Prophet.jl

A Julia proof of concept for Prophet-style forecasting components.

This package is a Julia proof of concept for Prophet-style forecasting with a Python Prophet-like public API:

• Stan-equivalent trend and likelihood utilities in src/turing/prophet-turing.jl.
• Stan.jl dependency and bundled Stan model access through stan_model_file().
• A Turing model that mirrors the bundled Stan model's priors and likelihood.
• A Flux/Turing variant that keeps the Prophet terms equivalent and adds a Bayesian neural residual.
• Embedded holidays loaded from data/generated_holidays.csv, with no runtime holiday-package dependency.
• Plotting through Makie (backend=:makie) and Gadfly (backend=:gadfly).

Python-Style Usage

Julia's package module is also named Prophet, so the constructor is ProphetModel:

import Prophet

m = Prophet.ProphetModel()
Prophet.add_country_holidays(m; country_name="US")
Prophet.add_seasonality(m; name="monthly", period=30.5, fourier_order=3)
Prophet.fit(m, df)  # df is a DataFrame with ds and y columns
future = Prophet.make_future_dataframe(m; periods=365)
forecast = Prophet.predict(m, future)
fig = Prophet.plot(m, forecast; backend=:makie)
components = Prophet.plot_components(m, forecast; backend=:gadfly)

Select the modeling backend at construction time:

m_stan = Prophet.ProphetModel(model_backend=:stan)
m_turing = Prophet.ProphetModel(model_backend=:turing)
m_neural = Prophet.ProphetModel(model_backend=:neural_turing) # also accepts :flux_turing

fit routes through the selected backend:

• :stan uses the bundled Stan model through CmdStan.
• :turing uses the Stan-equivalent Turing model.
• :neural_turing keeps the Prophet terms and adds the Flux/Turing Bayesian neural residual.

The exported prophet and neural_prophet Turing models are tested directly against the Stan-equivalent model math.

If you prefer exported names:

using Prophet

m = ProphetModel()
fit(m, df)
forecast = predict(m, make_future_dataframe(m; periods=365))

Diagnostics

The Python diagnostics flow is available with Julia keyword arguments:

using Dates

df_cv = Prophet.cross_validation(
    m;
    initial=Day(730),
    period=Day(180),
    horizon=Day(365),
)
df_p = Prophet.performance_metrics(df_cv)

Python Prophet supports parallel="dask" for distributed cross-validation. In Julia, use Dagger for the same scheduler role:

using Dates

df_cv = Prophet.cross_validation(
    m;
    initial=Day(730),
    period=Day(180),
    horizon=Day(365),
    parallel=:dagger,
)

parallel=:threads is also available for a simple single-process threaded run.

Examples

Run the backend comparison example from the repo root:

julia --project=. examples/backend_comparison.jl

It fits the same synthetic daily series with :stan, :turing, and :neural_turing, then prints a small forecast summary and side-by-side backend preview.

There are also standalone examples for each backend:

julia --project=. examples/stan_backend.jl
julia --project=. examples/turing_backend.jl
julia --project=. examples/neural_turing_backend.jl

Each standalone example exposes a reusable runner:

include("examples/stan_backend.jl")
model, forecast, summary = run_stan_backend_example()

include("examples/turing_backend.jl")
model, forecast, summary = run_turing_backend_example()

include("examples/neural_turing_backend.jl")
model, forecast, summary = run_neural_turing_backend_example()

Pluto notebook versions are available in examples/pluto/:

• examples/pluto/stan_backend.jl
• examples/pluto/turing_backend.jl
• examples/pluto/neural_turing_backend.jl

Holidays

Python Prophet depends on the Python holidays package. For this POC, the package reads the generated CSV directly:

using Prophet

make_holidays_df([2015, 2016], "US")
get_holiday_names("US")
supported_holiday_countries()

The maintenance script src/scripts/generate_holidays_file.jl normalizes an existing holidays CSV into the embedded format:

julia --project=. src/scripts/generate_holidays_file.jl

The input/output paths and year range can be overridden with:

• PROPHET_HOLIDAYS_INPUT
• PROPHET_HOLIDAYS_OUTPUT
• PROPHET_HOLIDAYS_START_YEAR
• PROPHET_HOLIDAYS_END_YEAR

Plotting

plot_forecast(model, forecast; backend=:makie)
plot_forecast(model, forecast; backend=:gadfly)
plot_forecast_component(model, forecast, "trend"; backend=:makie)

Tests

julia --project=. -e 'using Pkg; Pkg.test()'

GitHub Actions tests Julia 1.10 and the latest stable Julia release. CI also downloads and builds CmdStan 2.37.0, sets both CMDSTAN and JULIA_CMDSTAN_HOME, and verifies the stanc version during the test run.

The test suite is organized under Julia's standard test/ directory and includes parity-oriented coverage inspired by Python Prophet's test_prophet, test_diagnostics, test_serialize, test_utilities, and model tests. The backend-facing tests exercise :stan, :turing, and :neural_turing where the runtime cost is appropriate for CI.