Forecasting long duration floating holidays in online traffic

Abstract

Techniques for forecasting long duration floating holidays in online traffic are described. According to some embodiments, a machine learning service receives a request to train a time series forecast model on time series data of a user, receives an input for the time series forecast model that comprises a first feature weight that represents a first pivot day and a second feature weight that represents a second pivot day, performs a linear interpolation on the first feature weight and the second feature weight for a day between the first pivot day and the second pivot day to generate a linearly interpolated first weight of the first feature weight for a feature vector and a linearly interpolated second weight of the second feature weight for the feature vector, determines a first coefficient for the time series forecast model based at least in part on the time series data of the user, the linearly interpolated first weight of the first feature weight from the feature vector, and the linearly interpolated second weight of the second feature weight from the feature vector, generates, by the time series forecast model comprising the first coefficient, a prediction for a future day, and transmits the prediction to the user.