powerAI v3: A Kalman Filter for Your Fitness Curve

Roughly eighteen months ago we asked ourselves a question: how can neural networks deliver real value to an athlete? More precisely — what if, week after week, session after session, you could get what normally requires a lab visit or a Powertest? Across three model generations, that question became powerAI v3. This is the honest version of what v3 does better — and what it deliberately does not promise.

Why we built this

In endurance sport, you spend most of your time flying blind. You train, week after week — but how your performance is actually developing, you never see. And which training block produced which effect, even less so. There are clean studies on all of this, but they happen in the lab: controlled conditions, expensive equipment, out of reach for a normal athlete. That is exactly what we wanted to break open. From day one, our ambition was to make that clarity affordable — and available to everyone, not just to a test lab.

The road here: v1 and v2

v1 was our first neural network — the beginning of what would become many. For a first attempt it was surprisingly good. But it leaned on a demographic relationship that, in hindsight, made the numbers look too optimistic, and on monotonous indoor sessions it lost the beat.

v2 was a clean restart: more data, more robust test concepts, and the metric error from v1 corrected — the quality measure we used back then was, we admit it openly, too optimistic. v2 also handled poor or very flat heart-rate data far better. And the real leap: v2 attaches an uncertainty to every estimate. On a messy session, the model itself says: I'm not sure about this one. The model knows when it doesn't know.

But we never switched v2 on as the default for everyone. A few users could opt in, nothing more — because even then we knew there was much more on the table. So we went straight on to v3.

v3: the leap

v3 is the quantum leap, for two reasons.

The first: v3 reads your raw data directly. Earlier models were fed pre-chewed summary features. v3 learns second by second from the raw streams themselves — heart rate, power, temperature, cadence and more. Think of a weather model: it takes thousands of noisy readings — temperature here, air pressure there — and reconstructs the state of the entire atmosphere from them. That is exactly what v3 does with your body. From the raw interplay of heart rate, watts, temperature, cadence, stride rate and much more, it reconstructs the state of your engine. It built that foundation on 744,871 training streams before it ever estimated a single VO2max — it knows how a body responds to load.

The second reason is the real trick: the filtering.

The Kalman filter, in one sentence

Every single measurement carries noise. A ride in the heat, a bad night, a cranky chest strap — each session on its own is a shaky snapshot of your engine. No single value is the truth. Together, though, they form a clear picture.

That is precisely what a Kalman filter is built for. It fuses a chain of noisy measurements into one calm curve, weighing each new session against everything it already knows about you. The trick lies in how much it listens: in a quiet week it smooths hard and averages small outliers away. When your training volume ramps up and signals that something might genuinely be changing, it opens the gate and lets more movement through. That way it never swallows a real jump just because it likes to stay calm.

The filter itself is textbook. The more than 11,000 curated Powertest labels it is calibrated against are not.

How calm is the curve, really?

We measure this calmness with a single quantity: the typical swing of your estimated fitness from one week to the next — measured across 425 athletes, each with at least ten weeks of training. Important: this is not a deviation from a lab value. It is a measure of how steadily the curve itself runs over time.

In plain terms: your estimated VO2max varies by less than one point from week to week, your Critical Power by less than 6 W. Compared with v2, which sat around 1.95 ml/min/kg and around 11.7 W here, that is more than cut in half.

The last column, the "physical floor", is the limit no honest smoothing can get below — a real athlete's curve always carries some genuine week-to-week movement. At 0.886 we sit practically on that floor. That is the strongest evidence that we are ironing out the noise, not your form.

And every single weekly value comes with an honest uncertainty band. For the first athletes using v3 live, it averages around ±0.4 to 0.5 ml/min/kg and ±2 to 3 W. This band breathes: after a training break, when the filter has not heard from you in a while, it widens — and with every session that comes back in, it tightens again. So the curve does not just tell you where you stand; it tells you how sure it currently is.

For the first time, a week-to-week measurement of this quality is available to everyone — not in a lab, not in a study, but in your everyday training. Honestly: we know of no platform that delivers this.

Calm doesn't mean deaf

The obvious objection: is the curve only this smooth because the filter swallows real progress? No — and we checked that three ways:

• The filtered curve sits at the physical floor, not below it. It smooths right up to the limit of what is honest, and no further.
• On device-consistent comparison pairs, the filtered curve moves more in step with real Powertest jumps than the raw estimate does.
• Where a real Powertest shows a jump, the filtered curve follows it more closely than the unfiltered one.

Calm when nothing happens. Awake when something changes.

A real fitness curve

Take one athlete from our validation: a Powertest in May, then a rising curve over the following weeks up to a peak of 52.0 ml/min/kg VO2max and 316 W Critical Power in July — followed by a clean descent into recovery. No spikes, no jitter. A fitness curve the way a coach would draw it by hand. And because every point carries its confidence band, you see at a glance how much weight the curve deserves right now.

That is the real win: for the first time, you can read training effects straight off your performance curve. No more guessing whether the hard block did anything — you watch the curve respond to it.

What v3 is not

Two honest points, because they belong here.

First: the raw estimate from a single activity scatters considerably — only the fusion over time turns it into a calm curve. v3 is therefore not a replacement for a lab test or your Powertest. Quite the opposite.

Second: for the large majority of athletes, the curve sits in a perfectly reasonable band from the start. Rarely, though, the starting value can carry a larger offset — the curve then runs cleanly but sits, as a whole, a step away from your true level. There is a simple fix: a single Powertest calibrates the network to you — our digital twin model — and from then on the curve is anchored to you. Our recommendation, therefore: one Powertest at the start, and roughly one per year after that to stay calibrated.

Physiologically grounded

v3 estimates the two clean axes — aerobic capacity (VO2max) and Critical Power. VLamax, your thresholds and your training zones are derived from them with the established Mader model. The network delivers the calm axes; the physiology does the rest. No zoo of models — one coherent source.

Your Powertest is the anchor

The more often we measure you cleanly, the tighter your curve sits on you — not on a population average. The Powertest is exactly that anchor point: the one clean fixed star the entire time series is aligned to. That is the difference between "estimated from your data" and "measured on you".

→ Book a Powertest and give your curve a fixed star.

Where v3 stands today

As of today, v3 is rolling out. We are switching it on step by step, checking curve stability on live data against the picture from validation as we go. On individual athletes, a newer approach is already showing even smoother curves — we are validating that across the board before we put numbers on it. No marketing deadline; we will name the date when we see it.

And this calm curve is only the beginning. It becomes the foundation for the next step of our TrainAI — training that aligns with your real development instead of a population average.

FAQ

What's new in v3 compared to v2? v2 delivers a good single value with an uncertainty attached. v3 fuses those single values over time into one calm fitness curve — exactly what an adaptive plan needs.

What does the Kalman filter do? It treats every training session as a noisy measurement of your engine and fuses them into a smooth curve. In hard training weeks it allows more movement, so real changes in form come through.

Doesn't "calm" mean you swallow real progress? No. The curve sits at the physical floor and follows real Powertest jumps more closely than the raw estimate. It smooths noise, not form.

Is v3 more accurate than a Powertest? No. The Powertest remains the fixed star the whole v3 time series is aligned to — the most accurate single value. v3 does something different: it keeps your curve calm, current and honest between tests.

Where do the numbers come from? From more than 11,000 curated Powertest labels and pre-training on 744,871 activity streams. The week-to-week consistency is measured across 425 athletes.

Do I need a Powertest for v3 to work? No — v3 runs off your normal activities. But with a Powertest your curve sits on you instead of a population average, and one test per year keeps it calibrated.

When will v3 go live for me? It is rolling out now. We switch it on step by step as soon as live stability matches the validation picture.

Related articles

• VO2max table — where your estimate stands in context
• VLamax explained — the second engine
• Garmin VO2max accuracy — why watch numbers differ

Sources: internal powerAI codebase (model architecture, filter definition, benchmark pipeline). Data basis: pre-training on 744,871 activity streams, calibration against more than 11,000 curated Powertest labels, week-to-week consistency measured across 425 athletes (245,972 records).

Photo: Annalena Duschl