[Paper Review] Comprehensive Analysis of Time Series Forecasting Using Neural Networks.

Manie Tadayon, Yumi Iwashita|arXiv (Cornell University)|Jan 27, 2020

Time Series Analysis and Forecasting21 references4 citations

TL;DR

This paper proposes hybrid neural network architectures that integrate dynamic and static measurements for time series forecasting, leveraging clustering to enhance prediction efficiency and accuracy. Results show that feature-based clustering outperforms distance-based clustering in speed and effectiveness, while adding more predictors does not always improve forecasting performance.

ABSTRACT

Time series forecasting has gained lots of attention recently; this is because many real-world phenomena can be modeled as time series. The massive volume of data and recent advancements in the processing power of the computers enable researchers to develop more sophisticated machine learning algorithms such as neural networks to forecast the time series data. In this paper, we propose various neural network architectures to forecast the time series data using the dynamic measurements; moreover, we introduce various architectures on how to combine static and dynamic measurements for forecasting. We also investigate the importance of performing techniques such as anomaly detection and clustering on forecasting accuracy. Our results indicate that clustering can improve the overall prediction time as well as improve the forecasting performance of the neural network. Furthermore, we show that feature-based clustering can outperform the distance-based clustering in terms of speed and efficiency. Finally, our results indicate that adding more predictors to forecast the target variable will not necessarily improve the forecasting accuracy.

Motivation & Objective

• To develop neural network architectures that effectively combine dynamic and static measurements for improved time series forecasting.
• To evaluate the impact of anomaly detection and clustering techniques on forecasting accuracy.
• To compare feature-based and distance-based clustering methods in terms of efficiency and performance.
• To investigate whether increasing the number of predictors consistently improves forecasting results.

Proposed method

• Proposes multiple neural network architectures designed to process both dynamic and static time series measurements.
• Integrates clustering techniques—specifically feature-based and distance-based clustering—into the forecasting pipeline to group similar time series.
• Applies anomaly detection to preprocess data and improve model robustness before forecasting.
• Employs a hybrid modeling approach where static features are embedded and combined with dynamic sequences in the neural network architecture.
• Uses dynamic measurements (e.g., sequential data points) as input to recurrent or sequential neural networks for temporal pattern learning.
• Evaluates model performance using standard forecasting metrics, comparing clustering variants and predictor counts.

Experimental results

Research questions

• RQ1 How do hybrid neural network architectures that combine dynamic and static measurements affect time series forecasting accuracy?
• RQ2 Does clustering of time series data improve forecasting performance and prediction speed?
• RQ3 Can feature-based clustering outperform distance-based clustering in terms of efficiency and accuracy?
• RQ4 Does increasing the number of predictors lead to better forecasting results, or is there a diminishing return?

Key findings

• Feature-based clustering significantly improves prediction speed and efficiency compared to distance-based clustering.
• Clustering enhances both forecasting accuracy and overall prediction time, demonstrating a dual benefit.
• The integration of static and dynamic measurements through neural networks leads to more robust forecasting models.
• Adding more predictors does not necessarily improve forecasting accuracy, indicating potential diminishing returns with increased complexity.
• Anomaly detection contributes to model robustness, though its direct impact on accuracy is not quantified in the results.
• Hybrid architectures combining static and dynamic inputs yield superior performance over models using only dynamic inputs.