swift code runs unchanged against them.
Median error against true need, kcal a day
Show as a table
| After | Median error vs true need |
|---|---|
| Day one (the formula) | 320 kcal |
| 4 weeks | 115 kcal |
| 8 weeks | 42 kcal |
| 12 weeks | 25 kcal |
5,000 simulated people across five weighing and logging personas. The rule was tuned on one set of settings and judged on held-out ones (other random seeds, 5% and 25% under-logging, quiet and noisy scales, sparse weighers), with the pass criteria fixed before the held-out runs.
What it doesn't do well
- It's slow at first. At week 4, 46% of people are within 100 calories of their true need. Even a perfect scale-only rule would top out around 74% that early, because two weeks of weigh-ins can't tell fat from water.
- About 6% of steps move the wrong way, when scale noise looks like a real change. The next steps pull it back.
- Two big steps in a row can overshoot. In the simulation, 4 of 5,000 people ended week 12 more than 150 calories worse off than under the previous rule (and 8 ended that much better off).
- It's a simulation. The body model is well validated, but simulated people aren't real ones. We'll publish how corrections do on real weigh-ins once there are enough of them.
Every number in the rule, and the research behind it, is on Sources and methods. Your weigh-ins are stored on your phone, and the correction is worked out there.
Changes
- Page published.
- The noise gate: the target moves only when the gap is clearly bigger than the scale's own noise, with 500-calorie steps allowed when the evidence is strong. Median error at week 4: 131 → 115 kcal; wrong-way steps 7.2% → 6.3%.
Something here wrong or out of date? Write to support@getmarble.app.