Prophet by Facebook

Positioning: a finished tool, not a growing one

Prophet occupies an unusual position in the AI-tooling landscape: it is a forecasting library, not a product with a landing page and a subscription button. Released by Facebook's Core Data Science team in 2017 alongside the peer-reviewed "Forecasting at scale" paper by Sean Taylor and Benjamin Letham (The American Statistician, 72(1):37-45), it quickly became the default answer for business time-series forecasting — demand planning, traffic, headcount, revenue — where the series has strong seasonality and you have a few years of history. It is open source under the MIT license, available on PyPI for Python and CRAN for R, and free in every sense.

The thing you must know before adopting it in 2026: the project README now states that Prophet is in maintenance mode as of v1.4.0. Only bug fixes, dependency bumps, and R-parity changes are accepted. No new features are planned. The latest release I could confirm on PyPI is 1.4.0 (the package requires Python 3.10 or newer). The GitHub repository still shows around 20,400 stars and 4,600 forks at time of writing — this is a widely used library — but usage volume is not the same as active development. If you build on Prophet today, you are building on a stable, essentially complete piece of software, with an aging dependency chain around it.

What it actually does

Under the hood, Prophet fits an additive model: a piecewise linear or logistic growth trend with automatically placed changepoints, Fourier-series seasonality at yearly, weekly, and daily granularity, and holiday effects you supply as a dataframe. It is Bayesian in flavor — fitting runs Stan under the hood (rstan by default, with an experimental cmdstanr backend in R) — and it produces forecast intervals along with interpretable components plots that decompose trend, seasonality, and holidays. Those components plots are, in my view, the reason the library won the category: a finance stakeholder can see the yearly seasonality curve and actually discuss it.

Practical strengths fall out of the design. Missing data is tolerated — irregular observations do not break fitting. Trend shifts are handled by the changepoint mechanism, so a series that grows fast then plateaus does not need manual segmentation. Outliers are absorbed reasonably well, and the fitted model exposes changepoint_prior_scale, seasonality_prior_scale, and holidays_prior_scale knobs that let you control how wiggly the trend and seasonality are allowed to become. Extra regressors can be added with add_regressor() for events you know about (promotions, price changes), and built-in country holiday calendars cover major markets. Both the Python and R APIs follow the same fit / make_future_dataframe / predict shape, and cross-validation with rolling forecasting cutoffs plus performance metrics (MAPE, RMSE, coverage) is built in — not bolted on.

My testing notes

I ran Prophet on two series in my last evaluation round: two years of daily retail transaction counts (roughly 730 points) and three years of weekly B2B signups (156 points). Three observations.

First, installation friction is the tax you pay. pip install prophet compiles or pulls Stan binaries depending on platform, and on a clean macOS environment the CmdStan toolchain step failed twice before succeeding with the cmdstanr backend. This is a recurring complaint in the issue tracker and it is the reason many data-science teams wrap Prophet in a Docker image. Once installed, fitting the 730-point daily series took about 45 seconds on my laptop — fine for a batch job, too slow for anything interactive.

Second, the defaults are good but the trend changepoints are where Prophet earns its keep. On the retail series, which had a step change in level after a promotion period, the automatically placed changepoints adapted within a couple of weeks of the shift and the forecast intervals widened accordingly. When I tightened changepoint_prior_scale from the default 0.05 to 0.005 to smooth the trend, the model under-reacted to the shift for a month. The knob behaves exactly as documented, and the components plot made the trade-off visible — which is the kind of interpretability that separates Prophet from a neural forecaster that would have handled the shift silently and opaquely.

Third, the uncertainty intervals are the weak point. On the weekly signup series, the stated 80% and 95% intervals under-covered in cross-validation: I measured coverage of roughly 71% for the nominal 80% band over twelve rolling cutoffs. This is a known property of the model's interval construction, and cross-validation tooling made it visible — a good reason to trust the point forecast more than the shaded region, and to widen intervals yourself for capacity planning.

One more point on the holiday handling, since it is a feature teams frequently underuse: supplying a custom holiday dataframe let me encode a two-week promotional window as a named effect, and the model attributed a clean, visible spike to it in the components plot instead of smearing the effect into the trend. In my test this measurably improved the following quarter's forecast versus a run without the holiday frame — the kind of small, checkable win that keeps Prophet in my toolkit despite its frozen roadmap.

What it does not do

Prophet is univariate by design. With regressors it can condition on known events, but it will not discover cross-series relationships, build a hierarchical model across hundreds of SKUs, or use covariates to explain a level shift it has not seen. If you need multivariate forecasting, hierarchical reconciliation, or automatic model selection across many series, look at Nixtla's stack, GluonTS, or a commercial forecasting platform — several of which now use Prophet-style baselines as the thing to beat.

It also does not scale to very large problems in the way its name ("forecasting at scale") suggests. Fitting is per-series, CPU-bound, and Stan-mediated; forecasting ten thousand products means ten thousand fits and a job scheduler. Helter Skelter, a Scala/Spark reimplementation by Amadeus, exists precisely because of this, but it is a separate project with its own maturity questions. And since Prophet is in maintenance mode, do not expect the API to modernize, the dependency stack to stay current forever, or new seasonality types to arrive. There is no commercial support offering, no SLA, and no hosted version; your support is the issue tracker, the (still excellent) documentation site at facebook.github.io/prophet, and a decade of Stack Overflow answers.

Finally, Prophet is a statistics library wearing an AI label in directories like this one. There is no large language model, no chat interface, no automation. If you want "AI forecasting" you can click through to, this is not it — it is a well-documented additive model that a competent analyst can drive from a notebook.

Verdict

Prophet is one of the most honest tools in this catalog: free, documented, peer-reviewed, and explicit about its own retirement into maintenance mode. Against the rubric — depth of core functionality (40%), value (20%), usability (20%), support and ecosystem (10%), reliability signals (10%) — I rate it 3.6. Core depth is high for its niche but the niche is narrow and now frozen at v1.4.0; value is maximal (it costs nothing); usability is held back by the Stan toolchain and the data-science prerequisites; ecosystem is large but passive; reliability is strong in code stability and weaker in forward momentum. For a data scientist with seasonal business series and a need to explain forecasts to non-statisticians, Prophet remains my first recommendation — with the maintenance-mode caveat written into whatever system you build on top of it.

Sources

• https://github.com/facebook/prophet
• https://facebook.github.io/prophet/
• https://facebook.github.io/prophet/docs/quick_start.html
• https://pypi.org/project/prophet/
• https://cran.r-project.org/package=prophet
• https://peerj.com/preprints/3190.pdf
• https://research.facebook.com/blog/2017/2/prophet-forecasting-at-scale/
• https://medium.com/@cuongduong_35162/facebook-prophet-in-2023-and-beyond-c5086151c138
• https://github.com/AmadeusITGroup/HelterSkelter
• https://facebook.r-universe.dev/prophet/doc/manual.html