#!/usr/bin/env python
# coding: utf-8

# # Bitcoin analysis and forecast based on historical trade data

# ### Readme
# 
# This is a Bitcoin analysis and forecast made 04 March 2017 to get me started with FB Prophet. The outcome and prediction should not be taken too seriously. Of course, if the prediction and forecast is correct, I'll take all the honor and glory and you should hire me as a great data analyst. However, the analysis shows some interesting stuff nonetheless.
# 
# The trade data is from 2011-07-17 to 2018-03-04 and was modified in the way that a mean "y" column was added.
# 
# 
# #### Forecasts and graphs
# 
# * 1 year forecast based on all trade data
# * Seasonality and holiday effects based on all trade data and holidays
# * 1 year forecast based on all trade data without outlier 
# * 60 day forecast based on last year trade data
# * 60 day forecast based on last year trade data without anomaly 
# 
# 
# #### Insights
# 
# * Best day to buy seems to be Wednesday
# * Best day to sell seems to be Sunday
# * Best month to buy seems to be April/May
# * Best time to sell seems to be around new year
# * End of February respective beginning of March looks like a bull trap
# 
# 
# #### Further actions
# 
# We should check back latest 2019-03-04 to see if the forecast was correct and to make a new one ;)

# ### Resources
# #### Prophet
# https://research.fb.com/prophet-forecasting-at-scale/
# #### Bitcoin trade data CSV
# https://data.bitcoinity.org/markets/price/all/USD?c=e&t=l

# In[1]:


import warnings
import datetime
import copy
import pandas as pd
from fbprophet import Prophet

warnings.filterwarnings('ignore')

dateparse = lambda dates: [pd.datetime.strptime(d, '%Y-%m-%d %H:%M:%S %Z') for d in dates]

df = pd.read_csv('bitcoin_trade_data_combined.CSV', parse_dates=['ds'], date_parser=dateparse, usecols=['ds', 'y'])
df.head()


# We make a copy and extract the last 365 datapoints

# In[2]:


dc = copy.deepcopy(df)
df_oy = dc[-365:]
df_oy.head()


# We declare some special events that seems to bias the rate.

# In[3]:


chinese_new_year = pd.DataFrame({
  'holiday': 'cny',
  'ds': pd.to_datetime(['2018-02-16', '2017-01-28', '2016-02-08',
                        '2015-02-19', '2014-01-31', '2013-02-10',
                        '2012-01-23']),
  'lower_window': -20,
  'upper_window': 20,
})
xmas = pd.DataFrame({
  'holiday': 'xmas',
  'ds': pd.to_datetime(['2011-12-24', '2012-12-24', '2013-12-24',
                        '2014-12-24', '2015-12-24', '2016-12-24',
                        '2017-12-24']),
  'lower_window': -20,
  'upper_window': 20,
})
holidays = pd.concat((chinese_new_year, xmas))


# In[4]:


m = Prophet(daily_seasonality=False, holidays=holidays, holidays_prior_scale=0.05)
m.fit(df)
future = m.make_future_dataframe(periods=365)
future.tail()


# In[5]:


forecast = m.predict(future)


# In[6]:


forecast[['ds', 'yhat', 'yhat_lower', 'yhat_upper']].tail()


# In[7]:


m.plot(forecast);


# In[8]:


m.plot_components(forecast);


# Now let's assume that the end of 2017 run was an outlier. We remove the outlier and see how the forecast changes

# In[9]:


outlier_from = pd.Timestamp('2017-11-15')
outlier_to = pd.Timestamp('2018-01-15')
for x in range(1, len(df['ds'])):
    if (df['ds'][x] > outlier_from and df['ds'][x] < outlier_to):
        df['y'][x] = None
m_wo = Prophet(daily_seasonality=False, yearly_seasonality=False)
m_wo.fit(df)
future_wo = m_wo.make_future_dataframe(periods=365)
forecast_without_outliers = m_wo.predict(future_wo)
m_wo.plot(forecast_without_outliers);


# Now we want to make a 60 day forecast based on last years data to show how important perspective is.

# In[10]:


m_oy = Prophet(daily_seasonality=False, yearly_seasonality=False)
m_oy.fit(df_oy)
future_oy = m_oy.make_future_dataframe(periods=60)
forecast_oy = m_oy.predict(future_oy)
m_oy.plot(forecast_oy);


# What happens if we forecast without the anomaly? Let's have a look.

# In[11]:


for i, row in df_oy.iterrows():
    if (row['ds'] > outlier_from and row['ds'] < outlier_to):
        df_oy['y'][i] = None
m_wo = Prophet(daily_seasonality=False, yearly_seasonality=False)
m_wo.fit(df_oy)
future_wo = m_wo.make_future_dataframe(periods=60)
forecast_without_outliers = m_wo.predict(future_wo)
m_wo.plot(forecast_without_outliers);


# In[12]:


forecast_without_outliers[['ds', 'yhat', 'yhat_lower', 'yhat_upper']].tail()