Mental skills and archery talent: Musa 2018 study

The chapter, part of a short book on machine learning in sport, uses the answers of 50 archers to a questionnaire on psychological coping skills in competition, the Athletic Coping Skills Inventory (ACSI), to tell high-performing from low-performing archers. First the group was split into two classes by an automatic clustering algorithm (k-means); then two neural networks tried to recover each archer's class.

The better network, using a hyperbolic tangent activation function, classified 94% of the archers correctly (sensitivity 94.29%, specificity 93.33%, Matthews coefficient 0.8608). The other, with a logistic function, reached 84% by putting almost everyone in the high class: it misclassified 8 of the 15 archers in the low class. The figures imply classes of 35 and 15 archers. In the plots the high class scores higher on six of the seven scales and on shooting; the exception is freedom from worry, which is higher in the low class. The authors conclude that psychological skills set the better archers apart and that the network could help identify talent.

The questionnaire yields seven scores: concentration, coping with adversity, peaking under pressure, coachability, freedom from worry, goal setting with mental preparation, and confidence with achievement motivation. The archers completed it before a shooting test; in a pilot run the reliability of the scales ranged from 0.67 to 0.82. The number of classes, two, was chosen from the silhouette score in Orange 3.11. The networks had 8 inputs, 26 neurons in the hidden layer and one output, and were assessed with five-fold cross-validation.

The chapter does not describe the archers (age, sex, level, bow) or the shooting test, apart from a plot axis that refers to a single end. The same authors have published other analyses with the same questionnaire, the same number of archers and two classes of the same size, on 50 young Malaysian archers aged 13 to 20: in all likelihood these are the same data.

The chapter reports a classification exercise more than a study of archers. The two classes do not come from an external criterion, such as competition level or results over time: the algorithm created them from the same data. There are seven psychological scales, but the networks have eight inputs and the chapter does not say what the eighth is; if it were the shooting score, the network would partly be recovering information it already had. A network with 26 neurons trained on 50 cases risks memorising the sample, and cross-validation on such small numbers cannot rule that out.

The shooting score is a single end, which says little about an archer's level. One figure in the table does not add up: precision for the second network is given as 12%, whereas the stated numbers would put it at around 81%. Freedom from worry being higher in the low class contradicts the authors' own reading, since on the same page they describe freedom from worry as important for an archer.

For archers and coaches the chapter says something limited: in a group of young archers, those who scored higher on almost all the questionnaire's psychological skills were also in the class with better shooting. It does not say whether training those skills improves shooting, or whether the questionnaire can predict who will become a strong archer: that would require following archers over time.

The 94% accuracy measures the network's ability to recover two groups built from the same data, not its ability to spot talent. The finding on freedom from worry, higher among the archers in the low class, is the only one that runs the other way, and the chapter offers no consistent explanation for it.