import pandas as pd
import numpy as np
import seaborn as sns
from scipy import stats
import matplotlib as mpl
import matplotlib.pyplot as plt

import scipy as stats

from fbprophet import Prophet

data = pd.read_csv('Sales_by_day.csv', parse_dates=['ds'])
txns = pd.DataFrame(data[['ds','# Txns']])
txns['y'] = np.log10(txns['# Txns'])
del txns['# Txns']
txns

#forecasting tool used to fit model & plot data
m = Prophet()
m.fit(txns)

#create future dataframe with number of periods to be forecasted
future = m.make_future_dataframe(periods=336)
future

forecast = m.predict(future)
forecast[['ds', 'yhat', 'yhat_lower', 'yhat_upper']].tail()

%matplotlib inline
m.plot(forecast);

m.plot_components(forecast);

forecast

#Get R-squared to measure model variation
stats.corrcoef(x=txns['y'], y=forecast.loc[:724,'yhat'])

events_15 = pd.read_csv('2015_events.csv', parse_dates=['ds'])
events_16 = pd.read_csv('2016_events.csv', parse_dates=['ds'])
events_17 = pd.read_csv('2017_events.csv', parse_dates=['ds'])
events_future = pd.read_csv('2017_events_future.csv', parse_dates=['ds'])
events_future2 = pd.read_csv('2017_events_future2.csv', parse_dates=['ds'])


past_events = events_15.append(events_16)
all_past_events = past_events.append(events_17)
all_events = all_past_events.append(events_future)
all_events2 = all_past_events.append(events_future2)
all_events2

#read in dates with Scratcher Event & Holidays listed
dates = pd.read_csv('changepoint_holiday.csv',parse_dates=['Scratcher Holiday','Ad Changepoint','Year Changepoint','Holidays'])
#set up different event/holiday dataframes, their 'impact' windows and combine sets
scratch_hol = pd.DataFrame({
    'holiday':'holiday',
    'ds': pd.to_datetime(dates['Scratcher Holiday'].dropna()),
    'lower_window':0,
    'upper_window':7,
})
reg_hol = pd.DataFrame({
    'holiday':'holiday',
    'ds': pd.to_datetime(dates['Holidays'].dropna()),
    'lower_window':-2,
    'upper_window':2,
})
new_dates = pd.DataFrame({
    'ds':pd.to_datetime('2017-11-23'), 
    'holiday':'holiday', 
    'lower_window':-2, 
    'upper_window':2,},
    index=[21])
reg_hol = reg_hol.append(new_dates)

holidays = scratch_hol.append(reg_hol)
holidays.sort( ascending=False)

#columns = {'ds', 'holiday', 'lower_window', 'upper_window'}
#rows = '2017-08-11', '2017-10-19', '2017-11-10', '2017-12-01','2017-12-15'

future_events = pd.read_csv('future_events.csv', parse_dates=['ds'])
future_events
#new_holidays = holidays.append(future_events)
#new_holidays

future_events = pd.read_csv('future_events.csv', parse_dates=['ds'])
future_events
new_holidays = holidays.append(future_events)
new_holidays

m1 = Prophet(holidays=all_events2, holidays_prior_scale=.25)
forecast3 = m1.fit(txns).predict(future)
m1.plot(forecast3);

m1.plot_components(forecast3);

forecast3.to_csv('daily_fcst_all_events2.csv')

forecast1.to_csv('fcst_daily_txns.csv')
txns.to_csv('data_daily_txns.csv')

#Get R-squared to measure model variation
stats.corrcoef(x=txns['y'], y=forecast1.loc[:724,'yhat'])

forecast.to_csv('prophet_txns_fcst.csv')
forecast1.to_csv('prophet_txns_fcst1.csv')

#m1 = Prophet(holidays=holidays, holidays_prior_scale=.25)
forecast1_sales = m1.fit(sales).predict(future)
m1.plot(forecast1_sales);

m1.plot_components(forecast1_sales);

week_data = pd.read_csv('week_data.csv', parse_dates=['ds'])
week_data.describe()

week_bskt = week_data[['ds','# Txns','Sales $']]
week_bskt['y'] = week_bskt['Sales $']/week_bskt['# Txns']
del week_bskt['Sales $']
del week_bskt['# Txns']
week_bskt

week_bskt

week_txns = week_data[['ds','# Txns']]
week_txns['y'] = np.log10(week_txns['# Txns'])
del week_txns['# Txns']
week_txns

week_txns.to_csv('week_txns.csv')

m_week = Prophet(holidays=new_holidays, changepoints=['2016-01-30','2017-02-04'],changepoint_prior_scale=.25, holidays_prior_scale=.25)

m_week.fit(week_txns)

future_week = m_week.make_future_dataframe(periods=44, freq='W-SAT')

fcst_week = m_week.predict(future_week)
m_week.plot(fcst_week);
#m_week.plot_components(fcst_week);

fcst_week.to_csv('fcst_week.csv')

fcst_week

#Get R-squared to measure model variation
stats.corrcoef(x=week_txns['y'], y=fcst_week.loc[:111,'yhat'])

week_sales = week_data[['ds','Sales $']]
week_sales['y'] = np.log10(week_sales['Sales $'])
del week_sales['Sales $']
week_sales

m_bskt = Prophet(holidays=all_events)
m_bskt.fit(week_bskt)
fcst_week_bskt = m_bskt.predict(future_week)
m_bskt.plot(fcst_week_bskt);

fcst_week_bskt.to_csv('fcst_week_bskt_all_events.csv')

m_week.plot_components(fcst_week_sales);
#m_week.yearly_seasonality(fcst_week_sales)

m_week.plot_components(fcst_week_sales);

fcst_week
fcst_week.to_csv('fcst_week_sales.csv')
week_sales.to_csv('data_week_sales.csv')

#Get R-squared to measure model variation
stats.corrcoef(x=week_txns['y'], y=fcst_week.loc[:111,'yhat'])

m_week.fit(week_bskt)
fcst_week_bskt = m_week.predict(future_week)
m_week.plot(fcst_week_bskt);

fcst_week_bskt.to_csv('fcst_week_bskt.csv')

mg1_data = pd.read_csv('mg1_week_data.csv', parse_dates=['ds'])
mg2_data = pd.read_csv('mg2_week_data.csv', parse_dates=['ds'])
mg3_data = pd.read_csv('mg3_week_data.csv', parse_dates=['ds'])
#mg1_data.info()
#mg2_data.describe()
mg3_data.describe()

#mg1 = pd.DataFrame(mg1_data[['ds','Txns']])
#mg1['y'] = np.log10(mg1['Txns'])
#del mg1['Txns']
#mg1.drop(mg1.index[-1],inplace= True)
mg1

#mg2 = pd.DataFrame(mg2_data[['ds','Txns']])
#mg2['y'] = np.log10(mg2['Txns'])
#del mg2['Txns']
#mg2.drop(mg2.index[-1],inplace= True)
mg2

#mg3 = pd.DataFrame(mg3_data[['ds','Txns']])
#mg3['y'] = np.log10(mg3['Txns'])
#del mg3['Txns']
#mg3.drop(mg3.index[-1],inplace= True)
mg3

m_week.fit(mg1)
fcst_mg1_txns = m_week.predict(future_week)
fcst_mg1_txns

m_week.plot(fcst_mg1_txns);

m_week.plot_components(fcst_mg1_txns);

m_week.fit(mg2)
fcst_mg2_txns = m_week.predict(future_week)
fcst_mg2_txns

m_week.plot(fcst_mg2_txns);

m_week.plot_components(fcst_mg2_txns);

m_week.fit(mg3)
fcst_mg3_txns = m_week.predict(future_week)
fcst_mg3_txns

m_week.plot(fcst_mg3_txns);

m_week.plot_components(fcst_mg3_txns);