運動賽事預測模型在近年運動市場蓬勃發展之下，已漸漸成為一個重要的議題。建立一個好的運動賽事預測模型，在運動博弈、球隊管理營運、媒體轉播管理上，均具有非常大的助益。現存相關研究較少使用機器學習方法進行運動賽事預測模型之建構。本研究提出一個以機器學習方法為基礎之運動賽事預測模型，使用美國職業籃球聯盟（National Basketball Association，NBA）2018-2019賽季所有賽事之攻守數據做為資料庫，預測每場賽事最終球隊總得分。本研究使用相關研究較常採納之十五項攻守數據，自最具公信力之籃球數據網站Basketball-Reference.com擷取NBA 2018-2019賽季所有賽事之球隊攻守數據。攻守數據經過標準化後，將資料進行特徵工程（Feature Engineering），相關研究均以單一賽事延遲（Game-Lag）作為特徵建構方式，本研究採納六場賽事延遲，提高研究之完整性。本研究建構之預測模型包括二種預測程序（Process）。第一個程序由二階段預測模型組成：第一階段模型由四種機器學習方法及一種無母數回歸方法建構，經過特徵建構後之十三個預測變量進入第一階段模型進行建模及預測，並且在第一階段模型分析時，選擇預測效果較佳的賽事延遲數量。第二階段模型由特徵選擇（Feature Selection）開始，本研究採用三種內嵌式（Embedded）特徵選擇方法，從十三個預測變量中，經由排序法（Ranking），選出六個重要預測變量，進行第二階段模型建模及預測。最後將五種方法、二階段模型共計十種不同預測結果進行比較。透過實證資料，本程序探討不同賽事延遲對預測模型之影響，得到預測效果較佳之賽事延遲資訊，並且發現二階段模型能夠取得較佳之預測結果。第二個程序著重於特徵工程中，對於特徵之組成，本研究採用適應性權重（Adaptive Weighting），將距離目標賽事較近之參考數據賦予較高之權重，並且與不同之賽事延遲進行配對，完成不同之特徵組合（Feature set），作為五種機器學習方法組成之預測模型之輸入變項，經由本程序之實證結果，得到預測效果較佳之適應性權重及賽事延遲之組合。本研究藉由二種程序所得到之實證結果，證實適當的選擇賽事延遲及適應性權重，對於建構機器學習為基礎之籃球賽事比數預測模型具有提升的效果。
Sports outcome prediction became an important topic due to the rapid growth of sports market. A good sports outcome prediction model is a great encouragement on sports lottery, team/club management and operation, and broadcasting management. Few existing research used machine learning method to construct the prediction model. This research proposed a sports outcome prediction model based on machine learning methods. This model obtains promising performance on basketball games and expected to apply on another sports event.
This research build a sports outcome prediction model based on machine learning (ML) methods. Using statistics from all the basketball games of National Basketball Association (NBA) 2018-2019 season as our dataset to predict the final score of candidate basketball game.
This research collected 15 basketball statistics which most commonly used by related research from Basketball-Reference.com, the most reliable basketball statistic provider. The data normalization shall be implemented before the process of analyze since they different variables have different scales. This research propose two process in the proposed prediction scheme. First process , namely “feature selection with fixed weighting process” is a two-stage model start with applying game-lag information with fixed weighting on feature construction. Most related literature applied single game-lag information as their feature construction. This process constructed the feature from one to six game lag. First stage model constructed by four machine learning methods and one nonparametric regression method. Thirteen features are used in the first stage model construction and prediction. The proper game-lag selection was determined by the best prediction performance on first stage model. Second stage model started by feature selection. This research used embedded feature selection by three methods. Six features are selected by ranking methodology and used as predictor in second stage model. Finally, ten results, from five methods on two stage model each, are compared. The empirical results of this process reveal the impact of different game lag information on the performance of prediction model. Thus to determine the proper game lag selection. We also obtain better prediction performance through two stage model. The second process, namely “adaptive weighting process” start with applying adaptive weighting technique with game-lag information on feature engineering. The nearer datapoint to the target game is assigned with higher weighting. The prediction model of this process involve five machine learning algorithms. The empirical results shows that implementing adaptive weighting techniques with game-lags has the positive influence on the prediction performance.
