prophet-forecasting

Business users need to know future demand—how many orders will arrive tomorrow, next week, next month. Traditional statistical methods (ARIMA, SARIMA) require manual parameter tuning, handle missing data poorly, and don’t automatically detect changepoints.

Facebook Prophet is a time-series forecasting library designed for business data with strong seasonal patterns. It automatically detects trend changes, models weekly/yearly seasonality, and handles missing data gracefully—unlike ARIMA which requires complete data series.

1. Training data: Aggregate daily counts (delivered orders) with dates as ds and counts as y
2. Model configuration:
   • changepoint_prior_scale: Controls trend flexibility (higher = more adaptive)
   • seasonality_mode="multiplicative": Seasonal patterns scale with trend level
   • Enable weekly and yearly seasonality
3. Custom regressors: Add is_weekend flag as external regressor
4. Forecast generation: m.make_future_dataframe(periods=30) creates future dates
5. Output: Point estimates (yhat) with confidence intervals (yhat_lower, yhat_upper)
6. Evaluation: MAPE, MAE, RMSE on in-sample fit

m = Prophet(
    changepoint_prior_scale=0.15,
    seasonality_mode="multiplicative",
    weekly_seasonality=True,
    yearly_seasonality=True,
)
m.add_regressor("is_weekend")
m.fit(df)
future = m.make_future_dataframe(periods=30)
forecast = m.predict(future)

• Automatic changepoint detection: Finds where trend changes significantly
• Handles missing data: Unlike ARIMA which requires complete series
• Multiplicative seasonality: Seasonal effects scale with trend—realistic for growing businesses
• Interpretable components: Trend, weekly seasonality, yearly seasonality
• Confidence intervals: Provides uncertainty bounds for decisions

• Related: Data Warehouse — forecasts stored in warehouse
• Related: ETL Pipeline — ML forecast runs after ETL
• Related: Apache Airflow — Airflow orchestrates forecast scheduling

• In-sample evaluation only: MAPE/MAE/RMSE on training data overestimates quality—need holdout set
• No baseline comparison: Can’t tell if Prophet beats naive forecast without comparison
• Limited regressors: Only is_weekend used—could benefit from holidays, promotions
• Single model: No model versioning or A/B testing