Add Regressor Weather Forecast Data

[c3-VincentBrisse]

To help with forecasting water consumption, I would like to add a regressor with weather data. I have three years of data.
During cross validation, I first train on the first two years, then predict on the next two weeks, then train on the first two years plus one week, then predict on the next two weeks etc...
To avoid data leakage in validation, I should only use weather forecast data, while to have a better training I should obviously use actual weather data. Is there a way to implement this with prophet ?

[bletham]

This came up in a discussion a couple years ago. I made an issue for it in #442, but it isn't something that there's any real interface for. I think the best thing to do will be to write out the CV loop yourself.

Which fortunately I do not think is too difficult. This is all of the code for doing cross validation for a single cutoff:

prophet/python/fbprophet/diagnostics.py

Lines 167 to 197 in 3c69ce3

	# Generate new object with copying fitting options
	m = prophet_copy(model, cutoff)
	# Train model
	history_c = df[df['ds'] <= cutoff]
	if history_c.shape[0] < 2:
	raise Exception(
	'Less than two datapoints before cutoff. '
	'Increase initial window.'
	)
	m.fit(history_c, **model.fit_kwargs)
	# Calculate yhat
	index_predicted = (df['ds'] > cutoff) & (df['ds'] <= cutoff + horizon)
	# Get the columns for the future dataframe
	columns = ['ds']
	if m.growth == 'logistic':
	columns.append('cap')
	if m.logistic_floor:
	columns.append('floor')
	columns.extend(m.extra_regressors.keys())
	columns.extend([
	props['condition_name']
	for props in m.seasonalities.values()
	if props['condition_name'] is not None])
	yhat = m.predict(df[index_predicted][columns])
	# Merge yhat(predicts), y(df, original data) and cutoff
	return pd.concat([
	yhat[predict_columns],
	df[index_predicted][['y']].reset_index(drop=True),
	pd.DataFrame({'cutoff': [cutoff] * len(yhat)})
	], axis=1)

Right now it is calling predict on df[index_predicted][columns]. And you would just need to change that to plug in the forecasted values for the regressor in the window [cutoff, cutoff + horizon].

That does CV for a single cutoff, and then you would just loop over cutoffs and concatenate all of the results.

[c3-VincentBrisse]

Thank you that answers my question !