StevenReitsma

train = pd.read_csv('train.csv', parse_dates=['Date'])
train.columns = train.columns.str.lower()
features = pd.read_csv('features.csv', parse_dates=['Date'])
features.columns = features.columns.str.lower()
features[['markdown1', 'markdown2', 'markdown3','markdown4', 'markdown5']] = features[['markdown1', 'markdown2', 'markdown3','markdown4', 'markdown5']].fillna(0)
stores = pd.read_csv('stores.csv')
stores.columns = stores.columns.str.lower()
df_base = (
    train
    .merge(features, how='inner')

StevenReitsma

class WeekNumberCreator(object):
    def __init__(self, datecol):
        self.datecol = datecol
    def transform(self, X):
        return X.assign(weeknumber=lambda df: df.set_index(self.datecol).index.week)

StevenReitsma

class SmoothedVarCreator(object):
    def __init__(self, var, groupby_keys, alpha_list):
        self.var = var
        self.groupby_keys = groupby_keys
        self.alpha_list = alpha_list
    def transform(self, X):
        for alpha in self.alpha_list:
        func = {
            '%s_sm%s' % (self.var, int(alpha * 10)): lambda df:
            X.groupby(self.groupby_keys)[self.var].apply(lambda x: x.ewm(alpha=alpha, min_periods=0).mean())

StevenReitsma

class LagCreator(object):
    def __init__(self, groupby_keys, lagdict, drop_cols=True, suffix=''):
        self.groupby_keys = groupby_keys
        self.lagdict = lagdict
        self.drop_cols = drop_cols
        self.suffix = suffix
    def transform(self, X):
        cols_to_drop = list(set([item for sublist in 
                            self.lagdict.values() for item in sublist]))
        for week in self.lagdict:

StevenReitsma

from bdranalytics.pandaspipeline.transformers import PdFeatureChain
pipeline_pd = PdFeatureChain([
    ('wna', WeekNumberAdder(datecol='date')),
    ('svc', SmoothedVarCreator(var='weekly_sales', groupby_keys=['store', 'dept'], alpha_list=[0.1, 0.3])),
    ('lc1', LagCreator(groupby_keys=['store', 'dept', 'weeknumber'], lagdict=lagdict_yearly, drop_cols=False, suffix='ly')),
    ('lc2', LagCreator(groupby_keys=['store', 'dept'], lagdict=lagdict))
])
df_features = pipeline_pd.transform(df_base).dropna()

StevenReitsma

from bdranalytics.model_selection.growingwindow import IntervalGrowingWindow
igw = IntervalGrowingWindow(test_start_date='2011-03-01', timestamps='index', test_size='7 days')
cv = igw.split(df_features.set_index('date'))
multiple_ts_split = list(cv)

StevenReitsma

def force_float_df(df):
    def force_float_col(x):
        try:
            return x.astype(float)
        except:
            return pd.factorize(x)[0]
    return df.apply(force_float_col)
forward_sale_cols = ['weekly_sales1', 'weekly_sales2']
drop_cols = ['date']
all_drop_cols = forward_sale_cols + drop_cols

StevenReitsma

Index(['store', 'dept', 'weeknumber', 'weekly_sales_2', 'weekly_sales_1',
       'weekly_sales_sm1_1', 'weekly_sales_sm3_1', 'cpi0', 'fuel_price0',
       'isholiday0', 'markdown10', 'markdown20', 'markdown30', 'markdown40',
       'markdown50', 'temperature0', 'unemployment0', 'weekly_sales0',
       'isholiday1', 'temperature1', 'weekly_sales_1ly1', 'isholiday2',
       'temperature2', 'weekly_sales_1ly2'],
      dtype='object')

StevenReitsma

def neg_sq_loss(labels, preds):
    a = -0.1
    err = preds - labels
    grad = 2 * err * (np.sign(err) + a)**2
    hess = 2 * (np.sign(err) + a)**2
    return grad, hess

StevenReitsma

from sklearn.multioutput import MultiOutputRegressor
import lightgbm as lgb
gbm = lgb.LGBMRegressor(
    objective=neg_sq_loss,
    n_jobs=-1,
    num_leaves=51,
    learning_rate=0.02,
    n_estimators=500,
    categorical_feature=[0, 1, 2]
)