Forecasting Time Series Data with FbProphet in Python

    FbProphet, an open source software released by Facebook, provides a procedure for forecasting time series data based on an additive model. In this tutorial, I'll briefly explain how to forecast time series data by using FbProphet API in Python.

    The tutorial covers:

1. Preparing time series data
2. Defining the model and forecasting
   
3. Performance evaluation
4. Source code listing
   

    Let's start by loading the required packages for this tutorial.

import yfinance as yf
import pandas as pd

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

 

You can install 'yfinance' and 'fbprophet' packages by using pip. 

 

pip install yfinance 

pip install fbprophet 

 

 

Preparing time series data

 

    Time series data is a sequence of data points indexed in time order. We can get this type of data by using Yahoo finance API. In this tutorial, we'll use Facebook historical stock price data to analyze. Below code shows how to download and prepare it for use.

 

tickerSymbol = "FB"
tickerData = yf.Ticker(tickerSymbol)
tickerDf = tickerData.history(start="2015-01-01", end="2020-01-07")

tickerDf.head() 

               Open       High        Low  ...    Volume  Dividends  Stock Splits
Date                                         ...                                   
2014-12-31  79.540001  79.800003  77.860001  ...  19935400          0             0
2015-01-02  78.580002  78.930000  77.699997  ...  18177500          0             0
2015-01-05  77.980003  79.250000  76.860001  ...  26452200          0             0
2015-01-06  77.230003  77.589996  75.360001  ...  27399300          0             0
2015-01-07  76.760002  77.360001  75.820000  ...  22045300          0             0 

 

    The  prophet model requires data frame input data with columns 'ds' and 'y'. 'ds' is time stamp data and 'y' numeric data. We'll create input data as shown below.

 

df = pd.DataFrame(columns=["ds", "y"])

df["ds"] = tickerDf.index
tickerDf = tickerDf.reset_index()
df["y"] =  tickerDf['Close'] 

df.head()

 ds          y
0 2014-12-31  78.019997
1 2015-01-02  78.449997
2 2015-01-05  77.190002
3 2015-01-06  76.150002
4 2015-01-07  76.150002 

 

 

Defining the model and forecasting

 

    We define model by using Prophet class and fit the model on 'df' input data.   

 

model = Prophet()
model.fit(df)
 

    Prediction days can be set by using make_future_dataframe() method of the model.

future = model.make_future_dataframe(periods=180)

    Now, we can predict the future period with trained model.

forecast = model.predict(future)
forecast.head()

 ds      trend  ...  multiplicative_terms_upper       yhat
0 2014-12-31  77.918570  ...                         0.0  70.933857
1 2015-01-02  77.967258  ...                         0.0  70.955889
2 2015-01-05  78.040291  ...                         0.0  71.149336
3 2015-01-06  78.064635  ...                         0.0  71.453739
4 2015-01-07  78.088979  ...                         0.0  71.712704  

   FbProphet provides 'plot' method to visualize forecast data. We can easily plot the forecast data.

plot.plot(model, forecast)
 

 

 

    To visualize components we use plot_components method.  

 

plot.plot_components(model, forecast) 

 

Performance Evaluation

 

     We can check prediction performance by using cross_validation() method. Defined cutoffs in a historic data will be used to measure prediction error rates. We apply cross validation and extract the statistics of the prediction.

 

df_cv = cross_validation(model, initial='360 days', period='180 days', horizon = '365 days')
df_p = performance_metrics(df_cv, rolling_window=0.1) 

df_p.head() 

horizon         mse       rmse        mae      mape     mdape  coverage
0 36 days  168.357384  12.975260   9.822421  0.064845  0.054113  0.301136
1 37 days  179.032791  13.380314  10.178312  0.067101  0.056766  0.289773
2 38 days  188.598973  13.733134  10.482477  0.069057  0.057384  0.272727
3 39 days  193.924787  13.925688  10.671479  0.070444  0.058132  0.264773
4 40 days  199.392740  14.120649  10.897853  0.071910  0.059398  0.250000

    Finally, we'll visuaze Mean Absolute Percent Error (MAPE) metrics by using plot_cross_validation_metric() method. 

 

plot.plot_cross_validation_metric(df_cv, metric='mape')

 

 

    In this tutorial, we've briefly learned how to forecast time series data by using FbProphet API in Python.  The full source code is listed below. 

 

 

Source code listing

import yfinance as yf
import pandas as pd

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

tickerSymbol = "FB"
tickerData = yf.Ticker(tickerSymbol)
tickerDf = tickerData.history(start="2015-01-01", end="2020-01-07")
tickerDf.head()

df = pd.DataFrame(columns=["ds", "y"])

df["ds"]  = tickerDf.index
tickerDf = tickerDf.reset_index()
df["y"] =  tickerDf['Close']

df.head()

model = Prophet()
model.fit(df)

future = model.make_future_dataframe(periods=180)
future.head()

forecast = model.predict(future)
forecast.head()

plot.plot(model, forecast)
plot.plot_components(model, forecast)


df_cv = cross_validation(model, initial='360 days', period='180 days', horizon = '365 days')
df_p = performance_metrics(df_cv, rolling_window=0.1)
df_p.head() 

plot.plot_cross_validation_metric(df_cv, metric='mape')

 

 

References:

1. FB Prophet API