Prophet

Prophet (Taylor and Letham, The American Statistician 2018) is Meta's automatic forecasting procedure for univariate time series with strong seasonal effects. It decomposes a series into trend, yearly/weekly/daily seasonality modeled with Fourier terms, and holiday effects from built-in country calendars, fitting everything through Stan with full posterior uncertainty intervals. The ds/y dataframe API and sklearn-style fit/predict made it the default first forecaster for demand, capacity, and growth teams: it is robust to missing data and outliers, handles trend changepoints, and ships diagnostics like cross-validation and component plots out of the box.

Meta placed Prophet in maintenance mode as of v1.4.0: only bug fixes and dependency bumps are accepted, with no new features planned, and the team points to Nixtla's statsforecast/neuralforecast and NeuralProphet for state-of-the-art accuracy. It remains MIT-licensed with R and Python packages on CRAN and PyPI, and at 20k+ GitHub stars it is still the most widely adopted open forecasting package in the industry.

Author-reported: strong results on Meta's internal forecasting tasks in the 2018 paper; no formal head-to-head benchmark numbers against modern foundation models published by the authors. Independent comparisons generally place it behind deep-learning methods on complex series, ahead on short seasonal business series with holidays.