Prophet (forecasting library)
Prophet is an open-source software library for forecasting time series data, released by Facebook (now Meta) in February 2017 for Python and R.[3] It fits an additive regression model in which a trend with automatically chosen changepoints is combined with yearly, weekly and daily seasonality and user-supplied holiday effects, and the project describes it as working best on series with strong seasonal patterns and several seasons of history.[4] The method was described by Sean J. Taylor and Benjamin Letham in the paper "Forecasting at Scale", published in The American Statistician in 2018.[1] Independent benchmarks have repeatedly found it less accurate than classical methods such as exponential smoothing and ARIMA.[23][24][25] As of November 2026 the project is in maintenance mode: since version 1.4.0 the maintainers accept only bug fixes, dependency updates and R parity changes, with no new features planned.[4]
| Field | Value |
|---|---|
| Developer | Facebook Core Data Science team (Meta)[4] |
| Authors of the method | Sean J. Taylor, Benjamin Letham[1] |
| Initial public release | February 2017[3] |
| Latest release | 1.5.0 (October 2026)[5] |
| Languages | Python and R, with model fitting in Stan[3] |
| Python package | prophet on PyPI (named fbprophet before v1.0)[21][6] |
| R package | prophet on CRAN (version 1.1.7, behind the GitHub and PyPI releases)[22][4] |
| License | MIT License[21][22] |
| Status | Maintenance mode since v1.4.0[4] |
| Repository | github.com/meta/prophet[4] |
History
Taylor and Letham announced Prophet on the Facebook research blog on February 23, 2017. They presented it as a tool meant to help both experts and non-experts produce forecasts quickly, on the grounds that demand for forecasts inside the company exceeded the number of analysts able to make them. The post gave capacity planning and goal setting against a baseline as examples of the forecasting that large organizations like Facebook depend on, and said Prophet had become a key part of producing trustworthy forecasts at Facebook for decision-making and even for product features. It added that the core fitting procedure was written in Stan so that the Python and R versions shared the same code.[3] The README credits the release to Facebook's Core Data Science team.[4]
The accompanying paper was published in The American Statistician, volume 72, issue 1, pages 37-45, in 2018.[1] In it the authors framed "scale" not as a computing problem but as a people problem: many analysts without time series training making forecasts, a wide variety of series with their own quirks, and too many forecasts for anyone to check by hand. They wrote that computation and storage had been "relatively straightforward" by comparison.[2]
The Python package was published as fbprophet until version 1.0 (April 2021), which renamed it prophet.[6] Version 1.1 (June 2022) dropped the PyStan 2 dependency, made cmdstanpy the only Stan backend in Python, and began shipping prebuilt binaries for macOS, Linux and Windows so that users no longer had to compile the model themselves.[7] Version 1.1.5 (October 2023) added an option to scale the target with min-max scaling and a separate mode for holiday effects.[8] Releases from 2024 onward were mostly bug fixes and dependency updates; version 1.4.0 (August 2026) dropped support for Python versions older than 3.10, added static typing, and added optional nested Prophet models that forecast the future values of extra regressors.[9] The README states that Prophet is in maintenance mode as of v1.4.0.[4] As of November 2026 the GitHub repository sits under the meta organization with about 20,500 stars.[4]
| Date | Version or event | Notes |
|---|---|---|
| 2017-02-23 | Public release | Announced on the Facebook research blog[3] |
| 2017-03-05 | v0.1 | First tagged release[5] |
| 2018 | Paper published | The American Statistician 72(1)[1] |
| 2020-03-03 | v0.6 | Python 2 no longer supported[4][5] |
| 2021-04-02 | v1.0 | PyPI package renamed from fbprophet to prophet[6] |
| 2022-06-24 | v1.1 | cmdstanpy sole Python backend; prebuilt binaries; Python 3.7 minimum[7] |
| 2023-10-10 | 1.1.5 | scaling='minmax' option, holidays_mode, macOS arm64 binaries[8] |
| 2026-01-27 | v1.3.0 | Support for pandas 3.0 and NumPy 2.4[5] |
| 2026-08-01 | v1.4.0 | Python 3.10 minimum, static typing, nested regressor models; maintenance mode begins[9][4] |
| 2026-10-05 | v1.5.0 | Bug fixes in cross-validation, performance metrics and fitting; CPython 3.14 wheels[4][5] |
Model
Prophet treats forecasting as a curve-fitting problem. The paper describes a decomposable model with three components plus noise:[2]
Here $g(t)$ is the trend (non-periodic change), $s(t)$ is periodic seasonality, $h(t)$ is the effect of holidays and other irregular events, and $\epsilon_t$ is an error term assumed to be normally distributed. The authors compare the formulation to a generalized additive model in which time is the only regressor, and they acknowledge that it gives up some of the inferential advantages of a generative time series model such as ARIMA. In exchange, they argue, it handles several seasonal periods easily, does not need regularly spaced observations or interpolation of missing values, fits fast enough for interactive use, and has parameters that analysts can interpret.[2]
Trend
Prophet offers a piecewise linear trend and a saturating (logistic) growth trend.[2] The library also offers a flat trend for series dominated by seasonality or by external regressors.[17]
In the linear model the growth rate can change at a set of $S$ changepoints $s_1, \dots, s_S$. With base rate $k$, offset $m$, a vector of rate adjustments $\delta$, and an indicator vector $a(t)$ whose $j$-th entry is 1 once $t \ge s_j$, the trend is
where $\gamma_j = -s_j \delta_j$ keeps the curve continuous.[2] The logistic version replaces this with
where $C(t)$ is a carrying capacity that the user supplies and that may change over time. Here the offset adjustments $\gamma_j$ are not $-s_j \delta_j$: the paper gives a separate recursive formula for them so that the logistic segments still join at each changepoint. The paper motivates this with Facebook user growth, whose ceiling in a region might be the number of people with internet access.[2] In the library the capacity is given as a cap column, and a saturating minimum can be set with a floor column.[18]
Changepoints can be listed by hand, for example at product launches, but by default Prophet places many candidate changepoints and puts a sparse Laplace prior on the rate changes, $\delta_j \sim \text{Laplace}(0, \tau)$, so that most of them end up near zero. The documentation describes this prior as equivalent to L1 regularization.[2][10] By default the library places 25 candidate changepoints, spread uniformly over the first 80% of the history, and sets the prior scale changepoint_prior_scale to 0.05.[10]
Seasonality
Seasonal effects are modelled with a truncated Fourier series of period $P$ (in days):[2]
The $2N$ coefficients get a normal prior, $\beta \sim \text{Normal}(0, \sigma^2)$, which smooths the fit. The paper reports that $N = 10$ for yearly seasonality and $N = 3$ for weekly seasonality worked well for most problems, and these are the library defaults.[2][11] Weekly and yearly terms are fitted by default when the history covers more than two cycles, and daily seasonality is added for sub-daily data; users can add other periods such as monthly or quarterly with add_seasonality, and can make a seasonality conditional on another variable, such as a weekly pattern that differs between sports seasons.[11] Seasonality is additive by default, and a multiplicative mode is available for series whose seasonal swings grow with the trend.[12]
Holidays and regressors
Holidays and one-off events enter as indicator variables, one per event, each with its own coefficient under a normal prior.[2] Users pass a table of event names and dates covering both the history and the forecast period, optionally with a window of days before and after each date; an event with no future dates is fitted but left out of the forecast. Built-in country holiday lists come from the Python holidays package.[11] Additional regressors can be added to the linear part of the model, but their future values must be known or forecast separately.[11] The documentation's COVID-19 case study uses the same mechanism to treat lockdown periods as one-off "holidays", so that the drop in foot traffic is not absorbed into the trend.[19]
Fitting and uncertainty
The model is short enough to write in a few lines of Stan. By default Prophet finds a maximum a posteriori estimate with Stan's L-BFGS optimizer; full Bayesian inference with Markov chain Monte Carlo is optional and, according to the documentation, takes minutes rather than seconds.[2][13] Before fitting, the library divides y by its maximum absolute value in the history (or applies min-max scaling when asked).[17]
Uncertainty intervals (80% by default) come from simulating future trend changes on the assumption that they will occur with the same average frequency and magnitude as in the history, plus observation noise. Seasonal uncertainty is included only when MCMC sampling is used. The paper calls this assumption fairly strong and the documentation calls it probably untrue; both say the intervals should not be expected to have exact coverage.[2][13]
| Parameter | Default | Effect |
|---|---|---|
changepoint_prior_scale | 0.05 | Flexibility of the trend[10][14] |
n_changepoints | 25 | Number of candidate changepoints[10] |
changepoint_range | 0.8 | Share of history where changepoints may occur[10] |
seasonality_prior_scale | 10 | Strength of seasonality; 10 means little regularization[11][14] |
holidays_prior_scale | 10 | Strength of holiday effects[11] |
| Yearly / weekly Fourier order | 10 / 3 | Wiggliness of each seasonal curve[11] |
seasonality_mode | additive | Additive or multiplicative seasonality[12] |
interval_width | 0.8 | Width of uncertainty intervals[13] |
mcmc_samples | 0 | 0 means MAP fitting; above 0 runs MCMC[13] |
Analyst-in-the-loop approach
The paper presents the model as one half of a workflow. Analysts are expected to adjust a small set of interpretable settings using domain knowledge: capacities from market-size estimates, changepoints at known product changes, relevant holidays and seasonal periods, and the smoothing parameters for trend and seasonality.[2] The other half is automated evaluation. Because time series observations cannot be randomly split as in ordinary cross-validation, the authors propose "simulated historical forecasts": the model is refit at a series of cutoff dates and its forecasts are compared with what actually happened, with a cutoff roughly every half-horizon as a rule of thumb. Forecasts with large errors relative to simple baselines, errors at the same date across all methods (suggesting outliers), or sudden jumps in error between cutoffs are flagged for a person to look at.[2]
The library implements this as cross_validation and performance_metrics, which report MSE, RMSE, MAE, MAPE and other measures as a function of the forecast horizon. By default the initial training window is three times the horizon and cutoffs fall every half horizon, and the work can be run in parallel with processes, threads or Dask.[14] The documentation recommends using this procedure for hyperparameter tuning, mainly of changepoint_prior_scale and seasonality_prior_scale.[14]
The paper's own evidence for the approach is limited. Its accuracy comparison uses a single series, the daily number of events created on Facebook, with nine simulated forecast dates; on that series Prophet with default settings had lower mean absolute percentage error than auto.arima, ets, seasonal naive and TBATS from the R forecast package at every horizon.[2] The authors wrote that they had "only anecdotal empirical evidence" that analyst input improves accuracy.[2]
Usage
The Python API follows the scikit-learn pattern: create a Prophet object, call fit on a pandas dataframe with a date column ds and a numeric column y, build a future dataframe with make_future_dataframe, and call predict, which returns the forecast yhat with yhat_lower, yhat_upper and the individual components. The R package provides a prophet function that fits the model and returns an object for predict and plot. The documentation's running example is the log of daily page views of the Wikipedia article on Peyton Manning.[20]
from prophet import Prophet
m = Prophet() # default trend, seasonality and priors
m.fit(df) # df has columns ds (dates) and y (values)
future = m.make_future_dataframe(periods=365)
forecast = m.predict(future)
The documentation also lists cases where the model needs care. Outliers can inflate the uncertainty intervals or distort seasonality, and the recommended fix is to remove them, since Prophet tolerates missing values.[15] The model is continuous in time, so fitting monthly data and asking for daily forecasts gives unreliable results.[16] Fitted Python models should be saved with the built-in JSON serialization rather than pickle, and a model cannot be updated with new data, only refit (optionally warm-started from earlier parameters).[17]
Reception and criticism
Accuracy has been the main point of dispute. In July 2017, forecasting researcher Nikolaos Kourentzes, who wrote that Prophet had appeared in quite a few online discussions since its release, tested it on the M3 competition data (3,003 yearly, quarterly, monthly and other series) with rolling-origin evaluation and mean absolute scaled error. Prophet ranked last of the eight forecasts he compared (seven methods, one of them in two settings) in every subset, by both mean and median error; its mean error was about 24% to 44% worse than exponential smoothing (ETS). Kourentzes noted that M3 lacks the high-frequency, multi-seasonal data Prophet was designed for and that more testing on such data was needed, but concluded that for typical lower-frequency business forecasting it did not perform well.[23]
The textbook Forecasting: Principles and Practice by Rob J. Hyndman and George Athanasopoulos covers Prophet in its third edition. In its quarterly cement production example, Prophet's test-set forecasts were worse than both ETS and ARIMA. The authors credit Prophet with being fast to estimate and completely automated but conclude that it rarely forecasts more accurately than the alternatives.[24] A 2021 study in the journal Forecasting on wholesale food prices for an Italian food wholesaler reached a similar verdict: Prophet was "quick and easy to use, but considerably less accurate" than ARIMA and LSTM models.[25]
Other criticism concerns the model's structure. The authors of NeuralProphet (2021) argued that Prophet lacks local context, which matters for near-term forecasts, and that its Stan backend makes it hard to extend. Their PyTorch-based successor keeps Prophet's components and adds autoregression and covariate modules; they reported accuracy gains of 55% to 92% over Prophet on short- to medium-term forecasts.[26] The Prophet paper itself flags some of the same trade-offs, noting that it gives up the inferential advantages of a generative model and that its uncertainty intervals do not have exact coverage.[2] The maintainers' own documentation now says they have no plans to further develop the underlying model, and it points users looking for state-of-the-art accuracy to statsforecast and other packages from the Nixtla group, and to NeuralProphet.[17]
Related software
Prophet's model has been reimplemented and wrapped by other projects. NeuralProphet re-creates the components in PyTorch with added neural network modules.[26] Forecasting: Principles and Practice fits Prophet through the R package fable.prophet.[24] The Prophet README also links Helter Skelter, a simplified reimplementation in Scala on Apache Spark for distributed computing.[4]
Continue reading
- ARIMAARIMA (Autoregressive Integrated Moving Average) is a class of statistical models for analyzing and forecasting time series data, specified by three non-negative integer orders written as ARIMA(p, d, q): p is…Mentioned 5 times above
- Time Series AnalysisTime series analysis is the statistical study of ordered observations when dependence across time is part of the signal rather than a nuisance to ignore.Listed in See also
- Temporal dataTemporal data is data where each observation is tagged with a timestamp, so the order in which observations arrive carries meaning.Listed in See also
- Bayesian statisticsBayesian statistics is a statistical paradigm in which probability expresses a degree of belief that is updated as evidence arrives, using Bayes' theorem.Listed in See also
- OverfittingOverfitting is a failure of generalization: a fitted model or a model-selection procedure performs better on the observations used to develop it than on new observations from the population of interest.Listed in See also
- Anomaly DetectionIn machine learning, anomaly detection is the process of identifying data points, events, or observations that deviate significantly from normal patterns in a dataset.Listed in See also
More to read
References
- ^1 ^2 ^3 ^4Taylor, Sean J.; Letham, Benjamin. "Forecasting at Scale". The American Statistician, 72(1): 37-45, 2018. doi.org/...00031305.2017.1380080
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17Taylor, Sean J.; Letham, Benjamin. "Forecasting at Scale" (preprint). Prophet repository. github.com/...prophet_paper_preprint.pdf
- ^1 ^2 ^3 ^4 ^5Taylor, Sean J.; Letham, Ben. "Prophet: forecasting at scale". Meta Research blog, February 23, 2017. research.facebook.com/...phet-forecasting-at-scale
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Meta. "Prophet: Automatic Forecasting Procedure" (README). GitHub, accessed November 18, 2026. github.com/...prophet
- ^1 ^2 ^3 ^4 ^5Prophet maintainers. "Releases". GitHub, accessed November 18, 2026. github.com/...releases
- ^1 ^2 ^3Prophet maintainers. "v1.0" release notes. GitHub, April 2, 2021. github.com/...v1.0
- ^1 ^2Prophet maintainers. "v1.1" release notes. GitHub, June 24, 2022. github.com/...v1.1
- ^1 ^2Prophet maintainers. "1.1.5" release notes. GitHub, October 10, 2023. github.com/...1.1.5
- ^1 ^2Prophet maintainers. "v1.4.0" release notes. GitHub, August 1, 2026. github.com/...v1.4.0
- ^1 ^2 ^3 ^4 ^5Prophet documentation. "Trend Changepoints". Meta, accessed November 18, 2026. meta.github.io/...trend_changepoints
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Prophet documentation. "Seasonality, Holiday Effects, And Regressors". Meta, accessed November 18, 2026. meta.github.io/..._holiday_effects,_and_regressors
- ^1 ^2Prophet documentation. "Multiplicative Seasonality". Meta, accessed November 18, 2026. meta.github.io/...multiplicative_seasonality
- ^1 ^2 ^3 ^4Prophet documentation. "Uncertainty Intervals". Meta, accessed November 18, 2026. meta.github.io/...uncertainty_intervals
- ^1 ^2 ^3 ^4Prophet documentation. "Diagnostics". Meta, accessed November 18, 2026. meta.github.io/...diagnostics
- ^Prophet documentation. "Outliers". Meta, accessed November 18, 2026. meta.github.io/...outliers
- ^Prophet documentation. "Non-Daily Data". Meta, accessed November 18, 2026. meta.github.io/...non-daily_data
- ^1 ^2 ^3 ^4Prophet documentation. "Additional Topics". Meta, accessed November 18, 2026. meta.github.io/...additional_topics
- ^Prophet documentation. "Saturating Forecasts". Meta, accessed November 18, 2026. meta.github.io/...saturating_forecasts
- ^Prophet documentation. "Handling Shocks". Meta, accessed November 18, 2026. meta.github.io/...handling_shocks
- ^Prophet documentation. "Quick Start". Meta, accessed November 18, 2026. meta.github.io/...quick_start
- ^1 ^2Python Package Index. "prophet". PyPI, accessed November 18, 2026. pypi.org/...prophet
- ^1 ^2CRAN. "prophet: Automatic Forecasting Procedure". The Comprehensive R Archive Network, accessed November 18, 2026. cran.r-project.org/package=prophet
- ^1 ^2Kourentzes, Nikolaos. "Benchmarking Facebook's Prophet". Nikolaos Kourentzes (blog), July 29, 2017. kourentzes.com/...benchmarking-facebooks-prophet
- ^1 ^2 ^3Hyndman, Rob J.; Athanasopoulos, George. "12.2 Prophet model". Forecasting: Principles and Practice, 3rd edition. OTexts. otexts.com/...advanced
- ^1 ^2Menculini, Lorenzo; Marini, Andrea; Proietti, Massimiliano; Garinei, Alberto; Bozza, Alessio; Moretti, Cecilia; Marconi, Marcello. "Comparing Prophet and Deep Learning to ARIMA in Forecasting Wholesale Food Prices". Forecasting, 3(3): 644-662, 2021. doi.org/...forecast3030040
- ^1 ^2Triebe, Oskar; Hewamalage, Hansika; Pilyugina, Polina; Laptev, Nikolay; Bergmeir, Christoph; Rajagopal, Ram. "NeuralProphet: Explainable Forecasting at Scale". arXiv:2111.15397, 2021. arxiv.org/...2111.15397
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
2 revisions · v2 · 2,964 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check before publication on November 18, 2026: 96 claims checked against 27 cited and independent sources; 1 material and 4 minor corrections applied before publication.
Cite this page: AI Wiki. "Prophet (forecasting library)." aiwiki.ai, updated 18 Nov 2026, fact-checked 18 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/prophet