Prophet Cheat Sheet

01Install & Importonce per env

pip install prophet★
PyPI name since v1.0 (was fbprophet).
conda install -c conda-forge prophet
Often the smoothest install (bundles the compiler).
from prophet import Prophet★
The estimator — sklearn-style fit/predict.
from prophet.plot import add_changepoints_to_plot
Plot helpers: also plot_plotly.
from prophet.diagnostics import cross_validation
And performance_metrics for evaluation.

02The Data Contractds & y, always

df.columns = ['ds', 'y']★
The only two columns Prophet requires.
df['ds'] = pd.to_datetime(df.ds)★
ds = date/timestamp; YYYY-MM-DD or with time.
# y must be numeric
The measurement to forecast (univariate).
df.loc[outliers, 'y'] = None
Set outliers to NaN — Prophet still fits.
df['y'] = np.log(df.y)
Log-transform tames exponential growth.

03Core Pipelinethe whole job

m = Prophet()★
Defaults auto-detect yearly/weekly seasonality.
m.fit(df)★
Fit on history — takes a second or two.
future = m.make_future_dataframe(periods=365)★
A ds-only frame extended N days ahead.
m.make_future_dataframe(24, freq='MS')
Month-start / any pandas freq for non-daily data.
forecast = m.predict(future)★
Fills yhat + interval for every row.
Prophet().fit(df).predict(future)
Chainable — the sklearn one-liner.

04Read the Forecastforecast columns

forecast['yhat']★
The point prediction.
yhat_lower · yhat_upper★
Uncertainty interval (default 80%).
trend
The g(t) trend component.
yearly · weekly · daily
Each seasonality's contribution.
holidays · <regressor>
Event & regressor effects, per column.
additive_terms · multiplicative_terms
Summed non-trend pieces.
forecast[['ds','yhat','yhat_lower','yhat_upper']].tail()
The slice you actually report.

05Visualizesee the fit

m.plot(forecast)★
History dots + forecast line + interval band.
m.plot_components(forecast)★
Trend + each seasonality broken out.
add_changepoints_to_plot(fig.gca(), m, forecast)
Overlay where the trend bends.
from prophet.plot import plot_plotly
Interactive, zoomable version.
plot_yearly(m) · plot_weekly(m)
Inspect one seasonality alone.

06Seasonalitys(t) · Fourier

Prophet(seasonality_mode='multiplicative')★
When swings grow with the trend (default additive).
Prophet(yearly_seasonality=True)
'auto' / True / False / int order.
m.add_seasonality(name='monthly', period=30.5, fourier_order=5)★
Add a custom cycle (period in days).
Prophet(yearly_seasonality=20)
Raise Fourier order (default 10 yearly / 3 weekly).
Prophet(seasonality_prior_scale=10)
Flexibility — lower dampens / smooths.
add_seasonality(..., condition_name='is_on')
Conditional season — needs a boolean column.
# more Fourier terms = overfittingcare
N terms ⇒ 2N variables per cycle.

07Trend & Changepointsg(t) · the #1 knob

Prophet(changepoint_prior_scale=0.05)★
↑ = flexible (overfit) · ↓ = smooth (underfit).
Prophet(n_changepoints=25)
Candidate points — rarely worth changing.
Prophet(changepoint_range=0.8)
Place them in the first 80% of history.
Prophet(changepoints=['2020-03-01'])
Specify bend dates by hand.
Prophet(growth='flat')
No trend at all — seasonality only.
# sparse L1 priorhow
Many candidates, few actually used.

08Saturating Growthlogistic · cap/floor

09Holidays & Eventsh(t)

m.add_country_holidays(country_name='IN')★
Built-in sets: US, IN, UK, BR…
pd.DataFrame({'holiday':…, 'ds':…})★
Custom events need columns holiday + ds.
lower_window · upper_window
Spread the effect ±days around the date.
Prophet(holidays=holidays_df)
Pass your event frame at construction.
from prophet.make_holidays import make_holidays_df
Generate a frame per country / province.
Prophet(holidays_prior_scale=10)
Lower it if holiday spikes overfit.

10Extra Regressorsexternal drivers

m.add_regressor('temperature')★
Fold in a known external variable.
# column in fit AND future dfmust
Prophet won't invent future regressor values.
add_regressor('x', mode='multiplicative')
Additive (default) or scales with trend.
add_regressor('x', standardize='auto')
Z-scores unless the column is binary.
add_regressor('x', prior_scale=0.5)
Regularize the coefficient.
regressor_coefficients(m)
Inspect the fitted betas.

11Uncertainty Intervalsthe band

Prophet(interval_width=0.80)★
Default is 80%, not 95% — widen if needed.
Prophet(uncertainty_samples=1000)
MC draws for the band; set 0 to skip & speed up.
Prophet(mcmc_samples=300)
Full Bayes → uncertainty in seasonality too.
# band = trend + noisenote
Seasonality is treated as certain unless MCMC.

12Cross-Validationrolling origin

df_cv = cross_validation(m, horizon='365 days',
initial='730 days', period='180 days')
performance_metrics(df_cv)★
rmse · mae · mape · mdape · smape · coverage.
plot_cross_validation_metric(df_cv, metric='mape')
Error vs. horizon, rolling-averaged.
cross_validation(m, ..., parallel='processes')
Fit each cutoff in parallel.
# never k-fold a time seriestime
Cutoffs move forward — don't shuffle.

13Hyperparameter Tuningwhat to sweep

14Save & Loadpersist a model

from prophet.serialize import model_to_json★
Also model_from_json.
json.dump(model_to_json(m), open('m.json','w'))
Save the fitted model as JSON.
m = model_from_json(open('m.json').read())
Reload without refitting.
# don't pickle Prophet modelsavoid
Pickle breaks across versions — use JSON.

★Prophet() defaultskey params

prophet cheat sheet