Prophet

Prophet is a forecasting tool developed by Facebook designed to handle time series data with strong seasonal effects and several seasons of historical data. It excels at automatically detecting holidays, trends, and multiple seasonality patterns with minimal manual tuning.

When to Use Prophet

Prophet is best suited for:

• Business metrics with daily observations spanning months or years
• Time series with strong seasonal patterns (daily, weekly, yearly)
• Data with multiple seasonality cycles (e.g., day-of-week + time-of-year)
• Series with holiday effects and special events
• When you have external regressors (additional features)
• Scenarios where you need robust handling of missing data and outliers

Strengths

• Automatic detection of yearly, weekly, and daily seasonality
• Handles missing data and outliers gracefully
• Intuitive parameters that map to business domain knowledge
• Built-in support for holidays and special events
• Provides uncertainty intervals with seasonality
• Works well with irregular time series (missing timestamps)
• Accepts additional regressors (exogenous variables)
• Fast fitting, even with large datasets
• Produces interpretable components (trend, seasonality, holidays)

Weaknesses

• Requires sufficient historical data (ideally 1+ years for yearly seasonality)
• May overfit on short time series (< 2 seasonal cycles)
• Not ideal for high-frequency data (sub-hourly) without aggregation
• Additive/multiplicative choice requires domain knowledge
• Can produce unrealistic extrapolations if trend parameters are misconfigured
• Limited to univariate forecasting (one target at a time)

Parameters

Common Time Series Parameters

All time series models share these parameters:

• Timestamp Column (required): Column containing dates/times
• Target Column (required): Numeric value to forecast
• Extra Regressors (optional): Additional feature columns (e.g., 'temperature', 'discount')
• Frequency (optional): Time spacing (D, H, W, M). Auto-inferred if not specified
• Forecast Steps (required, default=1): How many periods to predict
• Aggregation (default='mean'): Method for duplicate timestamps
• Fill Method (default='interpolate'): How to handle missing values
• Smoothing (default=false): Whether to apply smoothing
• Add Lags (optional): List of lag periods to use as features (e.g., [1, 7] for yesterday and last week)
• Add Rolling Mean (optional): Rolling mean window sizes (e.g., [7, 14] for 7-day and 14-day averages)

Prophet-Specific Parameters

Changepoint Prior Scale

• Type: Float
• Default: 0.05
• Min: 0.001
• Description: Controls trend flexibility. Higher values make the trend more flexible to fit sudden changes
• Tuning: Increase if trend changes are underfit; decrease if trend is overfitting noise
• Example: 0.5 for volatile trends, 0.01 for smooth trends

Seasonality Prior Scale

• Type: Float
• Default: 10.0
• Min: 0.01
• Description: Controls the flexibility of seasonality components. Higher values allow more variation in seasonal patterns
• Tuning: Increase for strong seasonal variation; decrease for subtle seasonality
• Example: 15.0 for highly seasonal data, 1.0 for weak seasonality

Seasonality Mode

• Type: String
• Default: 'additive'
• Options: ['additive', 'multiplicative']
• Description: How seasonality combines with trend
  • Additive: Seasonal effect is a constant addition (e.g., +10 units every December)
  • Multiplicative: Seasonal effect is proportional to trend (e.g., +10% every December)
• Guidance: Use multiplicative when seasonal variation grows with the trend level

Growth

• Type: String
• Default: 'linear'
• Options: ['linear', 'logistic', 'flat']
• Description: The type of growth trend
  • Linear: Unbounded growth/decline
  • Logistic: Growth with saturation (requires cap)
  • Flat: No trend
• Example: Use 'logistic' for market saturation scenarios

N Changepoints

• Type: Integer
• Default: 25
• Min: 0
• Description: Number of potential changepoints where trend rate can change. More changepoints = more flexible trend
• Guidance: Increase for very long or volatile histories (100+ for multi-year data)

Changepoint Range

• Type: Float
• Default: 0.8
• Min: 0.0
• Max: 1.0
• Description: Proportion of history in which trend changepoints will be estimated. 0.8 means changepoints only in first 80%
• Guidance: Prevents overfitting to recent noise by excluding the tail from changepoint detection

Uncertainty Interval

• Type: Float
• Default: 0.8
• Min: 0.0
• Max: 1.0
• Description: Width of the prediction interval (e.g., 0.95 for 95% confidence)
• Example: 0.95 for wide intervals, 0.5 for narrow intervals

Yearly Seasonality

• Type: String
• Default: 'auto'
• Options: ['auto', 'true', 'false']
• Description: Whether to model yearly seasonal patterns
  • auto: Enabled if data spans 2+ years
  • true: Force yearly seasonality
  • false: Disable yearly seasonality

Weekly Seasonality

• Type: String
• Default: 'auto'
• Options: ['auto', 'true', 'false']
• Description: Whether to model weekly seasonal patterns
  • auto: Enabled if data spans 2+ weeks
  • true: Force weekly seasonality
  • false: Disable weekly seasonality

Daily Seasonality

• Type: String
• Default: 'auto'
• Options: ['auto', 'true', 'false']
• Description: Whether to model daily seasonal patterns (intraday)
  • auto: Enabled if data has sub-daily frequency
  • true: Force daily seasonality
  • false: Disable daily seasonality

Hyperparameter Tuning Parameters

• Tune Hyperparameters (default=false): Enable automatic optimization
• Tuning Method (default='grid'): grid or random search
• Number of Trials (default=20): For random search
• CV Folds (default=3): Time series cross-validation folds
• Scoring Metric (default='rmse'): rmse, mae, or mape

Configuration Tips

Quick Start Configuration

For most business time series, start with:

- Seasonality Mode: 'additive'
- Growth: 'linear'
- Yearly Seasonality: 'auto'
- Weekly Seasonality: 'auto'
- Daily Seasonality: 'auto'

Handling Different Patterns

Strong Trend with Seasonality:

• Set changepoint_prior_scale=0.1 for flexible trend
• Use seasonality_mode='multiplicative' if variation grows with trend

Stable with Strong Seasonality:

• Keep default changepoint_prior_scale=0.05
• Increase seasonality_prior_scale=15.0

Volatile/Noisy Data:

• Decrease changepoint_prior_scale=0.01
• Decrease seasonality_prior_scale=5.0

Adding External Features

Use Extra Regressors for:

• Marketing spend
• Promotions/discounts
• Weather data
• Economic indicators

Prophet will automatically model their effects on the target.

Handling Growth Saturation

If your metric has a ceiling (e.g., market size):

1. Set growth='logistic'
2. You'll need to specify capacity in your data

Seasonal Period Selection

• Daily frequency: Enable weekly and yearly seasonality
• Hourly frequency: Enable daily, weekly, and yearly seasonality
• Monthly frequency: Only yearly seasonality makes sense

Common Issues and Solutions

Issue: Trend Overshoots Unrealistically

Solution:

• Decrease changepoint_prior_scale to 0.01 or 0.001
• Use growth='flat' if no trend expected
• Consider growth='logistic' with a cap for bounded growth

Issue: Missing Seasonal Patterns

Solution:

• Ensure you have 2+ complete seasonal cycles in training data
• Force seasonality: set yearly_seasonality='true' or weekly_seasonality='true'
• Increase seasonality_prior_scale to 15.0 or 20.0

Issue: Overfitting to Noise

Solution:

• Decrease changepoint_prior_scale to 0.01
• Decrease n_changepoints to 10-15
• Set changepoint_range=0.7 to exclude recent volatile data

Issue: Poor Performance on Short History

Solution:

• Prophet needs sufficient data (at least 2 seasonal cycles)
• For < 6 months of data, consider simpler models like ARIMA or Exponential Smoothing
• Disable seasonality that lacks sufficient history

Issue: Forecast Ignores Recent Trend Change

Solution:

• Increase changepoint_prior_scale to 0.1 or 0.5
• Increase changepoint_range to 0.9 to allow recent changepoints
• Add relevant external regressors that explain the change

Issue: Multiplicative vs Additive Seasonality

Solution:

• Use multiplicative if seasonal swings grow with the overall level (e.g., retail sales)
• Use additive if seasonal swings are roughly constant (e.g., temperature)
• Compare both modes using cross-validation metrics

Issue: Holidays Not Captured

Solution:

• Prophet has built-in holidays for many countries
• Ensure your timestamp column is in proper date format
• Add custom holidays using Prophet's API if needed

Example Use Cases

• Daily website traffic: Weekly + yearly seasonality, captures day-of-week and holiday effects
• Hourly energy consumption: Daily + weekly seasonality, handles intraday and weekly patterns
• Monthly retail sales: Yearly seasonality with promotional regressors
• App downloads with marketing: Trend + weekly seasonality + marketing spend as regressor
• Restaurant revenue: Weekly seasonality (weekend peaks) + holiday effects + weather regressors