Skip to content

Latest commit

 

History

History
265 lines (212 loc) · 7.19 KB

TimeSeries_Prophet.md

File metadata and controls

265 lines (212 loc) · 7.19 KB
Raw

TIME SERIES: Facebook's Prophet Library

alt text Eduardo Muñoz

Documentation

Forecasting at Scale by Sean Taylor and Benjamin Letham

Installation

https://facebook.github.io/prophet/docs/installation.html#python

pip3 install fbprophet

1. Reading our Dataset

  • The input to Prophet is always a dataframe with two columns: ds and y.
  • The ds (datestamp) column should be of a format expected by Pandas, ideally YYYY-MM-DD for a date or YYYY-MM-DD HH:MM:SS for a timestamp.
  • The y column must be numeric, and represents the measurement we wish to forecast.
df = pd.read_csv('../UPDATE-TSA-NOTEBOOKS/Data/BeerWineLiquor.csv')
date beer
0 1/1/1992 1509
1 2/1/1992 1541
2 3/1/1992 1597
3 4/1/1992 1675
4 5/1/1992 1822 
# Change the names
df.columns = ['ds','y']
# Make sure ds column is a pandas datetime object
df['ds'] = pd.to_datetime(df['ds'])
ds y
0 1992-01-01 00:00:00 1509
1 1992-02-01 00:00:00 1541
2 1992-03-01 00:00:00 1597
3 1992-04-01 00:00:00 1675
4 1992-05-01 00:00:00 1822
  • By default Prophet is going to expect daily data

2. Train Test Split

print(len(df)) # to know how many rows we have
train = df.iloc[:576]
test = df.iloc[576:]

3. Model

from fbprophet import Prophet
m = Prophet(seasonality_mode='additive') # 'multiplicative' Find out what is better for your model: RMSE
m.fit(train)

4. Predictions

# In Stead of doing df.index.freq = 'MS' as we do in Statsmodel
future = m.make_future_dataframe(periods=12,freq='MS')
preds = m.predict(future)
preds.head() # to find out more about the output
# yhat, yhat_lower and yhat_upper are probably the most important terms of this DF
preds[['ds','yhat_lower','yhat_upper','yhat']].tail(24)

4. Plotting the Results

ax = preds.plot(x='ds',y='yhat',label='Predictions',legend=True,figsize=(12,6))
test.plot(x='ds',y='y',label='Test Data',legend=True,ax=ax,xlim=('2018-01-01','2019-01-01'));
m.plot(preds);

# Trend and Seasonality
m.plot_components(preds);

5. Evaluate the Model

5.1 RMSE

from statsmodels.tools.eval_measures import rmse
# Alternative:
# from sklearn.metrics import mean_squared_error
RMSE = rmse(test['y'],preds['yhat'][-12:])
RMSE # it should be interesting to compare this value with test['y'].mean()

5.2 Prophet's Metrics

from fbprophet.diagnostics import cross_validation, performance_metrics
from fbprophet.plot import plot_cross_validation_metric

# The initial period should be long enough to capture all of the components of the model, 
# in particular seasonalities and extra regressors: at least a year for yearly seasonality,
# at least a week for weekly seasonality, etc.

# Initial training period
initial = 5 * 365
initial = str(initial) + ' days'
# Period lenght that we are going to perform the cross validation. 
# How many times to fold?
period = 5 * 365
period = str(period) + ' days'
# Horizon of prediction for essentially each fold. 
# How far out do you want to forecast for each period?
horizon = 365
horizon = str(horizon) + ' days'

df_cv = cross_validation(m,initial=initial,period=period,horizon=horizon)
performance_metrics(df_cv)

plot_cross_validation_metric(df_cv,metric='rmse');

6. Forecast

from fbprophet import Prophet
m = Prophet()
m.fit(df)
# In Stead of doing df.index.freq = 'MS' 
future = m.make_future_dataframe(periods=12,freq='MS')

forecast = m.predict(future)
m.plot(forecast);
m.plot_components(forecast);

Trend Line

# 1. Loading Libraries
import pandas as pd
import matplotlib.pyplot as plt
from fbprophet import Prophet
# 2. Reading the Dataset
df = pd.read_csv('../UPDATE-TSA-NOTEBOOKS/Data/HospitalityEmployees.csv')
df.columns = ['ds','y']
df['ds'] = pd.to_datetime(df['ds'])
# 3. Fitting the Model
m = Prophet()
m.fit(df)
future = m.make_future_dataframe(periods=12,freq='MS')
# 4. Predcitions or Forecasting
forecast = m.predict(future)
# 5. Main Changes in Trend Line
# It shows the major points where the trend line happened to change
from fbprophet.plot import add_changepoints_to_plot
fig = m.plot(forecast)
a = add_changepoints_to_plot(fig.gca(),m,forecast);

Changing Seasonality

What kind of model will give us better results? Add or Mult

Split Train Test and Watch RMSE for ADD or MULT if you are not sure about your Dataset

Additive Model

# 1. Loading Libraries
import pandas as pd
import matplotlib.pyplot as plt
from fbprophet import Prophet
# 2. Reading the Dataset
df = pd.read_csv('../UPDATE-TSA-NOTEBOOKS/Data/airline_passengers.csv')
df.columns = ['ds','y']
df['ds'] = pd.to_datetime(df['ds'])
# 3. Model
m = Prophet(seasonality_mode='additive')
m.fit(df)
future = m.make_future_dataframe(periods=50,freq='MS')
forecast = m.predict(future)
# 4. Plot
fig = m.plot(forecast);
fig = m.plot_components(forecast);
from fbprophet.plot import add_changepoints_to_plot
fig = m.plot(forecast)
a = add_changepoints_to_plot(fig.gca(),m,forecast);

Multiplicative Model

# 1. Loading Libraries
import pandas as pd
import matplotlib.pyplot as plt
from fbprophet import Prophet
# 2. Reading the Dataset
df = pd.read_csv('../UPDATE-TSA-NOTEBOOKS/Data/airline_passengers.csv')
df.columns = ['ds','y']
df['ds'] = pd.to_datetime(df['ds'])
# 3. Model
m = Prophet(seasonality_mode='multiplicative')
m.fit(df)
future = m.make_future_dataframe(periods=50,freq='MS')
forecast = m.predict(future)
# 4. Plot
fig = m.plot(forecast);
fig = m.plot_components(forecast);
from fbprophet.plot import add_changepoints_to_plot
fig = m.plot(forecast)
a = add_changepoints_to_plot(fig.gca(),m,forecast);