A Multi-Horizon Quantile Recurrent Forecaster
This work addresses the need for accurate probabilistic forecasting in domains like e-commerce and energy, though it appears incremental as it builds on existing methods like sequence-to-sequence networks and quantile regression.
The authors tackled the problem of probabilistic multi-step time series forecasting by developing a framework that combines sequence-to-sequence neural networks, quantile regression, and direct multi-horizon forecasting, with a new forking-sequences training scheme to enhance stability and performance. They demonstrated its effectiveness in predicting demand on Amazon.com and in public competitions for electricity price and load forecasting.
We propose a framework for general probabilistic multi-step time series regression. Specifically, we exploit the expressiveness and temporal nature of Sequence-to-Sequence Neural Networks (e.g. recurrent and convolutional structures), the nonparametric nature of Quantile Regression and the efficiency of Direct Multi-Horizon Forecasting. A new training scheme, *forking-sequences*, is designed for sequential nets to boost stability and performance. We show that the approach accommodates both temporal and static covariates, learning across multiple related series, shifting seasonality, future planned event spikes and cold-starts in real life large-scale forecasting. The performance of the framework is demonstrated in an application to predict the future demand of items sold on Amazon.com, and in a public probabilistic forecasting competition to predict electricity price and load.