Forecasting at Scale

When Taylor and Letham say "at scale," they do not mean computational scale — fitting thousands of time series is straightforward to parallelize. They mean three kinds of human scale:

1. Many people need to make forecasts, most without training in time series methods.
2. Many varieties of time series, each with idiosyncratic features like product launches or regional holidays.
3. Many forecasts must be monitored, compared, and corrected efficiently.

The traditional pipeline — hand an expert a single time series to study for days — collapses when an organization needs hundreds of forecasts per week. The authors' solution is an analyst-in-the-loop system: an interpretable model whose parameters an analyst can adjust with domain knowledge, combined with automated evaluation tools that surface the forecasts most likely to need human attention.

Prophet models a time series as the sum of three components plus noise. Think of it as three transparent layers stacked on top of each other, each capturing a different pattern. The growth layer $g(t)$ captures the long-term trajectory — is the metric going up, flattening, or declining? The seasonality layer $s(t)$ captures repeating cycles — weekly patterns, yearly patterns, or both. The holiday layer $h(t)$ captures irregular spikes and dips from holidays and special events. Whatever is left over goes into the noise term $\epsilon_t$.

This additive structure is what makes the model interpretable: you can look at each layer separately and understand what is driving the forecast.

The growth component $g(t)$ captures the non-periodic trend — the big picture of where the metric is heading. Prophet offers two growth models depending on the nature of the data.

Logistic growth is used when there is a natural upper limit — a carrying capacity. For example, the number of Facebook users in a country is bounded by the number of people with internet access. The classic logistic curve has the S-shape of population growth: slow start, rapid rise, then saturation.

Linear growth is used when no natural saturation exists — for example, revenue or page views that could in principle grow indefinitely. This is simply a piecewise line with slope changes at key moments.

In both cases, Prophet allows the growth rate to change at specific points called changepoints. A product launch might accelerate growth; a competitor's entry might slow it. Rather than fitting a single curve that cannot bend, Prophet fits a flexible piecewise curve that can adapt to shifts in the underlying dynamics.

Business metrics reflect human behavior, and humans are creatures of routine. People check social media more on weekdays, shop more before holidays, and travel more in summer. These repeating patterns are the seasonality component $s(t)$.

Prophet approximates seasonality using a Fourier series — a sum of sine and cosine waves at different frequencies. Think of it like tuning a radio: each frequency captures a different rhythmic cycle. For weekly seasonality, Prophet uses $N=3$ Fourier terms (capturing the broad weekday vs weekend pattern). For yearly seasonality, $N=10$ terms are typical (capturing finer distinctions like summer vacations and back-to-school periods).

The key advantage: you do not need to hand-specify what day-of-week effects look like. The Fourier basis automatically learns the shape from data.

Holidays like Thanksgiving (fourth Thursday in November) or the Super Bowl do not fall on the same date each year, so smooth seasonal curves cannot model them. Prophet handles this with a separate holiday component.

The analyst provides a table of holiday names and dates — past and future. Each holiday $i$ gets its own parameter $\kappa_i$ that measures its impact on the metric. The model can also include windows of days around a holiday. For instance, people behave differently in the days leading up to Christmas and after it.

This is one of Prophet's most practical features: by simply listing the holidays relevant to a country or business, the model automatically learns their effects without any statistical tuning.

Prophet frames forecasting as a curve-fitting problem rather than a generative time series model. This is a deliberate trade-off: it gives up the temporal dependence modeling of ARIMA (where each observation depends on the previous ones) in exchange for speed, flexibility, and interpretability.

The model is implemented in Stan and can be fit in two ways. The default is MAP estimation using L-BFGS — the same optimizer used in logistic regression — which takes seconds even for years of daily data. When full uncertainty quantification is needed, Prophet supports full Bayesian inference via Hamiltonian Monte Carlo sampling in Stan.

The critical design choice is that the model has a small number of analyst-interpretable parameters. The table below summarizes them. An analyst does not need to understand Fourier series or Laplace priors — they just turn a knob labeled "how flexible should the trend be?" or "how strong is the seasonality?"

How do you know if your forecast model is any good? You cannot use standard cross-validation because time series observations are not exchangeable — you cannot randomly shuffle dates.

Prophet introduces Simulated Historical Forecasts (SHFs): you pick multiple cutoff dates in the past, pretend you only have data up to that point, generate a forecast for the next $H$ days, then compare the forecast to what actually happened. This simulates the errors you would have made if you had used this method at that point in history.

By computing SHFs across multiple cutoff dates and comparing against baseline methods (naive, seasonal naive, ARIMA, exponential smoothing), Prophet can automatically surface forecasts that are performing poorly and flag them for human review. This is what makes the analyst-in-the-loop system work at scale: instead of checking every forecast, the analyst focuses on the ones the system identified as problematic.

ARIMA and Prophet represent fundamentally different approaches to time series. ARIMA is a generative model that explicitly accounts for the temporal dependence structure — each value depends on previous values. It is mathematically elegant and has strong inferential properties, but it requires regularly spaced data, careful differencing to achieve stationarity, and expertise to select orders $(p, d, q)$.

Prophet treats forecasting as curve fitting. It does not model the autocorrelation structure at all. Instead, it directly models the patterns humans care about — trend, seasons, and holidays — and leaves the residuals as unstructured noise. This makes it less rigorous theoretically, but far more practical for the analyst-in-the-loop scenario where speed, interpretability, and flexibility matter more than inferential guarantees.

Prophet can also handle missing data naturally (no interpolation needed), irregular spacing, and multiple seasonality periods — all of which are painful with ARIMA.