Future Me: How Contura Predicts the Body You’re Actually Working Toward

What your body could become. If you commit to your goals starting today, where could you realistically be in six months? In a year? Future Me shows you the version of yourself that’s actually within reach — not a fantasy, but a goal worth working toward. Keep using Contura, and we’ll refine the picture with your real trends.

Most “future body” features in fitness apps fall into one of two traps. They either show you a generic stock photo of someone fitter than you (motivating, but dishonest), or they linearly extrapolate your last few weeks of progress and predict you’ll weigh 12 kg by Christmas (mathematically valid, physiologically nonsense).

Future Me was built to do something harder: produce a prediction that is personal, physiologically plausible, and visually honest — and to keep getting better the more you log.

This article goes under the hood of the prediction engine. It’s a long read, but if you’ve ever wondered whether the body in your Future Me preview is actually grounded in something real — yes, it is. Here’s how.

The core problem

You log measurements over time — weight, waist, hip, chest, arm, thigh, calf, neck, wrist. Some users log obsessively. Some log once and disappear for two months. Some only ever measure their waist.

From this messy, unevenly sampled, often noisy data, the engine has to answer:

What is this person actually trying to do?
How fast are they actually changing?
Which of those changes are real, and which are measurement noise or a bad day?
Where are they likely to be in 1 month, 3 months, 6 months, and 1 year — assuming they keep doing roughly what they’re doing now?
What does that body look like in three dimensions?

A naive answer to any one of those questions will produce a bad prediction. Future Me answers all of them in a single coordinated pass.

Architecture: a six-phase pipeline

The engine is structured as a sequence of phases. Each one is responsible for a specific type of reasoning, and each one feeds the next:

Build the time series — group every recorded measurement by body part, sort chronologically, and reconstruct each part’s history independently.
Profile the data — figure out how much we actually know about this user. Three samples in a week is a different problem than 18 months of weekly weigh-ins.
Infer intent — what is the user trying to achieve? Weight loss, muscle gain, recomposition, focused work on a specific area, or just maintenance?
Predict each body part — the heart of the engine, where the per-part forecast is computed.
Apply intent-aware adjustments — small, deliberate aesthetic nudges, applied only to brand-new users with very limited data.
Fill the gaps — derive any missing body parts from the predicted ones using anthropometric constraint propagation, so the 3D model is always complete.

The output is a full set of predicted circumferences, an inferred intent, a confidence rating, and a divergence analysis — everything the renderer needs to draw your future self.

Phase 1–2: How much do we actually know about you?

Before the engine predicts anything, it asks a question most apps skip: do we have enough information to trust our own forecast?

It scores three things:

Span — how long has this user been recording? A week of data and a year of data warrant very different strategies.
Sampling density — how many measurements, in total, are in the history?
Body-part coverage — are they tracking a full picture, or just one or two parts?

These factors are weighted (with the largest weight on time span, because trends only become real with time) and combined into a single confidence score. That score sorts every user into one of three operating modes:

Mode	Who falls here	How the engine reasons
Starter	Brand-new users with almost no history	Ideal-target gap closure — see “What would a healthy version of this body look like?”
Developing	Some history, but trends not yet stable	A blend of personal trend and ideal-target reasoning
History-rich	Months of consistent data	Pure trend extrapolation, with safety mechanisms

This matters because Future Me does not pretend to know things it can’t know. A user who logs once doesn’t get a confident trend line — they get a sensible projection toward what their body could realistically become. A user with a year of consistent data gets a real forecast based on the trajectory they’re already on.

Phase 3: Inferring intent

Future Me decides for itself what you’re trying to do, based on your data.

For new users, it uses BMI and current proportions as a proxy for likely goal — overweight users are usually trying to lose, very lean users are usually trying to gain, and people near their healthy range are usually working on body recomposition. This is a deliberately humble assumption, and it gets revised the moment real data arrives.

For users with history, intent is read directly from the data:

Recent trends in weight, waist, and arm are analyzed over the last several weeks.
If a particular body part is being tracked unusually frequently, the engine recognizes that the user is focusing on that part and labels the intent accordingly (focusedPart(arm), for example).
Otherwise, the relative motion of the major signals — is weight dropping? are arms growing while weight holds steady? is the waist shrinking with weight stable? — determines whether the user is in weight loss, muscle gain, or a maintenance phase.

This inferred intent then feeds back into the prediction itself. Muscle-gain intent strengthens the engine’s confidence in arm and chest growth. Weight-loss intent strengthens its confidence in waist and hip reduction. The model is goal-aware, not just motion-aware.

Phase 4: The per-part prediction engine

This is where most of the work happens. Each body part is predicted independently but informed by the others. For data-rich users, the prediction is built up through several layers of reasoning, each one designed to neutralize a specific failure mode that simpler models fall into.

Outlier filtering

Real measurement data is full of bad points — a tape held too tight, a measurement taken right after a heavy meal, a number entered into the wrong field. Before any trend is computed, isolated outliers are filtered out using a robust local-window statistic that’s mathematically resistant to spikes. A safety valve prevents the filter from being too aggressive: if removing outliers would discard too large a fraction of the user’s data, the engine backs off and keeps the original series. Cleaning data is supposed to clarify the truth, not invent one.

A smoothed view of “where you are now”

Even your most recent measurement isn’t the truth — it’s a noisy sample of the truth. The engine computes a recency-weighted estimate of your current value, giving more weight to recent measurements but allowing earlier ones to anchor the estimate. This becomes the starting point for projection.

Dual trend estimation

A single trend calculation is a single point of failure. Future Me computes the trend two different ways and combines them:

A recency-weighted rate, which captures momentum and is responsive to genuine recent change.
A robust median-of-slopes rate, which is mathematically immune to outliers and ignores dramatic single-day swings.

Blending these gives the engine both responsiveness and stability — it can react to a real new trend without being fooled by one bad week. The blend is then modulated by a trend consistency factor: if your data is clean and monotonic, the trend is trusted more; if it’s zigzagging, the trend is discounted.

Anatomy-aware weight coupling

Body parts don’t change independently. If your weight is dropping, your waist is almost certainly going with it. Your wrist almost certainly isn’t.

Future Me encodes this directly. When a weight time series exists, the weight trend is partially blended into every other body part’s predicted trajectory, with each body part having its own coupling strength. The waist is highly coupled to weight; the hip and thigh are moderately coupled; the chest and arm are weakly coupled; the wrist barely couples at all. These coefficients are anatomically grounded — they reflect the well-established fact that some body sites are dominated by adipose tissue (which moves with weight) and others are dominated by skeletal structure (which doesn’t).

When the inferred intent is muscle gain, the coupling pattern shifts: the arm and chest get stronger upward bias to reflect that hypertrophy doesn’t follow the same rules as fat-mass change.

Volatility damping

If your measurement history is noisy — high coefficient of variation across recent points — the engine reduces how much it trusts any computed trend. Inconsistent data should not produce confident predictions.

The asymmetric prior — the safety core of the algorithm

This is the most important design decision in Future Me, and it’s the one that separates it from every linear-extrapolation predictor we’ve seen.

The engine treats convergent trends (motion toward a healthy reference) and divergent trends (motion away from a healthy reference) asymmetrically.

If your data shows you moving toward a healthier body, the engine trusts the trend and projects it forward, with damping appropriate to the strength of the evidence.
If your data shows you moving away from a healthy reference, the engine demands stronger evidence before it accepts the trajectory. Without that evidence, it defaults to a gentle convergence: the predicted body drifts back toward the healthy reference rather than running away from it.

In plain language: a user losing weight toward their goal range is trusted. A user gaining weight is shown a forecast that doesn’t blindly extrapolate further gain unless the data is unmistakable.

This mirrors how a responsible trainer thinks. Optimistic about progress, cautious about negative trajectories, never the source of a doom prediction the user didn’t ask for. It’s also why Future Me will never show a user a horror-show forecast based on three bad weeks. The math refuses to.

Horizon damping

The further out you forecast, the less anyone can know — and the engine knows this. Predicted change is progressively damped as the horizon stretches from one month to one year. Long-horizon predictions also pay attention to the user’s confidence score: a one-year forecast for a user with sparse data is damped more aggressively than one for a user with eighteen months of weekly logs.

Plateau attraction

Bodies don’t follow straight lines. They settle. Beyond the linear projection, Future Me applies an attraction term that gently pulls the predicted trajectory toward a height-scaled, gender-aware healthy plateau. The attraction grows with horizon length and is capped, so it never dominates real personal data — it just ensures that long-term predictions converge into physiologically sensible territory rather than running off into unrealistic numbers.

Sub-linear long-term change

Each body part has a physiologically plausible maximum daily rate of change. Beyond a certain horizon, the maximum cumulative change scales sub-linearly (with the square root of time) rather than linearly. This encodes a well-known reality of body change: initial progress is fast (water weight, beginner gains, glycogen), and subsequent progress is much slower. Linear extrapolation of early progress is the single most common reason fitness predictions fail. Future Me does not make that mistake.

Physiological clamping

After every other layer has had its say, the final predicted values are clamped to height-scaled physiological ranges. No predicted weight outside a plausible BMI band. No predicted waist that’s anatomically impossible. The output is always a body a real human could have.

Phase 5: Aesthetic adjustment for new users

For users with almost no data, Future Me applies a small, intent-aware nudge — slightly slimmer waist for users on a weight-loss intent, slightly more developed shoulders/arms/chest for users on a muscle-gain intent, where biologically appropriate. This is deliberate and openly disclosed: brand-new users have no trend data, so the prediction has to draw from something. We chose to draw from a gentle, healthy aspirational target rather than from nothing.

This nudge does not apply to history-rich users. Once you have real data, your real data is the source of truth. The engine will not dress up your prediction.

Phase 6: Filling in the missing parts

Most users don’t measure every body part. The engine uses a constraint-propagation solver — informed by anthropometric ratios established in human-proportion research (waist-to-hip relationships, chest-to-arm scaling, limb proportionality, and so on) — to derive any unmeasured part from the parts that were measured.

The result is always a complete, internally consistent body. The 3D Future Me model never has missing dimensions, and the proportions always make anatomical sense.

Validation: does this actually work?

Sophistication on its own is meaningless. We tested aggressively.

The engine is covered by a test suite of 39 cases across 7 distinct test suites, run against 10 realistic body profiles spanning male and female users from petite/lean (BMI ~19) to overweight (BMI ~34), heights from 158 cm to 185 cm, and weights from 48 kg to 110 kg.

Starter-mode behavior (no historical data): Overweight profiles in the BMI 30–34 range are predicted to move toward healthy ranges (BMI 25–28), while users already near their healthy range are correctly flagged as fit and given small refinement-style predictions rather than dramatic transformations.

History-rich behavior (six months of declining-weight data, male, 175 cm, 82 → 75 kg): The engine extrapolated a continued downward trend across all four horizons, with the magnitude of change increasing with horizon length but moderated by horizon damping and plateau attraction — exactly the curve a responsible trainer would draw.

Maintenance behavior (flat measurement history): All predicted values stayed within roughly 5–10% of current values — correctly recognizing that the right prediction for “this person isn’t changing” is “this person will stay roughly the same.”

Stress testing:

100 randomly generated profiles (heights 140–210 cm, weights 40–130 kg, both genders): every prediction produced valid output with sensible BMI ranges and positive circumferences.
64 extreme combinations (8 heights × 8 weights × 4 horizons): every case produced valid output.
Edge profiles (30 kg, 200 kg, 140 cm tall, 210 cm tall): all clamped to physiological ranges. No nonsense.

There are no profile-class blind spots, no horizon at which the engine breaks down, and no edge case at which it produces a body a human couldn’t have.

What Future Me will not do

We think it’s worth being explicit about the model’s deliberate refusals:

It will not predict a runaway negative trajectory. Three bad weeks do not produce a one-year disaster forecast.
It will not predict superhuman progress. Two great weeks do not produce a one-year transformation forecast.
It will not invent measurements you didn’t take. Body parts you don’t track are derived from what you do track, with full anatomical consistency.
It will not extrapolate past physiological reality. Every output is clamped to a body a real human could have at your height and gender.
It will not lie about its confidence. The engine reports back its own confidence rating, and divergence analysis is surfaced when your trajectory and the model’s trajectory disagree.

What Future Me will do

It will give you, on the day you first open the feature, a sensible portrait of the version of you that’s actually within reach. Then, every time you log a measurement, that portrait will sharpen. With one month of data, it will start reflecting your actual trajectory. With three months, it becomes specifically yours. With six or twelve months of consistent logging, you have a Future Me that is, in a meaningful and validated sense, a real forecast — not a daydream.

That’s the version of yourself worth working toward.

Frequently Asked Questions

Is the Future Me prediction based on AI? Yes — Future Me is powered by a multi-stage prediction engine that combines anthropometric reasoning, robust statistical trend estimation, and goal inference. It’s an AI system in the sense that it reasons about your body the way a knowledgeable trainer would. It’s not a single black-box neural network; it’s a transparent, validated pipeline whose every stage was designed to defeat a specific failure mode.

Why does my prediction change when I log new data? Because that’s the point. Future Me is designed to refine itself with every measurement. The more you log, the more your prediction reflects your actual trajectory rather than a default starting estimate.

What if I haven’t logged enough data yet? The engine handles this explicitly. New users get a prediction grounded in healthy anthropometric targets for their height and gender, not a confident trend extrapolation it doesn’t have the data to make. The forecast becomes more personal as you build history.

How accurate is Future Me at one year? Long-horizon predictions are deliberately damped. A one-year forecast is a directional guide, not a calendar date. Honest uncertainty is built into the model.

Does Future Me predict gender-specific changes? Yes. The engine uses gender-aware healthy reference proportions and intent-aware coupling, so a male user’s muscle-gain prediction emphasizes upper-body development while a female user’s prediction respects different anatomical proportions. These differences are grounded in published anthropometric reference data.

Why does my waist react more to weight changes than my wrist? Because anatomically, that’s how bodies work. The engine encodes this directly — body parts dominated by adipose tissue change with weight; body parts dominated by skeletal structure don’t.

Can Future Me show a worse version of me? Only if your data unambiguously and persistently shows a negative trajectory. The asymmetric prior built into the engine treats negative trends with extra skepticism, requiring strong evidence before predicting further decline. This is intentional.

Where do the “ideal” reference body proportions come from? From population-level anthropometric survey research, scaled to your height and gender. They’re reference points, not commands. Future Me uses them as anchors, not as targets we’re trying to push you toward.

In one sentence

Future Me is a validated, anatomy-aware, goal-aware body prediction engine that shows you the version of yourself you’re actually working toward — and gets sharper the more you log.

Open Contura. Log a measurement. See your future.