mojolearn.ProphetForecaster#

class mojolearn.ProphetForecaster(growth='linear', changepoints=None, n_changepoints=25, changepoint_range=0.8, yearly_seasonality='auto', weekly_seasonality='auto', daily_seasonality='auto', seasonality_mode='additive', seasonality_prior_scale=10.0, holidays_prior_scale=10.0, changepoint_prior_scale=0.05, max_iter=2000, numeric_mode=None)[source]#

Bases: object

Prophet-style additive or multiplicative forecaster with linear growth.

Models a series as a piecewise-linear trend with changepoints, plus Fourier seasonalities and optional holiday indicator regressors, and fits the maximum a posteriori parameters with float32 L-BFGS (Prophet uses Stan, so results agree with the prophet package to a tolerance, not exactly). Prophet’s conventions are kept: y is scaled by max|y| and time to [0, 1] over the history; n_changepoints are placed evenly over the first changepoint_range of the history (the first dropped); per-feature prior scales; Prophet’s linear initialization.

Automatic seasonalities follow Prophet: yearly (period 365.25 days, order 10) when the history spans two years; weekly (7 days, order 3) when it spans two weeks at sub-weekly spacing; daily (1 day, order 4) when it spans two days at sub-daily spacing. Add others with add_seasonality().

Time t is given in days (floats) or as numpy datetime64 (converted to days since 1970). y may be one series of shape (n_obs,) or a batch of shape (batch_size, n_obs) sharing t.

Not supported: logistic and flat growth, uncertainty intervals and sampling (mcmc_samples, interval_width), extra regressors beyond holiday indicator columns, conditional seasonalities, float64 y.

Parameters:

  • growth ({“linear”}, default “linear”) – Trend type; other values raise NotImplementedError.

  • changepoints (array-like or None, default None) – Explicit changepoint times (same units as t); they must fall within the training data. None places n_changepoints automatically.

  • n_changepoints (int, default 25) – Number of automatic changepoints.

  • changepoint_range (float, default 0.8) – Fraction of the history in which automatic changepoints are placed.

  • yearly_seasonality, weekly_seasonality, daily_seasonality (“auto”, bool or int, default “auto”) – "auto" applies Prophet’s rule above, True forces the seasonality on with its default order, False turns it off, and an int turns it on with that Fourier order.

  • seasonality_mode ({“additive”, “multiplicative”}, default “additive”) – How seasonalities and holidays combine with the trend.

  • seasonality_prior_scale (float, default 10.0) – Prior scale of the seasonal coefficients.

  • holidays_prior_scale (float, default 10.0) – Prior scale of the holiday coefficients.

  • changepoint_prior_scale (float, default 0.05) – Prior scale of the trend changes (smaller is smoother).

  • max_iter (int, default 2000) – Maximum L-BFGS iterations.

  • numeric_mode (str or None, default None) – Numeric mode for this model ('identical', 'deterministic' or 'fast'); None uses the process default (see mojolearn.set_numeric_mode()).

Variables:

  • seasonalities (dict) – Seasonalities in use, by name: period, fourier_order and prior_scale.

  • changepoints_t_ (ndarray, float32) – Changepoint times on the scaled [0, 1] time axis.

  • params_ (ndarray of shape (batch_size, 3 + n_changepoints + n_features), float32) – Fitted parameters per series.

  • objective_ (ndarray of shape (batch_size,), float32) – Final objective value per series.

  • n_iter_ (ndarray of shape (batch_size,), int64) – L-BFGS iterations per series.

  • trend_ (ndarray) – Trend component at the times of the last predict() call.

Examples

>>> import numpy as np, mojolearn
>>> t = np.arange(730, dtype=np.float64)
>>> y = (10 + 0.01 * t + np.sin(2 * np.pi * t / 7)).astype(np.float32)
>>> model = mojolearn.ProphetForecaster().fit(t, y)
>>> model.predict(np.arange(730, 760, dtype=np.float64)).shape
(30,)
__init__(growth='linear', changepoints=None, n_changepoints=25, changepoint_range=0.8, yearly_seasonality='auto', weekly_seasonality='auto', daily_seasonality='auto', seasonality_mode='additive', seasonality_prior_scale=10.0, holidays_prior_scale=10.0, changepoint_prior_scale=0.05, max_iter=2000, numeric_mode=None)[source]#

Methods

add_seasonality(name, period, fourier_order)

Add a custom seasonality before fit().

fit(t, y[, holidays])

Fit the model.

predict(t[, holidays])

Predict at the given times.
add_seasonality(name, period, fourier_order, prior_scale=None)[source]#

Add a custom seasonality before fit().

Parameters:

  • name (str) – Seasonality name. Using "yearly", "weekly" or "daily" replaces the automatic one.

  • period (float) – Period in days, greater than 0.

  • fourier_order (int) – Number of Fourier terms, at least 1.

  • prior_scale (float or None, default None) – Prior scale; None uses seasonality_prior_scale.

Returns:

self

fit(t, y, holidays=None)[source]#

Fit the model.

Parameters:

  • t (array-like of shape (n_obs,)) – Times in days (floats) or numpy datetime64, sorted ascending, no NaN.

  • y (array-like of shape (n_obs,) or (batch_size, n_obs)) – Observations, float32 and finite; at least 2 points.

  • holidays (array-like of shape (n_obs, n_holidays) or None, default None) – Holiday indicator columns.

Returns:

self

Raises:

  • TypeError – If y is float64.

  • ValueError – If t holds NaN or is not sorted, y is not finite, the lengths differ, there are fewer than 2 points, or explicit changepoints fall outside the training data.

predict(t, holidays=None)[source]#

Predict at the given times.

Parameters:

  • t (array-like of shape (n_times,)) – Times in days or numpy datetime64, as for fit().

  • holidays (array-like of shape (n_times, n_holidays) or None, default None) – Holiday indicator columns for t; required when the model was fitted with holidays.

Returns:

ndarray of shape (n_times,) or (batch_size, n_times), float32 – The forecast yhat. The trend component is stored in trend_.