Archery talent and neural networks: Taha 2018

Almost certainly the same 50 young Malaysian archers as in the paper on SVM classifiers (same number, same make-up, same mean age), analysed here with a neural network: the model recognised the high- or low-potential group with 94% accuracy, 96.67% sensitivity and 90% specificity, an error rate of 6%. The values imply that the high group held 30 archers and the low group 20. The archers, 37 boys and 13 girls aged 13 to 20, had done six fitness tests (grip strength, vertical jump, standing broad jump, static balance, arm strength and core strength) and shot one end of six arrows at 50 metres.

As in the other paper, the two groups were formed first, by combining fitness tests and score, but here with a different algorithm, k-means. In the charts the high group does better on every measure. The authors conclude that the network classifies the archers' potential well and that this could help coaches spot talent.

Before testing the archers jogged for 5 to 10 minutes and did some stretching. The shooting test came before the fitness tests, on a range set up like a competition at 50 metres: four practice arrows and six scored. The fitness tests are the standard ones: sit-ups, push-ups, one-leg balance (standing stork), a dynamometer for grip, standing broad jump and vertical jump.

The classifier is a neural network with a single hidden layer and a sigmoid activation function. The authors tried from 1 to 50 neurons and kept 10. The analysis was run in Orange 2.7 and assessed on accuracy, sensitivity, specificity and error rate. The abstract refers to five-fold cross-validation, the methods section to ten-fold.

The same limitations apply as for the SVM paper, because the data are, in all likelihood, the same: the groups are built from the measures later used to recognise them, the score is a single end, nobody was followed over time, and "high potential" corresponds to no actual outcome. On top of that there is only one model here, with no comparison against simple methods such as regression, and with 50 archers a 10-neuron network has many parameters relative to the data. Nor is there a group of archers held back for a final check.

The text is short, a conference paper, and contains inconsistencies: five or ten folds for cross-validation, and k-means here against hierarchical clustering in the other paper to form groups from the same data, without saying whether the groups match. The bows the archers used are not stated. It should not be counted as an independent confirmation of the SVM paper: in all likelihood it is the same sample analysed in a different way.

For a coach the practical contribution is close to nil: the study shows that an algorithm can recover two groups built from the same numbers it is given, not that a battery of fitness tests can tell who will become a good archer. The 94% should not be read as the chance of spotting a talented archer.

One piece of context remains: in this group of young Malaysians, those who scored more over six arrows also tended to have better fitness results. That fits the idea that a base of strength helps a young archer, but with this design it is impossible to say how much it matters, or whether strength comes before practice or the other way round.