Through data analysis and predictive modelling I attempt to predict NHL regular season game outcomes by utilizing opposing teams prior performance. Through the analysis I will investigate established key performance indicators and their use as predictors of team success?
NHL data is available via REST API from https://statsapi.web.nhl.com a gitlab repository of all available endpoints and variables contained therein is available here: https://gitlab.com/dword4/nhlapi
All available regular season game data points from the beginning of the 2010 season up to the early 2018 season were retrieved via the API.
I utilized pythons pandas, json and requests libraries to extract and transform game data from the NHL API for analysis and modelling
A variety of NHL data is available covering individual player and team stats, league standings, game outcomes and live feed information highlighting key events during the progression of individual games. For simplicities sake in this project I restricted the scope of analysis to focus solely on team performance per game, variables include:
I undertook descriptive analysis of game features to better understand the data and to identify any trends which would be indicative of a teams success within a single game or over a period of successive games.
During the loading process of game data from the API a number of 404 errors were encountered indicating the specified game data was not available. We would expect to retrieve 1,230 games prior to the 2017 season and 1,371 2017 onwards (Las Vegas Golden Knights entered the league in 2017). The missing games were especially prevalent for games in the 2012-13 season. The 2018-19 season at the time of analysis is underway, results are retrieved regularly and appended to the historical game dataset.
Excluding current 2018-19 and 2012-13 seasons the distribution of game win percentage centres slightly above 50%. League leaders will rarely hold a game wining average of above 64% and lower end teams can generally be expected to win two out of five games (40%) averaged over a season.
Is there a significant difference for teams playing at their home arena compared to playing away?
There appears to be a very slight disadvantage to home teams, a home team in general would be expected to 51% of their away games and conversely 49% of their home games
Certain teams will perform better at home than others, illustrated below is the variance of home and away Win % from overall performance.
Lets take a look at goals scored for home and away teams
It appears that home teams score a higher number of goals per game more frequently than away teams, this would indicate a wider margin for home team wins than losses
By conducting a paired t-test on the distribution of home and away goals, we can establish if there is a statistically significant difference between the two populations
#conduct a dependent t-test on home and away goals
from scipy import stats
stats.ttest_rel(goals['Home'], goals['Away'])
Ttest_relResult(statistic=11.135112011475373, pvalue=1.2859551452133143e-28)
The pvalue of the t-test being less than 0.05 we can reject the null hypothesis that there is no difference home and away teams in respect of goals scored
Lets look at shots (on goal, missed and blocked), a popular metric is the Corsi which compares shots for and against and can be represented as a integer or percentage
For example Team A have 66 shots on goal Team B have 55:
Corsi Team A = 66-55 = 11
Corsi Team A% = 66/(66+55) = 55%
Corsi Team B = 55-66 = -11
Corsi Team B% = 55/(66+55) = 45%
If we compare teams Win% and Corsi% we can see a correlating trend between the two, by calculating the correlation coefficient we see a moderate correlation between the two:
from scipy.stats.stats import pearsonr
pearsonr(corsi_grouped['Corsi %'],corsi_grouped['Win %'])
(0.6184222739960428, 0.00020899389348708302)
(Coefficient, P value for testing of non-correlation)
Below is a regression plot of Corsi% vs Win%
Elo scores were originally devloped as a chess ranking system and have since been deployed in various competitive enviroments including videogames and sports.
The concept behind this rating process in simplified terms is:
In the chart below we see the progression of Elo ratings for various teams
Lets look at an example of Elo scores can be used to predict the outcome of a game:
Team A have a score of 1700 Team B have a score of 1450
a_score = 1700
b_score = 1450
prob_a_win = 1/(1+10**((b_score - a_score)/400))
prob_a_win = 0.808317673
prob_b_win = 1 - prob_a_win
prob_b_win = 0.191682327
A coefficient K is used to calculate the number of points to be deducted from the losing team, this is multiplied by the prior probability of a team winning, so in cases where a favourite wins a smaller number of points are deducted from the losing team.
Lets imagine with the above scenario the favourite Team A won and coefficient K is set to 10:
a_score = 1700
b_score = 1450
prob_a_win = 0.808317673
elo_change = k*(1-prob_a_win)
elo_change = 10*(1-0.808317673)
elo_change = 1.916823275
Conversely if Team B (the underdog) won:
a_score = 1700
b_score = 1450
prob_b_win = 0.191682327
elo_change = k*(1-prob_b_win)
elo_change = 10*(1-0.191682327)
elo_change = 8.083176725
The value of K can be chosen by comparing the RMSE on predictions for each value of K
An optimum K coefficient appears around a value of 6 or 7, we will apply this for prediciting upcoming games.
In the dashboard below I have applied the Elo model to upcoming games to provide a predicitve probability of wins and losses, I have alose included some of the KPI’s common to NHL statistical analysis
Over the course of the analysed period the Elo rating system has provided the greatest accuracy, while other models (SVC, Naive Bayes and Logistic regression) were somewhat close in predictive accuracy when assesed against a training test split, this accuracy was not reflected in live deployment.