This folder holds the methodological underpinning of Dr Paul Felton’s research programme — not coaching content. None of these six papers tells you what a bowler should do. What they tell you is how much weight the rest of the catalogue can bear.
If you read only one thing: McErlain-Naylor 2021 — A review of forward-dynamics simulation models. Felton co-wrote it, and it is the authors’ own frank account of what their method can and cannot deliver.
The one-paragraph version
Felton’s headline cricket findings come from a computer simulation of one elite fast bowler, built in 2D, driven by strength measurements taken on a dynamometer, and optimised to find a “best” technique. That chain has four links, and each one has a known, measured amount of slack in it. The kinematic outputs (angles, speeds) are tight — around 1° and 1–2%. The kinetic outputs (forces, loads) are loose — 11–18%. And the final step, from “the model reproduces what he did” to “this is what he should do”, has an error that the authors state plainly cannot be quantified at all.
Simulation vs measurement
Measurement tells you what happened. Simulation tells you what would happen if the model is right.
- Felton’s simulation reproduces recorded deliveries to 5.3–5.7% overall — inside the field’s own bar of “<10%, ideally <5%”.
- But that only proves the model can recreate a delivery that was actually bowled. The 2021 review says four separate times that this “does not guarantee that any subsequently optimized movement solution is possible by the individual.”
- Trust the reproduction. Discount the prediction. A simulation-derived “optimal technique” is a mechanically-justified direction of travel, not a target.
- The authors’ own preferred phrasing for a simulation result — and the right register for any cue in this catalogue — is: “more extended front ankle and knee joint angles; increased trunk flexion; a longer delay in the onset of arm circumduction.” Direction, not number.
Confidence ladder, highest to lowest:
| Claim type | Trust | Why |
|---|---|---|
| Trunk angle / body position at a moment | High (~0.9°) | Directly measured and closely reproduced |
| Centre of mass and ball release speed | High (0.1–1.7%) | Reproduced to within measurement noise |
| Ball speed effects of technique changes | Medium | Simulation, one athlete, but mechanically causal |
| Ground reaction / peak force numbers | Low–medium (11–18% error) | Pin-joint models can’t do compliance |
| Internal loading / spinal load estimates | Low | Compliance parameters set 4–5× measured values as compensation |
| “This is the optimal technique” | Indicative only | Error is explicitly unquantifiable |
Individual-specific vs general
This is the biggest single trap in the whole catalogue.
- The simulation work is built on n = 1 — one 18-year-old elite male fast bowler, 85.0 kg, 1.94 m. Every spring stiffness, every inertia value, every strength curve is his.
- The 2021 review is explicit: even for a world record holder with near-optimal technique, “the evaluation results could not be extrapolated to other models of sub-elite athletes.”
- What legitimately transfers: the cause-and-effect mechanism — that changing X tends to change Y, and why. That is genuinely more trustworthy than a correlation from a group study, because the simulation controls everything else.
- What does not transfer: the numbers. An optimum computed for one bowler is not a benchmark for yours.
- Separately: the group studies elsewhere in this catalogue have the opposite problem — they generalise across bowlers but cannot establish cause. Neither method gives you both. Read them as complements, not as confirmation of each other.
2D vs 3D — the clearest quantitative answer in the folder
Fast bowling is a genuinely three-dimensional action, and Felton models it in 2D. Why, and what does it cost?
Why 2D: because a torque-driven model needs the athlete’s real strength at every joint angle, and that can only be measured in vivo one axis at a time. No method exists for measuring subject-specific 3D strength at the hip or shoulder. So subject-specific optimisation is stuck in 2D — not by preference, but by measurement.
What it costs (2019 controlled comparison, same bowler, same model):
| Naive flat model | Hips freed | Hips + shoulders freed | |
|---|---|---|---|
| Ground reaction force | 18% | 12% | 11% |
| Vertical force specifically | 23.0% | 13.7% | 13.6% |
| Ball release speed | 3.8% | 3.2% | 1.7% |
| Trunk orientation | 0.9° | 1.2° | 0.9° |
| Overall | 8.9% | 6.4% | 5.7% |
What a coach or system-builder should take from this:
- A flat, side-on analysis is defensible for angles and speeds, and not defensible for forces. Even the best-corrected 2D model is ~11% out on force, and ~13.6% out on vertical force no matter what you do (that residual is a pin-joint limitation, not a geometry one).
- The fix is not “buy 3D”. Felton kept 2D dynamics but fed in the measured left–right hip and shoulder separation. A single side-on camera cannot recover that separation — so this is evidence that side-on footage alone is missing something, not licence to ignore the problem.
- If you can only fix one thing, fix the hips. Freeing hip joint centres recovered most of the force error; freeing shoulders mostly helped ball speed.
- Fast bowling is the worst case for planarity, precisely because it is a side-on action where hip and shoulder separation approaches its maximum. What’s harmless in running and jumping is not harmless here.
- A model can look perfect on the output you’re checking and be badly wrong elsewhere. All three variants above matched centre-of-mass velocity to 0.1–0.2%, including the 18%-wrong one. Validate against the quantity that is sensitive to your assumption.
Where the strength numbers come from (and why it matters)
The two Parkinson papers (2022, 2023) determine how the strength inputs to Felton’s simulations are measured. In plain terms: a simulation searches for the best technique the athlete’s muscles can actually deliver. If you tell it the athlete is weaker than they are, it hands back a more conservative “optimum” — and you’d never know.
- Measuring strength at a single joint angle — a very common protocol — underestimates true peak torque by a mean of 31–36% at the worst common angles, with worst cases of 96% (knee flexion at 90°) and 80% (knee extension at 270°). The errors are all negative: the bias is systematic, always toward “weaker”.
- Dynamometer axis misalignment adds 0.3–17% on top, plus 10–15° of discrepancy between intended and actual joint angle.
- Using ~5 measurement angles straddling the expected peak recovers peak torque to ~2–5% and optimal angle to ~0.1–5%. Returns diminish sharply beyond that.
- The stubborn one: the width of the strength curve — how fast strength falls off away from the optimum — remains 14% out for the knee flexors even with maximum sampling. That parameter is exactly what decides whether a fast, large-range movement is strength-limited. So any optimal-technique claim that hinges on strength at extreme joint angles carries that uncertainty.
- Ignoring two-joint muscles (the monoarticular assumption, used in most models) cost 19% error vs 3% in the one case where it was properly tested.
Contradictions and tensions flagged
TENSION — 2016 vs 2019, how bad the planar assumption really is. The 2016 ISBS paper concluded the massless-segment method “was suitable to reproduce predominately planar movements” and reported only the best model variant, with no plain-planar baseline. The 2017/2019 controlled comparison supplies that baseline and shows the unmodified planar assumption was substantially inadequate — 8.9% overall, 18% on force, 23% on vertical force, versus 5.7% and 11%. Not a reversal, but a material revision: planarity was a real problem, not a benign simplification. Quote 2019’s numbers, not 2016’s framing.
TENSION — 2019 vs 2021, is 2D permanent or temporary? Felton, Yeadon & King (2019) argue a full 3D model is “non-viable” and present planar-with-massless-segments as the solution. The 2021 review, with Felton as an author, reframes the same fact as a temporary limitation and names 3D torque functions at the hip and shoulder as the unlock. Same evidence, different weight: solution vs stopgap.
TENSION — the 2021 review criticises the compliance settings used in Felton’s own 2019 model. Felton’s 2019 model permitted wobbling-mass movement of 4.5 cm shank / 7 cm thigh / 10 cm trunk and 6 cm vertical front-foot displacement. The 2021 review (Felton co-author) argues such values “may be excessive”, citing measured soft-tissue displacement of up to 1.4 cm and heel-pad/shoe deformation of 11.5–12.7 mm — against simulation compliance allowances of up to 56 mm. The review’s own explanation is that models are made too soft in these places to compensate for being too rigid elsewhere (pin joints, foot arch, vertebrae). This is a self-acknowledged weakness in exactly the parameters that produce the force numbers — and it is consistent with the stubborn 11–14% residual force error.
TENSION — the 2021 review’s own best-practice standard is not met by the cricket optimisations. The 2021 review states that a predicted optimum “should be robust to perturbations and not simply a single greatest one-off performance”, cites the 1000-repetition noisy-optimisation approach as the standard, and notes its incorporation in torque-driven models “has been sporadic”. The cricket fast bowling optimisations are torque-driven and did not do this. By the authors’ own criterion, the cricket “optimal technique” results are single-best-performance optima, not noise-robust ones.
TENSION — the 2021 review’s confidence in dynamometry vs the 2022 Parkinson finding. The review presents in-vivo dynamometry as the thing that makes torque-driven models trustworthy, “providing assurance that torques exerted at each joint angle and velocity within any predicted optimal technique are realistic for the individual.” Parkinson et al. (2022) show that if the protocol uses a single conventional test angle, strength parameters can be wrong by 30–96%. Reconcilable — the review assumes a full multi-angle protocol, and Felton’s cricket work did use one — but that confident framing must not be transferred to any study measuring strength at one angle.
TENSION — the objective function determines the answer. Reported in the 2021 review as a cautionary example, not as a flaw in Felton’s work, but it applies to all of it: in gymnastics, minimising joint torques produced a solution that diverged from an elite athlete’s real movement, while maximising success under movement variability produced one close to it. Same model, different criterion, opposite verdict. Whenever you read “optimal”, ask optimal for what?
Not a contradiction but worth knowing — the 2016 paper contains an evident typographical slip (“Failure to do so may result in a simulation model providing insights into the mechanics of a movement”), where the intended meaning is plainly failing to provide insights. The restrictive reading is the correct one.
Evidence strength within the folder
2019 (peer-reviewed journal, full statistics) > 2021 (peer-reviewed review, but narrative and self-referential) > 2016 / 2017 / 2022 (conference papers, 2–4 pages) > 2023 (one-page poster abstract).
All six were obtained and read in full text from the source PDFs — no file in this folder relies on an abstract alone.