What Acceleration Data Actually Tells You (And What It Doesn't)
Summary
Split times tell you how long an athlete took. They cannot tell you why. What sprint acceleration data actually contains, what it structurally cannot answer, and how to tell the two apart before you build a training decision on top of it.
A coach sent me a spreadsheet a few weeks ago. Fourteen athletes, 0 to 10 m and 0 to 30 m splits, tested in January and tested again in April. Seven athletes were quicker over 10 m. Nine were quicker over 30 m. The email attached to it asked which block of training had done the work.
I could not answer that from the sheet. Neither could he. Not because the data was bad, and not because he had tested carelessly. Because split times, on their own, do not carry the information that question is asking for.
This is the widest gap I see between what sprint testing produces and what coaches believe it produces. Acceleration data is genuinely useful, and I have spent a large part of my working life building the equipment that collects it. It is also routinely asked to answer questions it has no mechanism for answering. What follows is what those numbers actually contain, what they do not, and how to tell which is which before you build a training decision on top of them.
A split time is an outcome, not a mechanism
Start with the obvious thing that gets forgotten anyway. A timing gate does one job. It records the instant an athlete's body interrupts a beam at a fixed position on the ground. That is it. Everything else you read off a testing sheet is arithmetic performed on a small set of position and time pairs.
A 0 to 10 m split of 1.78 s is not a measurement of acceleration. It is a measurement of elapsed time over a distance, which is the accumulated result of every step, every ground contact, every force orientation, and every postural decision the athlete made in that ten metres. If the number improves by 0.05 s, something in that accumulation changed. The split cannot tell you what.
That matters because the interventions available to you are mechanistic. You can change how much force an athlete produces. You can change how that force is directed. You can change step length or step frequency, contact time, shin angle, trunk position, or the timing of the transition out of the drive phase. A split time collapses all of those into one scalar and hands it back to you. Working backwards from a scalar to a mechanism is guessing, and it is worth being honest with yourself about when you are guessing.
What acceleration actually is, mathematically
Acceleration is the rate of change of velocity. To describe it properly you need a velocity-time curve, not two or three points on a distance-time line.
The field-friendly way to get there, and the approach most sprint profiling now rests on, is to fit a model. Human sprint acceleration from a standing start is well described by a mono-exponential function of velocity against time. That model has two parameters. One is maximum sprinting speed, the asymptote the athlete would approach given enough runway. The other is a time constant, usually written as tau, which governs how quickly they approach it. From those two numbers you can derive maximum theoretical horizontal force, maximum theoretical velocity, peak power, and the slope of the force-velocity relationship.
Samozino and colleagues validated this approach against force plate data in 2016 and reported low bias, under 5 percent, for maximal horizontal force, velocity, and power output, along with inter-trial standard errors of measurement also under 5 percent (Samozino et al., 2016). That is a strong result for a field method, and it is why the approach spread as fast as it did.
But notice what has happened. The output of this method is not a direct measurement of acceleration either. It is a two-parameter model fitted to a handful of timing points, with the mechanical variables derived from the model and from an aerodynamic estimate based on the athlete's height and mass. It is a very good model. It is still a model, and the quality of what comes out of it depends entirely on the quality and the appropriateness of what goes in.
The acceleration phase is not one thing
Here is where the collapsing gets expensive. Coaches habitually talk about "the acceleration phase" as a single entity, and then test it with a single split.
Nagahara and colleagues tracked step-by-step kinematics across a 60 m sprint using sixty infrared cameras and looked for structural transitions in the acceleration pattern. They found two. The acceleration phase divides into an initial, a middle and a final section, and the transitions are not arbitrary statistical artefacts. They correspond to identifiable mechanical changes. At the first transition the foot begins contacting the ground in front of the centre of gravity, the knee joint starts to flex during support rather than remaining relatively extended, and the rise in step frequency stops (Nagahara et al., 2014).
So an athlete moving through 0 to 30 m is not doing one thing well or badly. They are executing at least three mechanically distinct sub-tasks in sequence. A single 30 m split averages across all three. Two athletes can post the same 30 m time with completely opposite profiles: one who leaves the blocks well and stalls through the transition, another who is slow off the mark and continues building cleanly.
If you only have 0 and 30, those two athletes are identical on your sheet and should be receiving different training. That is not a subtle measurement issue. That is the sheet actively misleading you.
Where the numbers came from determines what they mean
The single largest source of nonsense in sprint testing is comparing numbers that were never comparable.
Haugen and Buchheit's review of sprint monitoring methodology laid this out in detail, and the size of the effects is worth sitting with. Timing system type, gate height, whether the system uses single or dual beam detection, the start position, the distance the athlete stands behind the first gate, the surface, the footwear, the wind, the time of day, and the warm-up protocol all move sprint times by amounts that are comparable to or larger than a season's worth of genuine adaptation (Haugen and Buchheit, 2016).
A single beam system can be triggered by a swinging arm or a lead thigh rather than the torso. Dual beam detection with proper reflector geometry exists precisely to reduce that error, and it is one of the reasons we build the G4 gates the way we do. But the deeper point is not about any one product. It is that a 10 m time collected with a standing start 0.5 m behind the gate, on a single beam system, is a different quantity from a 10 m time collected from a three-point start with the athlete's hand on the line on a dual beam system. Putting both in the same column and drawing a trend line through them is arithmetic applied to incommensurable measurements.
The flying start problem
There is a specific version of this worth calling out because it contaminates modelled profiles rather than just raw times.
Most sprint profiling protocols use a standing start with the athlete positioned some distance behind the first gate, to avoid false triggering. That gap means the athlete already has non-zero velocity when the clock starts. The mono-exponential model, however, assumes the clock starts at zero velocity.
Jovanovic ran a simulation study on exactly this, and a follow-up study applied the same question to athlete data. The uncorrected model introduces systematic bias into the estimated sprint parameters, and two correction approaches were proposed, one estimating the time correction and one estimating the flying distance, which substantially reduce or remove that bias (Jovanovic, 2024; Jovanovic et al., 2024).
The finding that should give everyone pause is the one flagged as unexpected. For maximum sprinting speed, the uncorrected model, despite being biased, was the most sensitive at detecting genuine change. The authors were blunt about the rest: every other parameter and model combination showed an unsatisfying level of sensitivity (Jovanovic et al., 2024). Bias and sensitivity are not the same property, improving one can cost you the other, and on current evidence the ability of these models to detect real change in anything other than maximum speed is not something to lean on hard.
That is an uncomfortable result and I do not want to smooth it over. The practical reading is this: pick a correction approach, document it, and never change it mid-season. Your absolute values may be systematically off by a known amount. Your change scores will at least be internally consistent, and change scores are what you actually make decisions on.
How big does a change have to be before it is real
Back to the coach's spreadsheet. Seven athletes quicker over 10 m, nine quicker over 30 m.
Typical within-session coefficients of variation for well-controlled electronic timing over short sprints sit in roughly the 1 to 2 percent range, and Haugen and Buchheit's review discusses the reliability and sensitivity considerations at length (Haugen and Buchheit, 2016). On a 1.78 s 10 m split, 1.5 percent is about 0.027 s. Two standard deviations of that noise is roughly 0.05 s before you have added any between-session variation from surface, footwear, temperature, or where in the training week the test fell.
A 0.06 s improvement on a 10 m split, measured once in January and once in April, is not obviously distinguishable from measurement noise. It might be real. You cannot tell from one pair of trials.
The fix is not more expensive equipment. It is more trials and an honest reference point. Three to five trials per athlete per session, keep the best or the mean but decide which in advance and never switch, and establish your own typical error from repeated testing on your own athletes with your own protocol on your own surface. Published reliability figures are a starting estimate, not a substitute for knowing your own noise floor.
What acceleration data cannot tell you
Even with clean protocol, sufficient trials, and appropriate modelling, there are questions this data structurally cannot answer. Being clear about the boundary is more useful than pretending it is not there.
It cannot separate force magnitude from force orientation. This is the big one, and Morin and colleagues demonstrated it directly. They computed the ratio of horizontal force to total ground reaction force during acceleration, and an index of force application technique defined as the slope of that ratio against speed. The index correlated significantly with 100 m performance. Net horizontal force averaged over the acceleration also correlated with performance. Total force averaged over the acceleration did not (Morin et al., 2011). The orientation of the force an athlete puts into the ground mattered more to sprint performance than the amount of force they were capable of producing.
Vertical force deserves a caveat here, because the same paper contains a result that is easy to misquote. Averaged across the acceleration phase, vertical force was not related to performance. Measured specifically at top speed, it was. The distinction is between the acceleration phase and the top speed phase, and it matters: the case for prioritising horizontal force is a case about acceleration, not a claim that vertical force is irrelevant to sprinting.
An athlete who improves their 10 m split may have got stronger, or may have learned to direct existing force more horizontally. Those two athletes need different programmes. Split times cannot distinguish them. Modelled force-velocity profiles get you partway there by separating the force-dominant end of the profile from the velocity-dominant end, which is precisely the use case Morin and Samozino set out for individualised training prescription (Morin and Samozino, 2016). But even the modelled ratio of forces is derived, not measured, and it inherits every assumption in the model.
It cannot tell you about left and right leg contribution. A sprint split is a whole-body outcome. An athlete with a meaningful asymmetry in horizontal force production between limbs can post an entirely unremarkable split time. This matters most in return-to-sport contexts, where the asymmetry is the thing you actually care about and the split time is the thing you happen to be able to measure.
It cannot attribute a change to a cause. Testing before and after a training block tells you the outcome changed. It does not tell you the block caused it. Growth, maturation, technical coaching from a different source, changed footwear, a better night's sleep and regression to the mean are all live alternative explanations, particularly in youth and particularly with a single pre and post measurement.
It cannot tell you about fatigue state unless you designed for it. A profile collected fresh and a profile collected at the end of a heavy week are answering different questions. Neither is wrong. Comparing them as if they were the same test is.
Where direct measurement changes the picture
There is a class of these limitations that comes down to the difference between inferring force and measuring it.
The Samozino method infers horizontal force from a velocity-time model plus body mass plus an aerodynamic estimate. That inference is well validated and it works, which is why it is now standard practice. But it is an inference, and it produces a single whole-body value per trial.
Instrumenting the resistance directly is a different approach to the same question. A load cell in a friction-resistance sled measures the horizontal force at the point of application rather than deriving it from a model of the athlete's motion. That is the design logic behind DynaSled, and the reason it can report left and right leg force contribution separately during a push: the force is being measured while it is being applied, not reconstructed afterwards from where the athlete ended up.
The honest framing is that these are complementary rather than competing. Timing gates measure time at a position with high precision and always will, and free sprinting is the criterion task for a sprinter. Instrumented resistance measures force during a constrained task. Force plates measure vertical ground reaction force in a fixed location. Each answers a different question, and none of them makes the others redundant. What matters is knowing which question you are asking before you pick the tool.
DynaSled is new to market and I am not going to claim otherwise or point to adoption numbers that do not exist yet. The measurement argument stands on its own regardless.
What to do with this on Monday
Six things, in order of how much they will improve your data per unit of effort.
Write down your protocol and stop changing it. Start position, distance behind the first gate, gate height, number of trials, rest between trials, warm-up, surface, footwear. Every one of those is a variable that moves your numbers more than a training block does. Fix them.
Add a split. Going from 0 and 30 to 0, 5, 10, 20 and 30 costs you very little and is the single highest-value change most testing setups can make. It lets you see whether a change happened in the initial drive, through the transition, or in the late acceleration section, which maps onto the phase structure Nagahara and colleagues identified. That is the difference between knowing something changed and knowing what changed.
Run enough trials to see through the noise. Three minimum, five is better. Decide in advance whether you are using best or mean and never switch.
Establish your own typical error. Test the same athletes twice within a week with no training stimulus in between. The spread you get is your noise floor. Any change smaller than roughly twice that is not something you should act on.
Model, but declare your assumptions. If you are computing force-velocity profiles, state which flying start correction you are using and keep it constant. Report the model parameters alongside the derived variables so a future reader, including future you, can tell what was measured and what was computed.
Ask what mechanism you are trying to observe before you test. If the answer is "whether they got faster", a split is fine. If the answer is "why they got faster" or "which limb is limiting them", a split will not get you there and no amount of statistical treatment will rescue it.
Key takeaway
Acceleration data is an outcome measure wearing the clothes of a mechanistic one. Split times tell you reliably and precisely how long an athlete took to cover a distance. Modelled sprint profiles extend that into useful mechanical estimates, validated to within a few percent of force plate criterion values, and they are among the best tools the field has for individualising speed training.
What none of it does is tell you why. The acceleration phase contains at least three mechanically distinct sections. Force orientation matters more to sprint performance than force magnitude, and a split time cannot separate the two. Between-limb contribution is invisible to a whole-body timing measurement. And a large fraction of the change you observe between two testing sessions is protocol variation and measurement noise rather than adaptation.
None of that is an argument against testing. It is an argument for testing with a clear question, a fixed protocol, enough trials to see past your own noise floor, and honesty about the boundary between what you measured and what you inferred. The coaches who get the most out of this data are not the ones with the most expensive equipment. They are the ones who know exactly what their numbers can and cannot support, and who stop the analysis at that line.
References
- Nagahara R, Matsubayashi T, Matsuo A, Zushi K. Kinematics of transition during human accelerated sprinting. Biology Open. 2014;3(8):689-699. https://journals.biologists.com/bio/article/3/8/689/1118/Kinematics-of-transition-during-human-accelerated
- Haugen T, Buchheit M. Sprint Running Performance Monitoring: Methodological and Practical Considerations. Sports Medicine. 2016;46(5):641-656. https://link.springer.com/content/pdf/10.1007/s40279-015-0446-0.pdf
- Samozino P, Rabita G, Dorel S, Slawinski J, Peyrot N, Saez de Villarreal E, Morin JB. A simple method for measuring power, force, velocity properties, and mechanical effectiveness in sprint running. Scandinavian Journal of Medicine & Science in Sports. 2016;26(6):648-658. https://onlinelibrary.wiley.com/doi/10.1111/sms.12490
- Morin JB, Edouard P, Samozino P. Technical Ability of Force Application as a Determinant Factor of Sprint Performance. Medicine & Science in Sports & Exercise. 2011;43(9):1680-1688. https://journals.lww.com/acsm-msse/Fulltext/2011/09000/Technical_Ability_of_Force_Application_as_a.11.aspx
- Morin JB, Samozino P. Interpreting Power-Force-Velocity Profiles for Individualized and Specific Training. International Journal of Sports Physiology and Performance. 2016;11(2):267-272. https://pubmed.ncbi.nlm.nih.gov/26694658/
- Jovanovic M. Bias in estimated short sprint profiles using timing gates due to the flying start: simulation study and proposed solutions. Computer Methods in Biomechanics and Biomedical Engineering. 2024;27(2):145-155. https://www.tandfonline.com/doi/full/10.1080/10255842.2023.2170713
- Jovanovic M, Cabarkapa D, Andersson H, Nagy D, Trunic N, Bankovic V, Zivkovic A, Repasi R, Safar S, Ratgeber L. Effects of the Flying Start on Estimated Short Sprint Profiles Using Timing Gates. Sensors. 2024;24(9):2894. https://pmc.ncbi.nlm.nih.gov/articles/PMC11086264/
Mark Fisher
More from Sports Science
