PP: A multi-horizon quantile recurrent forecaster

2017 NIPS, time series workshop

traditional methods: ARIMA.

Seq2Seq

quantile forecast;

RELATED WORK

DeepAR, probabilistic forecasting with encoder-decoder models. A seq2seq architecture with an identical encoder and decoder. 

METHOD

为什么要用encoder-decoder 模型，是因为这个模型可以对分布进行拟合，是quantile regression model 的重要条件。

Loss: quantile loss

structure: encoder, lstm; decoder lstm.