Predictive Analysis and Modelling of NBA Players' Performance

Uspješno predviđanje ishoda u NBA-u, kao i parametri koji utječu na te ishode, oduvijek su privlačili pažnju kako sportskih analitičara, tako i sportskih kladionica. Kvaliteta igre NBA igrača utjecajna je brojnim nasumičnim i nepoznatim faktorima kao što su psihološko stanje, socijalni život i ozljede. Koristeći javno dostupne podatke, kreiran je podatkovni skup u kojemu svaka instanca nosi informacije dostupne prije same utakmice. Nastup igrača je klasificiran kao „over“ ili „under“ ovisno o broju poena koje je postigao u promatranoj utakmici i trenutnom sezonskom prosjeku. Koristeći napredne statistike, kreiran je binarni klasifikator s ciljem budućeg razlučivanja između navedenih klasa. Klasifikator je na test setu pokazao rezultat od AUC 59.7%. Unatoč dominaciji ofenzivnih statistika (broj odigranih minuta, broj uzetih šutova, jačina napada), naknadno kreirane značajke uvelike su poboljšale točnost modela.NBA attracts a great deal of attention among sports analysts and sportsbooks regarding the prediction of various outcomes of each game, together with the parameters which affect them. Performance of NBA players is influenced by many unknown and random factors, such as players’ psychological condition, social life and injuries. The stated factors hinder game-to-game predictions of players’ performance in relation to the expectations set by their past performances. In this thesis publicly available statistics are leveraged in order to create a dataset pertaining to the performance of a single player during a single season. A comparison between points that a player has scored and his current season average was done in order to classify the player's performance as ‘over’ or ‘under’. Using various statistical data concerning previous games of the season, a binary classifier was trained in order to distinguish between those categories for future games. The classifier performed with an AUC ROC score of 59.7%. Although the most important features are predominated by offensive statistics (e.g. how many minutes the player plays, how many shots he takes and how strong the offense of his team is), newly generated features significantly improved the performance of the model