-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathaggregators.py
192 lines (176 loc) · 9.23 KB
/
aggregators.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import itertools
from collections import defaultdict
from functools import partial
import numpy as np
import scipy
from statsmodels.tsa.api import Holt, SimpleExpSmoothing
TOURNEY_START_DAY = 134
def _rpi(season_results, team, weights):
results = season_results[team]['results']
win_percent = sum(results) / len(results) if results else 0
opponents_win_percent = _opponents_win_percent(season_results, [o for o in season_results[team]['opponents'] if o != team])
opponents_opponents_win_percent = _opponents_opponents_win_percent(season_results, season_results[team]['opponents'])
w1, w2, w3 = weights[0], weights[1], weights[2]
return w1 * win_percent + w2 * opponents_win_percent + w3 * opponents_opponents_win_percent
def _opponents_win_percent(season_results, opponents):
win_percents = []
for opponent in opponents:
results = season_results[opponent]['results']
win_percent = sum(results) / len(results) if results else 0
win_percents.append(win_percent)
return sum(win_percents) / len(opponents) if opponents else 0
def _opponents_opponents_win_percent(season_results, opponents):
win_percents = []
for opponent in opponents:
win_percent = _opponents_win_percent(season_results, season_results[opponent]['opponents'])
win_percents.append(win_percent)
return sum(win_percents) / len(opponents) if opponents else 0
def modified_rpi(X, start_day, weights=(.15, .15, .7)):
stats = defaultdict(partial(defaultdict, partial(defaultdict, list)))
rpis = []
for row in X.itertuples(index=False):
if row.Daynum >= start_day:
wrpi = _rpi(stats[row.Season], row.Wteam, weights)
lrpi = _rpi(stats[row.Season], row.Lteam, weights)
if row.Wteam < row.Lteam:
rpis.append(wrpi - lrpi)
else:
rpis.append(lrpi - wrpi)
if row.Daynum < TOURNEY_START_DAY:
stats[row.Season][row.Wteam]['opponents'].append(row.Lteam)
stats[row.Season][row.Lteam]['opponents'].append(row.Wteam)
stats[row.Season][row.Wteam]['results'].append(1)
stats[row.Season][row.Lteam]['results'].append(0)
return np.array(rpis)
#TODO remove unimportant
def descriptive_stats(results, frequency_domain=False):
described = []
for stat in results.keys():
x = np.abs(np.fft.fft(results[stat])) if frequency_domain else results[stat]
described.append(min(x))
described.append(max(x))
described.append(np.percentile(x, 75) - np.percentile(x, 25)) # IQR
described.append(np.median(x))
described.append(np.mean(x))
described.append(np.var(x)) # second moment
described.append(scipy.stats.skew(x)) #third moment
described.append(scipy.stats.kurtosis(x)) # fourth moment
return described
#TODO remove unimportant
def time_series_stats(results, frequency_domain=False):
timed = []
for stat in results.keys():
x = np.abs(np.fft.fft(results[stat])) if frequency_domain else results[stat]
timed.append(x[-1])
timed.append(np.mean(x[-3:])) # simple 3 game moving average
timed.append(np.mean(x[-5:])) # simple 5 game moving average
m = np.mean(x)
timed.append(len(list(itertools.takewhile(lambda xi, mean=m: xi > mean, reversed(x))))) # monotonicity
timed.append(len(list(itertools.takewhile(lambda xi, mean=m: xi < mean, reversed(x)))))
if len(x) > 1:
timed.append(SimpleExpSmoothing(x).fit(smoothing_level=.3, optimized=False).fittedvalues[-1]) # exponential smoothing
timed.append(Holt(x).fit(smoothing_level=.5, smoothing_slope=.5, optimized=False).fittedvalues[-1]) # Holt's linear trend
else:
timed.append(x[0])
timed.append(x[0])
return timed
def statistics(X, start_day, stat_F, frequency_domain=False):
stats = defaultdict(partial(defaultdict, partial(defaultdict, list)))
compiled_stats = []
for row in X.itertuples(index=False):
if row.Daynum >= start_day:
wstats = stat_F(stats[row.Season][row.Wteam], frequency_domain=frequency_domain)
lstats = stat_F(stats[row.Season][row.Lteam], frequency_domain=frequency_domain)
if row.Wteam < row.Lteam:
compiled_stats.append(np.subtract(wstats, lstats))
else:
compiled_stats.append(np.subtract(lstats, wstats))
if row.Daynum < TOURNEY_START_DAY:
wposs = row.Wfga - row.Wor + row.Wto + .475 * row.Wfta #TODO try multiple poss proxies like rpi weights
lposs = row.Lfga - row.Lor + row.Lto + .475 * row.Lfta
# interaction features based on commonly used metrics
stats[row.Season][row.Wteam]['eff_field_goal_percent'].append((row.Wfgm + .5 * row.Wfgm3) / row.Wfga)
stats[row.Season][row.Lteam]['eff_field_goal_percent'].append((row.Lfgm + .5 * row.Lfgm3) / row.Lfga)
stats[row.Season][row.Wteam]['true_shooting'].append(.5 * row.Wscore / (row.Wfga + 0.475 * row.Wfta))
stats[row.Season][row.Lteam]['true_shooting'].append(.5 * row.Lscore / (row.Lfga + 0.475 * row.Lfta))
stats[row.Season][row.Wteam]['rebound_rate'].append(row.Wor / (row.Wor + row.Ldr))
stats[row.Season][row.Lteam]['rebound_rate'].append(row.Lor / (row.Lor + row.Wdr))
stats[row.Season][row.Wteam]['free_throw_rate'].append((row.Wfta / row.Wftm) if row.Wftm > 0 else 0)
stats[row.Season][row.Lteam]['free_throw_rate'].append((row.Lfta / row.Lftm) if row.Lftm > 0 else 0)
stats[row.Season][row.Wteam]['turnover_rate'].append(row.Wto / wposs)
stats[row.Season][row.Lteam]['turnover_rate'].append(row.Lto / lposs)
stats[row.Season][row.Wteam]['assist_rate'].append(row.Wast / row.Wfgm)
stats[row.Season][row.Lteam]['assist_rate'].append(row.Last / row.Lfgm)
stats[row.Season][row.Wteam]['block_rate'].append(row.Wblk / row.Lfga)
stats[row.Season][row.Lteam]['block_rate'].append(row.Lblk / row.Wfga)
stats[row.Season][row.Wteam]['steal_rate'].append(row.Wstl / lposs)
stats[row.Season][row.Lteam]['steal_rate'].append(row.Lstl / wposs)
stats[row.Season][row.Wteam]['score_rate'].append(row.Wscore / wposs)
stats[row.Season][row.Lteam]['score_rate'].append(row.Lscore / lposs)
stats[row.Season][row.Wteam]['foul_rate'].append(row.Wpf / lposs)
stats[row.Season][row.Lteam]['foul_rate'].append(row.Lpf / wposs)
return np.array(compiled_stats)
def _all_teams(data):
seasons = data['Season'].unique()
teams = {}
for season in seasons:
season_games = data.query('Season == %s' % season)
sorted_teams = sorted((season_games['Wteam'].append(season_games['Lteam'])).unique())
teams[season] = {X:i for i, X in enumerate(sorted_teams)}
return teams
def custom_ratings(X, start_day, rating_F):
teams = _all_teams(X)
seasons = np.unique(X['Season'])
stat_categories = {X:i for i, X in enumerate(['points'])}
stats = {season: np.zeros((len(teams[season]), len(teams[season]), len(stat_categories))) for season in seasons}
ratings = []
day = start_day - 1
adj_stats = None
for row in X.itertuples(index=False):
if row.Daynum < TOURNEY_START_DAY:
if row.Daynum > day:
adj_stats = rating_F(stats)
day = row.Daynum
wteam_id = teams[row.Season][row.Wteam]
lteam_id = teams[row.Season][row.Lteam]
stats[row.Season][wteam_id][lteam_id][stat_categories['points']] += row.Wscore
stats[row.Season][lteam_id][wteam_id][stat_categories['points']] += row.Lscore
if row.Daynum >= start_day:
team_ids = teams[row.Season]
wrating = adj_stats[row.Season][team_ids[row.Wteam]]
lrating = adj_stats[row.Season][team_ids[row.Lteam]]
if row.Wteam < row.Lteam:
ratings.append(np.subtract(wrating, lrating))
else:
ratings.append(np.subtract(lrating, wrating))
return np.array(ratings)
# https://www.kaggle.com/lpkirwin/fivethirtyeight-elo-ratings
def _elo_pred(elo1, elo2):
return 1 / (10 ** (-(elo1 - elo2) / 400) + 1)
def _elo_expected_margin(elo_diff):
return 7.5 + 0.006 * elo_diff
def _elo_update(welo, lelo, mov):
elo_diff = welo - lelo
pred = _elo_pred(welo, lelo)
mult = ((mov + 3) ** 0.8) / _elo_expected_margin(elo_diff)
update = 20 * mult * (1 - pred)
return update
def elo(X, start_day):
stats = defaultdict(partial(defaultdict, lambda: 1500))
elos = []
for row in X.itertuples(index=False):
welo = stats[row.Season][row.Wteam]
lelo = stats[row.Season][row.Lteam]
if row.Daynum >= start_day:
if row.Wteam < row.Lteam:
elos.append(welo - lelo)
else:
elos.append(lelo - welo)
if row.Daynum < TOURNEY_START_DAY:
mov = row.Wscore - row.Lscore
wadvantage = 100 if row.Wloc == 'H' else 0
ladvantage = 100 if row.Wloc == 'A' else 0
update = _elo_update(welo + wadvantage, lelo + ladvantage, mov)
stats[row.Season][row.Wteam] += update
stats[row.Season][row.Lteam] -= update
return elos