[论文解读] A Multi-Horizon Quantile Recurrent Forecaster

Ruofeng Wen, Kari Torkkola+2 more|arXiv (Cornell University)|Nov 29, 2017

Time Series Analysis and Forecasting参考文献 22被引用 257

一句话总结

提出 MQ-RNN，一种 Seq2Seq 框架，使用 forking-sequences 训练和针对不同 horizon 的上下文，产生时间序列的多地平线概率分位数预测，以处理静态/未来协变量和跨序列学习。

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.

研究动机与目标

• 开发一个通用的概率性多步时间序列回归框架。
• 将 Seq2Seq 神经网络与分位数回归以及 Direct Multi-Horizon forecasting 融合。
• 为编码器-解码器架构引入一个稳定的训练方案（forking-sequences）。
• 兼容时间变量与静态协变量、未来已知事件以及跨序列学习。
• 在大规模的 Amazon 需求数据和公开的电力预测任务上展示有效性。

提出的方法

• 使用基于 Seq2Seq 的编码器将历史摘要为隐藏状态。
• 通过全局 MLP 上下文和时域特定本地 MLP 产生 K×Q 的多地平线分位数预测输出。
• 通过最小化跨越 horizon 和 quantile 的分位损失之和来训练。
• 引入 forking-sequences 训练，在多个时间点创建解码器，参数共享以实现一次性多地平线学习。
• 通过 horizon-specific context 和全局 context 将已知未来信息和未来对齐特征纳入。
• 允许编码器扩展（如 NARX 风格跳跃、滞后输入、WaveNet 风格膨胀卷积），并使用静态/未来特征以提升性能。

实验结果

研究问题

• RQ1 Can Seq2Seq 架构被改造为直接输出时间序列的多地平线分位数预测？
• RQ2 Does a Direct Multi-Horizon approach with quantile regression provide more robust and calibrated probabilistic forecasts than recursive methods?
• RQ3 How can known future information and seasonal/event effects be aligned within a forecast framework?
• RQ4 Does forking-sequences training improve stability and efficiency for encoder-decoder time series models?
• RQ5 Can a single model learn across multiple related series and perform cold-start forecasting?

主要发现

• MQ-RNN achieves improved probabilistic forecast accuracy across horizons and quantiles compared to baselines.
• Forking-sequences training stabilizes learning and reduces training costs by sharing parameters across multiple forecast creation times.
• Incorporating horizon-specific context and a horizon-agnostic global context improves alignment of seasonality and events, yielding sharper forecast intervals.
• Alternative encoders (e.g., MQ-CNN with WaveNet-style dilations or lag-based inputs) can provide further performance gains.
• MQ-RNN variants outperform state-of-the-art on Amazon demand forecasting and GE FCom2014 electricity price/load forecasting tasks by leveraging multi-horizon quantile losses and known future information.
• Quantile forecasts (e.g., P10, P50, P90) provide useful calibration and sharpness across horizons.