Coping skills and archer potential: Taha 2018 SVM

Fifty young Malaysian archers, 37 boys and 13 girls aged 13 to 20, filled in a questionnaire on psychological skills in sport and then shot a single end of six arrows at 50 metres. The authors split them into two groups with an automatic clustering algorithm (k-means), calling them archers with high and low psychological preparation, and then trained two machine learning models, so-called support vector machines (SVM), to recognise each archer's group from the seven scales of the questionnaire.

The linear model assigned the right group in 92% of cases, so it misplaced 4 archers out of 50: it classified everyone in one group correctly and 73.33% of the other. The second model, with a radial kernel, reached 70% by putting every archer in the same group, which is why its sensitivity is zero. In the charts the high-preparation group scores higher on six of the seven scales, the exception being freedom from worry. The score for the end appears in only one of these charts, with no figures and no test comparing the groups.

The questionnaire is the Athletic Coping Skills Inventory (ACSI-28), which measures seven skills: coping with adversity, coachability, peaking under pressure, concentration, freedom from worry, confidence and achievement motivation, and goal setting and mental preparation. The archers completed it before the shooting test, which was set up as a simulated competition: four practice arrows and then six arrows at 50 metres, the only end recorded. The paper does not say what bows were used or what target face; it says the archers were in university and state development programmes.

To form the groups the authors used k-means with two clusters (Orange 2.7). The text is unclear about which variables went into the clustering: the figure shows the seven scales together with score, and the discussion says the score made it possible to cluster the archers. Two SVM models, one linear and one with a radial kernel, were then trained in MATLAB on the seven scales and assessed with five-fold cross-validation: in turn, four fifths of the archers are used to train the model and the remaining fifth to test it.

The main limitation lies in the design: the groups were built, at least in part, from the same questionnaire scales that the model then uses to recognise them. High accuracy therefore shows that the algorithm recovers a split made from the same numbers; it does not show that the questionnaire predicts who shoots better, still less who will become a strong archer, since nobody was followed over time and the potential in the title is never measured. The score comes from just six arrows, too few to separate archers of similar level, and the paper does not say which bows were used.

Some figures do not add up. Mean age is given as 17.0 years with a standard deviation of 0.56 (0.056 in the abstract), hard to square with ages from 13 to 20. For the linear model a specificity of 100% means that no archer was put in the positive group by mistake and therefore a precision of 100%, not the 89.74% reported; for the other model a precision of 70% cannot be calculated with zero sensitivity. The values imply groups of 35 and 15 archers, and the second model's 70% is exactly the share of the larger group. The authors themselves describe the study as preliminary and the sample as limited.

For an archer or a coach the study does not show that a psychological skills questionnaire can pick out talented young archers: the title promises more than was measured. The ACSI remains a well-known questionnaire that can help in talking with an archer about where they feel least solid under pressure, but this work adds no evidence that its scores go with arrows in the middle.

In the registry the study counts mainly as an example of a common problem: a machine learning method applied to very little data, with a numerical result that sounds precise and a starting question, which archers have more potential, that is left unanswered.