add alternative way of ovr computation

analysis/player-team-ovr-basketball/README.md

@@ -0,0 +1,4 @@
+avg.csv: 40+ seasons of exported season average stats.
+game.csv: 40+ seasons of exported individual games stats.
+
+This is an alternative rating approach which looks at team-level results to predict rating value

analysis/player-team-ovr-basketball/process.py

@@ -0,0 +1,133 @@
+import pandas as pd  
+import numpy as np  
+import matplotlib.pyplot as plt  
+from sklearn.linear_model import LinearRegression
+from collections import defaultdict
+
+y_map = { 'hgt': 'hgt',
+   'stre': 'str',
+   'spd': 'spd',
+   'jmp': 'jmp',
+   'endu': 'end',
+   'ins': 'ins',
+   'dnk': 'dnk',
+   'ft': 'ft.1',
+   'fg': '2pt',
+   'tp': '3pt',
+   'diq': 'diq',
+   'oiq': 'oiq',
+   'drb': 'drb',
+   'pss': 'pss',
+   'reb': 'reb' }
+
+
+game = pd.read_csv('game.csv')
+avg = pd.read_csv('avg.csv')
+
+# save homecourt
+hm_crt2 = defaultdict(list)
+for g in game.itertuples():
+    hm_crt2[g[1]].append(g[5])
+hm_crt = {k:[v[0],v[-1]] for k,v in hm_crt2.items()}
+
+# add keys for merge
+avg['fakeKey'] = avg['pid'].astype(str) + '_' + avg['Season'].astype(str)
+game['fakeKey'] = game['pid'].astype(str) + '_' + game['Season'].astype(str)
+
+# merge
+game_rate = game.merge(avg,on='fakeKey')
+
+# make minute-averaged ratings
+team_rating = defaultdict(lambda:np.zeros(17))
+team_score = {}
+team_minutes =defaultdict(float)
+
+real_gids = defaultdict(set)
+
+for row in game_rate.itertuples():
+    key = (row[1],row[5])
+    if row[5] == row[6]:
+        continue
+    MP = row[11]
+    rt = MP *np.array(row[92:])
+    ms,os = [int(_) for _ in row[7].split('-')]
+
+    team_rating[key] += rt
+    team_minutes[key] += MP
+    team_score[key] = ms-os
+
+    real_gids[key[0]].add(key[1])
+
+team_rating_n = {k: team_rating[k]/team_minutes[k] for k in team_rating}
+team_rating_ovr = {k: v[0] for k,v in team_rating_n.items()}
+
+game_res = []
+gt = []
+for gid,teams in hm_crt.items():
+    t = list(teams)
+    game_res.append( list(team_rating_n[(gid,t[0])] - team_rating_n[(gid,t[1])]) + [team_score[(gid,t[0])]] )
+    gt.append(gid)
+
+ratings_regression = list(game_rate.columns[91:])
+diff_df = pd.DataFrame(np.array(game_res),columns=ratings_regression + ['MOV'])
+
+# CRAP! normalize doesn't actually do zscore, so some of the stuff below is wrong! Might not matter much
+reg = LinearRegression()
+rating_vals = list(diff_df.drop(['MOV','Ovr','Pot'],axis=1).columns)
+reg.fit(diff_df[rating_vals], diff_df['MOV'])
+# print('Intercept: \n', reg.intercept_)
+# print('Coefficients: \n', reg.coef_)
+
+# Adjust old ovrs for the ratings we're skipping
+# Recompute Ovr because we want the unscaled version, so scaling can be applied on top in JS
+
+# Adjust old ovrs for the ratings we're skipping
+# Recompute Ovr because we want the unscaled version, so scaling can be applied on top in JS
+avg['OvrOld'] = (5 * avg['Hgt'] + 1 * avg['Str'] + 4 * avg['Spd'] + 2 * avg['Jmp'] + 1 * avg['End'] + 1 * avg['Ins'] + 2 * avg['Dnk'] + 1 * avg['FT.1'] + 1 * avg['2Pt'] + 3 * avg['3Pt'] + 7 * avg['oIQ'] + 3 * avg['dIQ'] + 3 * avg['Drb'] + 3 * avg['Pss'] + 1 * avg['Reb']) / 38
+
+# Scale to match old ovr
+mean_old = avg.OvrOld.mean()
+std_old = avg.OvrOld.std()
+
+ovr_new_unscaled = reg.predict(avg[rating_vals])
+mean_new = ovr_new_unscaled.mean()
+std_new = ovr_new_unscaled.std()
+
+factor_mult = std_old / std_new
+factor_add = mean_old
+# print('factor_mult: \n', factor_mult)
+# print('factor_add: \n', factor_add)
+
+avg['OvrNew'] = (ovr_new_unscaled-mean_new) * factor_mult + factor_add
+# print(dataset.Ovr)
+# print(dataset.OvrNew)
+
+def formatThree(num):
+    return str(np.format_float_positional(num, precision=3, unique=False, fractional=False, trim='k'))
+
+print(avg[['OvrOld', 'OvrNew']])
+
+# Output
+print('(')
+ratings = [_.lower() for _ in rating_vals]
+ratings2 = ['hgt', 'stre', 'spd', 'jmp', 'endu', 'ins', 'dnk', 'ft', 'tp', 'oiq', 'diq', 'drb', 'pss', 'fg', 'reb'];
+for i in range(len(ratings2)):
+    if i == len(ratings2) - 1:
+        end_part = ''
+    else:
+        end_part = ' +'
+    idx = ratings.index(y_map[ratings2[i]])
+    print('    ' + formatThree(factor_mult * reg.coef_[idx]) + ' * ratings.' + ratings2[i] + end_part)
+print(') + ' + formatThree(factor_add));
+
+
+# Plot
+avg.plot.hexbin(x='OvrOld', y='OvrNew', gridsize=20)
+plt.xlim(0, 100)
+plt.ylim(0, 100)
+plt.xlabel('Old Ovr')  
+plt.ylabel('New Ovr')  
+
+plt.plot([0, 100], [0, 100])
+
+plt.show()