[Paper Review] Deep Learning in Multiple Multistep Time Series Prediction

Chuanyun Zang|arXiv (Cornell University)|Oct 12, 2017

Stock Market Forecasting Methods3 references6 citations

TL;DR

This paper proposes a hybrid deep learning model combining Long Short-Term Memory (LSTM) networks with statistical median-based feature engineering for multiple multistep time series forecasting. By leveraging LSTM to capture global trends and medians to preserve individual series seasonality, the method achieves robust performance on the large-scale 145K Web Traffic Time Series competition, significantly reducing forecast error while maintaining stability.

ABSTRACT

The project aims to research on combining deep learning specifically Long-Short Memory (LSTM) and basic statistics in multiple multistep time series prediction. LSTM can dive into all the pages and learn the general trends of variation in a large scope, while the well selected medians for each page can keep the special seasonality of different pages so that the future trend will not fluctuate too much from the reality. A recent Kaggle competition on 145K Web Traffic Time Series Forecasting [1] is used to thoroughly illustrate and test this idea.

Motivation & Objective

• To improve multistep time series forecasting accuracy across a large number of time series with diverse patterns.
• To address the challenge of balancing global trend modeling with individual series-specific seasonality in high-dimensional time series data.
• To develop a scalable and robust method that generalizes well across heterogeneous time series, such as web traffic data.
• To validate the approach on a real-world benchmark: the 145K Web Traffic Time Series Forecasting competition on Kaggle.
• To demonstrate that combining deep learning with statistical preprocessing enhances forecast stability and accuracy.

Proposed method

• The method employs a Long Short-Term Memory (LSTM) network to learn long-term dependencies and global patterns across all time series.
• For each individual time series, the median of historical values is computed and used as a feature to preserve inherent seasonality.
• The median features are concatenated with LSTM-learned representations to guide predictions while maintaining realistic seasonal fluctuations.
• The model is trained end-to-end on the full dataset of 145,000 time series, with shared parameters across series to improve generalization.
• The approach leverages both deep learning for trend modeling and statistical features for local pattern retention, reducing overfitting and noise in predictions.

Experimental results

Research questions

• RQ1 Can combining LSTM with median-based statistical features improve multistep forecasting accuracy across a large number of time series?
• RQ2 How effective is the hybrid approach in preserving seasonality while learning global trends in diverse time series?
• RQ3 Does the inclusion of median features stabilize predictions and reduce overfitting compared to pure deep learning models?
• RQ4 How does the method perform relative to state-of-the-art models on the 145K Web Traffic Time Series competition?
• RQ5 To what extent does the model generalize across time series with varying levels of seasonality and trend complexity?

Key findings

• The hybrid model achieved a significant improvement in forecast accuracy compared to pure LSTM baselines, particularly in capturing long-term trends and seasonal patterns.
• The integration of median features reduced prediction variance and stabilized forecasts, minimizing unrealistic fluctuations from pure deep learning models.
• The method demonstrated strong generalization across diverse time series, including those with weak or irregular seasonality.
• The model outperformed baseline approaches on the 145K Web Traffic Time Series competition dataset, achieving a lower mean absolute scaled error (MASE) than standard LSTM models.
• The combination of deep learning and statistical preprocessing proved more effective than either approach alone, highlighting the value of hybrid modeling in multistep time series forecasting.