Archers' coping skills and k-NN: Taha 2018

Fifty young Malaysian archers filled in a questionnaire on psychological coping skills for competition and then shot one end of six arrows at 50 metres. The authors split the group in two with an automatic clustering algorithm (k-means), labelling the two sets highly and poorly psychologically prepared archers, and then tried six variants of a machine-learning classifier, k-nearest neighbour, to see how well the psychological scores alone could recover which group each archer belonged to.

The best variant, using cosine distance, classified 94% of the archers correctly, with a Matthews coefficient of 0.832; the worst, with 100 neighbours, stopped at 76%, putting every archer in the same group. According to the figures, the better-prepared group had higher mean values on six of the seven questionnaire scales, the exception being freedom from worry. The authors present the method as a way of identifying talent at little cost in time and money.

The archers, 37 boys and 13 girls aged 13 to 20, came from university and state development programmes in Malaysia. Before shooting they completed the Athletic Coping Skills Inventory, which gives seven scores: coping with adversity, coachability, peaking under pressure, concentration, freedom from worry, confidence and achievement motivation, and goal setting and mental preparation. The shooting was a simulated competition: four practice arrows and then six scored arrows at 50 metres.

According to the figure and the discussion, the two classes were formed with k-means on eight variables, namely the seven scales and the shooting score. The classifiers, with 1, 10 or 100 neighbours and different distance measures, were trained on the psychological variables alone and evaluated with five-fold cross-validation, measuring accuracy, sensitivity, specificity, precision, error rate and the Matthews coefficient.

The main problem is that the reasoning goes round in a circle. The two groups are not defined by an external outcome, such as competition score or national team selection, but by an algorithm that formed them from eight variables, seven of which are the psychological scales themselves; a classifier is then asked to recover them from the same scales. Under these conditions high accuracy is to be expected and says little about the ability to predict who will become a good archer. The shooting score, just six arrows, is only one of those eight variables, and the paper does not show how far apart the two groups were on the target.

The sample is small and unbalanced: from the percentages in the table the groups come to about 38 and 12 archers, and a classifier that put everyone in the larger group would already score 76%, the same as the worst variant. There are also signs of carelessness: the standard deviation of age is 0.56 in the text and 0.056 in the abstract, values hard to square with ages from 13 to 20; the method twice refers to SVM, a different algorithm, probably left over from an earlier paper by the same authors; and the sensitivity given for the cosine variant (97.87%) does not fit the other values in the same row. It is a short engineering conference paper with no individual data.

For an archer or coach the study offers no ready-made tool. It shows that in a group of young archers the seven scales of the coping skills questionnaire move together closely enough to separate those with high scores from those with low ones, which can be seen without any machine learning. It does not show that those scores predict future results, because nobody was followed over time and shooting was measured on six arrows.

The questionnaire itself can give a coach a starting point for talking with an athlete about concentration, confidence or handling pressure. Using it to decide who stays in a programme, on the strength of this study, would go beyond what the data allow.