A Multi-Horizon Quantile Recurrent Forecaster

TL;DR

This paper introduces a versatile probabilistic multi-step time series forecasting framework using Sequence-to-Sequence neural networks, quantile regression, and a novel training scheme, demonstrating effectiveness in real-world demand and electricity forecasting tasks.

Contribution

It presents a new multi-horizon forecasting approach combining neural networks, quantile regression, and a novel training scheme for improved stability and flexibility.

Findings

Effective in large-scale demand forecasting at Amazon.com

Performs well in electricity price and load prediction competitions

Accommodates various covariates and seasonal shifts

Abstract

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.