Flexible models of holiday lift

Holidays often have a strong influence on time series data. Popular methods for modeling holiday effects involve using indicator variables on official holiday dates or cyclical trigonometric indicators such as Fourier transforms. However, these can often fail to capture irregular variations common among holiday effects such as offset from the official date, sharpness, flatness, or asymmetry around the peak. We introduce a model of holiday effects as a function of continuous time and parameters determining form and intensity. We implement this model in Stan and show how it infers posterior distributions explaining observed and holdout data for simulated and real examples of various forms of holiday effects.