StephenCurry Shot Prediction

11 Feb 2019

Reading time ~1 minute

shots distribution

CLICK HERE: Notebook Version
CLICK HERE: Medium Blog for Coding

OVERVIEW

For this project, I used Beautifulsoup package to extract data from a NBA stats website regarding Stephen Curry’s shot points in Season 2014-2016. With the data, I built a neural network with Adam optimizer from scratch without any frameworks or packages, to predict whether a shot made by Stephen at a specific point will succeed or not.

DETAILS

The website I crawled hasn’t updated anymore, thus I only got the stats for season 2014-2015 and 2015-2016. The pachage beautifulsoup is quite easy to use:

extracting the data into JSON format

url = 'http://buckets.peterbeshai.com/api/?player=201939&season=2015'
data = requests.get(url, headers={
            "User-Agent": "Mozilla/5.0 (Windows NT 6.1; WOW64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/50.0.2661.102 Safari/537.36"})
cor =[]    
made = []

for shot in response.json():
    cor.append((shot['LOC_X'],shot['LOC_Y']))
    made.append(shot['EVENT_TYPE'])

The variable cor saved the coordinates of the positions where Stephen made shots and made is a binary class indicates whether he made it or not. I visualized the data into the graph shown at the top and the code is as follows:

plot the data on top of a court image

court_map = plt.imread('shotchart.png')
BB = (-300, 300, -50, 400)
def plot_data(df,BB,alpha=0.6,maps=court_map,s=10):
    fig, ax = plt.subplots(figsize=(15,10))
    #ax.scatter(df['pickup_longitude'],df['pickup_latitude'], alpha=alpha,c='r',s=s,zorder=1,marker='.')
    ax.scatter(df[0,np.where(y==1)], df[1,np.where(y==1)],zorder=2,alpha=alpha,c = 'b', s=2, cmap=plt.cm.Spectral, label='Made Shot')
    ax.scatter(df[0,np.where(y==0)], df[1,np.where(y==0)],zorder=2,alpha=alpha,c = 'r', s=2, cmap=plt.cm.Spectral, label='Miss Shot')
    ax.legend(loc=1)
    #ax.set_xlim(BB[1],BB[0])
    #ax.set_ylim(BB[3],BB[2]) 
    ax.xaxis.set_label_text('longitude')
    ax.yaxis.set_label_text('latitude')
    ax.set_title('Shooting Points')
    ax.imshow(maps, zorder=0,extent=BB)
plot_data(x,BB)

construct ANN architecture from scratch

Then, I created the ANN from scratch: built single forward propagation and backward propagation block, with repect to different activation functions (relu, sigmoid, tanh); define parameters initialization and parameters update with Adam functions; define cost function, etc. I wrote a blog about the whole coding process, please kindly find it here.

The model works well and the loss curve is placed below: