Tools & Research

Different Twins
Two Calculators,
Two Engines

Ten years ago, Hurdles: Eye, Tape Measure and Stopwatch gave coaches tables to put split times into context. Today, two new calculators go further: they can estimate a final time from touchdowns actually observed during a race. The stopwatch hasn't changed — what we can extract from it has.

A Finishing Time Tells Us How Fast, Not Necessarily How

Knowing that a woman ran the 100 m hurdles in 13.50, or a man ran the 110 m hurdles in 13.50, tells us the result — not how the performance was constructed. Two identical times can emerge from opposite distributions. One athlete gains time early and loses it after the last hurdles; another enters the rhythm more slowly, improves through the middle, and reaches the same finish time.

A final time combines everything. Touchdowns open the race up. A touchdown is the cumulative time from the start signal to the lead leg's contact with the track after clearing each hurdle — TD1 is the time to landing after H1, TD4 after H4, and so on through TD10.

Even a handful of accurately measured touchdowns can tell us a great deal. But the next question is inevitable: against what should we compare them? A split to H4 is fast or slow only relative to the athlete's level and the distribution we expect. This is where models become useful.

A Map That Can Be Read in Two Directions

Think of a map. The women's 100 m hurdles calculator resembles a large map that can be read both ways. Knowing the destination — the final time — we can see the race pattern generally associated with that level. Knowing several points the athlete actually passed, we can use the same map to find the final time most compatible with that route. For the 100 m hurdles, the reference world is the Kimura model: one statistical structure, read in either direction.

The 110 m hurdles calculator works differently. There is no single map that performs both tasks well, so it uses two specialised models: Shibayama describes the usual distribution of a race at a given level, and Iwasaki estimates the final performance compatible with observed touchdowns. The Iwasaki prediction is then placed on Shibayama's descriptive map so the associated race pattern can be shown.

With the Shibayama cartridge, we ask: what does a 13.50 race usually look like? With the Iwasaki cartridge: given these measured touchdowns, what final time fits them best? We then return to the descriptive cartridge to display the profile for that predicted time. We didn't force both calculators into an identical design — each tool is adapted to the literature available for its event.

Describing and Predicting Are Not the Same Thing

A descriptive model asks how a race is generally distributed for an athlete at a given level. A predictive model starts from actual measurements and asks which final time is statistically compatible with the observed touchdowns. The two are related but distinct — which is why Shibayama and Iwasaki can return slightly different values without contradicting each other.

What Regression Actually Means

Regression sounds more intimidating than the idea behind it. Take the 151 performances analysed by Iwasaki, Nunome and Nozawa, each with a final time and touchdowns from H1 to H10. Plot final time against TD4 and every performance becomes a point in a cloud: faster athletes generally land after H4 sooner, but the points don't line up perfectly — two 13.50 hurdlers may land several hundredths apart and build their race differently.

A linear regression finds the line that best represents that relationship: predicted TD = intercept + coefficient × final time. Enter 13.50 and it returns the statistically expected TD4 for that level in the studied population — an association, not a technical prescription. The scatter around the line is the residual variation, or noise.

R² (the coefficient of determination) tells us how informative that regression is. In Iwasaki's study, values ranged from around 0.53 to close to 0.98 across the different touchdowns — and predictive power rose steadily as the race unfolded. The closer to H10, the fewer widely different final times remain compatible with the observed split.

One Metre After the Start, One Metre Before the Finish

Push this to an extreme. Predicting the final result from the first metre after the start tells us almost nothing — we might be watching a pure sprinter who leaves the blocks brilliantly but has never hurdled, next to an experienced hurdler who doesn't need an aggressive start to qualify. After one metre, the sprinter could look more promising. The observation alone says almost nothing about the 110 m hurdles result.

Now move to one metre before the finish. If both athletes arrive together, almost all the race is behind them, and there's little room left for radically different results. Touchdowns follow the same logic: a TD1 consistent with 13.50 leaves nine hurdles still to run; a TD7 carries much more information; a TD10 leaves only the run-in. This doesn't mean the algorithm assigns TD10 a greater mathematical weight — rather, TD10 has a closer statistical relationship with final time, and so contains more predictive information.

Coaches recognise this instinctively. An excellent first hurdle is interesting but guarantees little. Arriving at H7 in a rhythm consistent with a very fast performance means much of the result has already been constructed — maintaining that compatibility through H10 narrows the possible finish times further. Mistakes remain possible: a hurdle can be hit, posture lost, the run-in mismanaged. The point is simply that less race remains to be run.

The Two Engines of the 110 m Hurdles Calculator

Shibayama et al. (2019) analysed 76 high-level male hurdlers across three groups — 12.94–13.38, 13.40–13.68 and 13.70–14.16 — using high-frequency video to study how interval times and distributions differ between levels. The calculator normalises and interpolates the group means: for a target of 13.50, it generates a representative pattern including approach, intervals, cumulative times and average speeds. These are benchmarks, not prescriptions, and the target model is used within the 12.94–14.16 range rather than extrapolated beyond its source population.

In Athlete Analysis, the question changes: not what a 13.50 race ought to look like, but which final time best matches the measured touchdowns. Iwasaki's study includes 151 performances, roughly 12.91–14.47 seconds, fitting a separate regression for each touchdown from TD1 through TD10, with leave-one-out cross-validation used to check generalisation. With several input touchdowns, the calculator searches for the final time that best fits those relationships overall — the most compatible performance within the data the model was built on, not a certainty.

A Practical Example: 13.50 Target, 13.46 Projection

The following numbers are invented for illustration, not the actual study coefficients.

Suppose the four regressions, considered separately, point towards 13.40, 13.44, 13.46 and 13.47. The calculator doesn't take the arithmetic mean — it looks for the final time that best fits the whole set of observations, and might return 13.46. The accurate reading is: the combination of measured touchdowns is most compatible with a performance of approximately 13.46 in the population the model was built on. It doesn't mean the athlete will certainly run 13.46.

Why Descriptive and Predictive Outputs Can Differ

If the coach enters a 13.50 target, Shibayama describes a representative 13.50 pattern. Actual touchdowns might lead Iwasaki to estimate 13.42 or 13.61. There's no contradiction — the models answer different questions, built on different samples, sizes and statistical procedures, so even an identical final time won't produce touchdown values that match to the hundredth.

It's tempting to calculate each interval by subtracting one predicted cumulative value from the next. Mathematically that can be done; methodologically it's less clean, since the two values come from independently fitted regressions and the paper didn't model that interval directly. Instead, once Iwasaki identifies the best-fitting final time, the calculator uses Shibayama's descriptive model to display the corresponding pattern: observed touchdowns → Iwasaki regressions → estimated final time → Shibayama race pattern. Each model is used for the purpose it was designed to serve.

What RMSE Tells Us

The projection also displays RMSE (root mean square error), summarising the distance between observed touchdowns and the model's predictions at the fitted result, in seconds. A low RMSE doesn't mean a better athlete — a 14.00 hurdler who closely matches the average 14.00 pattern can obtain a very low RMSE, while a 13.20 hurdler with an unusual distribution might produce a higher one. RMSE indicates compatibility with the model, not absolute performance, which is why we've avoided labels like "excellent" or "poor" for thresholds that haven't been specifically validated.

A More Unified Engine for the Women's 100 m Hurdles

The 100 m hurdles calculator relies on Kimura et al. (2026), which covers a wide range of performance levels — useful because a slower hurdler doesn't necessarily run the same race as a 12.50 performer with every split stretched by the same percentage. The moment maximum rhythm is reached, and how speed is maintained or lost, can vary with level, so simply scaling a 12.50 pattern to describe 16.50 would oversimplify the event.

The calculator normalises and interpolates Kimura's mean patterns between anchors from 12.57 to 18.49 seconds. Entering 14.20 generates an expected profile without merely multiplying splits by a constant, and Athlete Analysis compares measured touchdowns with the available patterns to find the best-fitting final time. The same statistical reference serves both directions.

Different Twins, Side by Side

A Calculator Is Not a Coach

The tools are reference points, not commands. The average pattern isn't necessarily ideal for every individual — anthropometrics, acceleration ability, clearance skills, strength and rhythm endurance can all lead to different successful solutions. If a 13.50 athlete consistently differs from the model while performing at a high level, the coach shouldn't force them to resemble the mean. Instead, the difference is worth investigating: a limitation, an individual characteristic, or perhaps a strength.

Athlete Analysis doesn't require all ten touchdowns. Coaches can enter only the points actually measured — H1, H4, H7 and H10, for instance — without inventing values for the rest. The model estimate stays distinct from the observed measurement, which keeps the tool usable even with straightforward video analysis or trackside timing.

Three Zones as an Operational Language

Both calculators summarise the central race in three zones: H1–H4 (Acceleration), H4–H7 (Max Rhythm) and H7–H10 (Rhythm Endurance). These are operational labels, not rigid physiological phases. H1–H4 captures how rhythm is built and stabilised after H1; H4–H7 describes the section often closest to sustainable maximum rhythm; H7–H10 assesses the ability to maintain speed and effectiveness. H10–finish is shown separately, since it combines the final landing and the run-in.

Imagine two athletes targeting 13.50, both finishing in 13.52 — two hundredths off, but their zone deviations tell different stories:

Athlete A enters the race slowly and recovers progressively; Athlete B builds an early advantage and gives it away over the last hurdles. Nearly identical final times can hide very different coaching questions.

Limits Matter as Much as Possibilities

These models are only useful within their boundaries: 12.57–18.49 for the implemented women's model, 12.94–14.16 for the Shibayama target, and 12.91–14.47 for the Iwasaki prediction. Extrapolating far beyond those ranges means asking a model to describe populations it has never observed. Regression also can't account for wind, hurdle contact, technical mistakes, track conditions, pacing choices or every athlete-specific characteristic — a result from the calculator is an informed estimate, not a verdict.

The aim is to use research to ask better questions of performance while preserving the coach's essential role. Few coaches arrive at training intending to apply a linear regression validated through leave-one-out cross-validation — but they can film a race, measure four touchdowns, enter them in a calculator, and immediately see where an athlete gains or loses relative to a target. Ten years ago we mainly had eye, tape measure, stopwatch and tables. Today, descriptive and predictive models can join that toolkit — and this is, as far as we can tell, only the beginning.

Main References

Kimura, K., Sato, K., Miyauchi, I., Hase, S., Kakehata, G., Goto, Y., & Iso, S. (2026). Temporal characteristics of 100-m hurdles performance based on race analysis — ranging 12–18 s. Scientific Journal of Sport and Performance, 5(2), 347–354. doi.org/10.55860/WMQR1877

Shibayama, K., Kijima, K., Morioka, Y., & Sakurai, K. (2019). The relationship between race pattern and performance of elite 110 m hurdlers. Japan Journal of Physical Education, Health and Sport Sciences, 64, 475–485. doi.org/10.5432/jjpehss.18120

Iwasaki, R., Nunome, H., & Nozawa, K. (2022). New predictive model of the touchdown times in a high level 110 m hurdles. PLOS ONE, 17(12), e0278651. doi.org/10.1371/journal.pone.0278651