Forecast Time Series with Prophet

Prophet Forecasting Expert is a skill for time series forecasting with Facebook Prophet, covering model configuration, parameter tuning, seasonality handling, and production deployment. It explains Prophet's additive decomposition into trend (piecewise linear or logistic growth with automatic changepoint detection), seasonality (Fourier series for weekly, yearly, and custom patterns), holidays (user-defined irregular events with their own prior scales), and an error term. Standard setup imports Prophet plus its plotting (plot_plotly, plot_components_plotly), diagnostics (cross_validation, performance_metrics), and serialization (model_to_json, model_from_json) modules:

import pandas as pd
import numpy as np
from prophet import Prophet
from prophet.plot import plot_plotly, plot_components_plotly
from prophet.diagnostics import cross_validation, performance_metrics
from prophet.serialize import model_to_json, model_from_json
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore')

Use it when forecasting a time series that has at least a year of daily data with clear seasonal patterns - Prophet's stated sweet spot. It provides a full data-preparation routine (renaming to Prophet's required ds/y columns, numeric coercion, dropping missing values, deduplicating same-timestamp rows by averaging, resampling to a regular frequency, and linear interpolation of gaps), so it fits workflows starting from irregular or messy raw data. It is not suited to series shorter than roughly a year, series without meaningful seasonality, or use cases needing a fundamentally different forecasting family (e.g. ARIMA-only or deep-learning sequence models) rather than Prophet's additive structure.

Input is a ds/y-formatted dataframe and a set of Prophet hyperparameters: growth ('linear' or 'logistic'), seasonality_mode ('additive' or 'multiplicative'), seasonality_prior_scale and holidays_prior_scale (flexibility, 0.01-10), changepoint_prior_scale (trend flexibility, 0.001-0.5), changepoint_range (fraction of history eligible for changepoints), n_changepoints, interval_width, and uncertainty_samples. Output is a fitted model and a forecast dataframe produced via make_future_dataframe() and model.predict(), including custom 90% confidence bounds computed from yhat_upper/yhat_lower. The skill also covers custom seasonalities (add_seasonality(name='monthly', period=30.5, fourier_order=5), a quarterly equivalent), custom holiday tables with lower_window/upper_window (e.g. a Black Friday effect), and external regressors added via add_regressor(name, prior_scale, standardize).

For validation it wires in cross_validation() with configurable initial/period/horizon windows (e.g. 730/90/365 days) and parallel='processes' for scaling, feeding performance_metrics() for a grid search over parameters like changepoint_prior_scale and seasonality_mode to minimize a chosen metric (e.g. MAPE). It also documents ensemble forecasting - bootstrap-sampling the training data across several Prophet models and averaging their predictions with a standard-deviation spread - and a diagnostics routine computing MAE/MAPE/RMSE plus residual-over-time, residual-histogram, Q-Q (via scipy.stats.probplot), and actual-vs-predicted plots.

Data scientists and analysts building production forecasting pipelines who need Prophet configured correctly for their seasonality and business constraints, plus the surrounding validation and deployment machinery: model persistence via model_to_json()/model_from_json(), periodic retraining, monitoring prediction-interval coverage and residual drift, and chunked/parallel processing for large-scale forecasting jobs.