Related Experiment Video
Updated: Jul 28, 2026
11:18
Closed-loop Neuro-robotic Experiments to Test Computational Properties of Neuronal Networks
Published on: March 2, 2015
A comparative study of autoregressive neural network hybrids
Tugba Taskaya-Temizel1, Matthew C Casey
1University of Surrey, School of Electronics and Physical Sciences Department of Computing, Guildford, UK. t.taskaya@surrey.ac.uk
Summary
Combining forecasting models, like autoregressive integrated moving average (ARIMA) and neural networks, may not improve accuracy. This study shows hybrid models can underperform individual time series forecasting methods.
Area of Science:
- Time Series Analysis
- Machine Learning
- Econometrics
Background:
- Hybrid forecasting models, combining linear and non-linear approaches (e.g., ARIMA and neural networks), are often assumed to yield superior predictions.
- This assumption risks overlooking complex interdependencies between linear and non-linear components, potentially misapplying models to residuals.
Purpose of the Study:
- To challenge the prevailing assumption that combined forecasting models consistently outperform individual methods.
- To investigate the performance of hybrid models against their constituent forecasting techniques.
Main Methods:
- Empirical evaluation using nine diverse datasets.
- Comparison of hybrid models (ARIMA with time-delay neural networks) against individual autoregressive linear and neural network models.
Main Results:
- The study demonstrates that combined forecasts do not necessarily outperform individual forecasts.
- In several instances, hybrid models significantly underperformed compared to the performance of their individual components.
Conclusions:
- The efficacy of combining forecasting models is not guaranteed and requires careful validation.
- Hybrid forecasting architectures may not always be superior and can potentially lead to reduced predictive accuracy.
