frac-regression-mlb.ipynb

Kernel: bayesian

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import pymc3 as pm

%matplotlib inline

df = pd.read_csv('datasets/mlb_2013-2016.csv')
df.head()

df.plot(kind='scatter', x='R', y='Winning %')

import seaborn as sns

# df.plot(kind='scatter', x='Season', y='Team Salary')
sns.violinplot(x='Season', y='Team Salary', data=df)

df.columns

# Let's predict 'Winning %' from the rest of the columns.
cols = list(df.columns)
cols.remove('Season')
cols.remove('Team')
cols.remove('Team Salary (in millions)')
cols.remove('League')
cols.remove('Wins')
cols.remove('Losses')

banned = ['Winning %', 'Wins', 'Losses']
y_cols = ['Winning %']
feat_cols = list(set(cols) - set(y_cols) - set(banned))

from sklearn.ensemble import RandomForestRegressor

rfr = RandomForestRegressor(n_estimators=50)

rfr.fit(df[feat_cols], df[y_cols])
plt.plot(rfr.feature_importances_)

important_cols = []
threshold = 0.03
impt_idxs = [i for i, k in enumerate(rfr.feature_importances_) if k > threshold]
for i in impt_idxs:
    important_cols.append(feat_cols[i])
important_cols

import theano.tensor as tt
with pm.Model() as model:
    weights = pm.Normal('weights', mu=0, sd=100**2, shape=(len(important_cols),))
    perc_losses = tt.dot(weights, df[important_cols].T)
    # weights = pm.Normal('weights', mu=0, sd=100**2, shape=(2,))
    # perc_losses = weights[0] * df['Team Salary'] + weights[1] * df['R']
    alpha = 1
    beta = 1 / perc_losses
    sd = pm.HalfCauchy('sd', beta=100)
    # like = pm.Beta('likelihood', alpha=alpha, beta=beta, observed=df[y_cols])
    like = pm.Normal('likelihood', mu=perc_losses, sd=sd, observed=df[y_cols])
    # like = pm.Beta('likelihood', mu=perc_losses, sd=sd, observed=df[y_cols])
    # like = pm.Bernoulli('likelihood', p=perc_losses, observed=df[y_cols])

with model:
    trace = pm.sample(50000, step=pm.Metropolis())

pm.traceplot(trace)

burnin = 30000
pm.traceplot(trace[burnin:])