ProphetDAG

Prophet allows you to add regressors to your forecast to help better predict the future. When adding a regressor you must know the values in the past (for training) and in the future (for prediction).

But often you want to use something as a regressor where you don't know what the future values will be. The Prophet documentation has this to say about this:

One can also use as a regressor another time series that has been forecasted with a time series model, such as Prophet. For instance, if r(t) is included as a regressor for y(t), Prophet can be used to forecast r(t) and then that forecast can be plugged in as the future values when forecasting y(t). A note of caution around this approach: This will probably not be useful unless r(t) is somehow easier to forecast then y(t). This is because error in the forecast of r(t) will produce error in the forecast of y(t)

Prophet does not propagate the uncertainty for r(t) into the forecast for y(t); this is where ProphetDAG can help.

See the notebook for a worked example.

The key parts of the example are as follows:

Define the DAG

from prophetDAG import ProphetDAG
import networkx as nx
from fbprophet import Prophet

graph = nx.DiGraph()
graph.add_nodes_from([(1,{'name':'total'})
                     ,(2,{'name':'Hamilton'})
                     ,(3,{'name':'Washington'})
                     ,(4,{'name':'Franklin'})])
graph.add_edge(2,1)
graph.add_edge(3,1)
graph.add_edge(4,1)

Create a Prophet Model for each node and attach the data

for (i,d) in zip(graph.nodes,[total,hamilton,washington,franklin]):
    graph.nodes[i]['df'] = d
    m = Prophet()
    m.add_country_holidays(country_name="US")
    m.history_dates = history_dates
    m.start = start
    m.t_scale = t_scale
    future = m.make_future_dataframe(periods=365)
    graph.nodes[i]['prophet'] = m
    graph.nodes[i]['future'] = future

Fit the model

p = ProphetDAG()
# Or specify the number of samples in the output
# p = ProphetDAG(n_samp=1000)
result = p.fit(graph)