Prophet v1 Baseline
Prophet is a forecasting model primarily used for time series data, designed to be interpretable, easy to use, and resilient to missing data and trend changes.
Model Details
Model Description
- Developed by: GAIA | Nickel5
- Model type: Additive model for time series forecasting
- Parameters: Forecasting parameters, including changepoint flexibility, seasonalities, and uncertainty intervals
Uses
Prophet is specifically designed to handle daily observations with seasonality (such as weekly or yearly patterns), holidays, and sudden changes in trend. It is used extensively for forecasting business metrics, such as sales, social media activity, or server loads, and can be tuned to capture custom seasonal patterns.
Direct Use
To directly use Prophet, you need to install the library and load a time series DataFrame with at least two columns:
- ds: Date/time column
- y: Target value column (e.g., sales, Dst index)
The Forecast Includes:
- yhat: The predicted value
- yhat_lower: Lower bound of the uncertainty interval
- yhat_upper: Upper bound of the uncertainty interval
Prophet is especially useful when interpreting time series components, such as trends, weekly seasonality, and holidays. You can specify additional holidays and seasonalities based on domain knowledge.
Downstream Use
Prophet is a versatile model for a wide range of time series applications and can be extended for complex seasonality or custom holiday impacts.
Bias, Risks, and Limitations
While Prophet is designed to be interpretable and flexible, there are a few considerations:
- Trend sensitivity: The model may be overly responsive to short-term changes if not carefully configured with changepoint controls.
- Uncertainty intervals: The model's forecast uncertainty intervals are sometimes wider or narrower than expected, depending on seasonal variation.
- Seasonality assumptions: By default, Prophet assumes weekly and yearly seasonality, which may not be suitable for non-standard periodicities without customization.
Recommendations/Future Work
- Refine seasonalities: Enhance the model by adding custom seasonalities for cases with non-standard cycles.
- Improve changepoint selection: Incorporate domain knowledge to set more precise changepoints.
- Adaptation to irregular data: Extend Prophet’s robustness for handling irregularly spaced time series.
Training Data
Prophet is typically trained on historical time series data and performs well with daily or sub-daily data containing trends and seasonality. Training data should cover at least one full seasonal cycle to allow the model to learn recurring patterns.
Installation:
