•  
  •  
 

Abstract

Time series data often faces missing values that degrade the quality of analysis and prediction accuracy. This study focuses on the evaluation of the influence of various imputation methods on the performance of deep learning models in time series forecasting. The imputation methods tested included mean, median, mode, Multiple Imputation by Chained Equations (MICE), Last Observation Carried Forward (LOCF), and K-Nearest Neighbors (KNN), which were integrated with RNN, GRU, LSTM, and Bi-LSTM models. The research process includes missing value scenarios, data imputation, model training, and evaluation using MAPE, RMSE, and R², with statistical significance assessed using paired t-tests and hyperparameter optimization using Particle Swarm Optimization (PSO) as a supporting approach. The results showed that the imputation method had a significant effect on model performance (p < 0.05), where LOCF and mean imputation achieved competitive performance, with MAPE as low as 0.73011, RMSE of approximately 3.23223, and R² values above 0.92 in the evaluated experiments. However, LOCF was not consistently the best-performing method across all datasets and experimental scenarios, as Mode and MICE achieved competitive or better results for several metrics under artificial missingness. LSTMs showed the highest sensitivity to imputation variations (p ≈ 0.0001). At the same time, GRU and Bi-LSTM exhibited varying levels of stability depending on the dataset and experimental conditions. Data deletion generally degraded forecasting performance, whereas LOCF provided competitive results, potentially benefiting datasets with stronger temporal continuity or higher temporal fluctuation. The main contribution of this study is to present a comprehensive comparison of deep learning models, namely RNN, GRU, LSTM, and Bi-LSTM, on time series data with missing values. Furthermore, this study systematically analyzes the influence of various imputation methods on model performance using metric-based evaluation and statistical validation. Ultimately, this research aims to provide practical guidance for selecting appropriate imputation techniques and deep learning models based on data characteristics to improve the accuracy and reliability of forecasting.

First Page

55

Last Page

69

clean version-Artikel_KEDS_Rev 1.docx (347 kB)
Revision Manuscript without Highlight

Artikel_KEDS_Rev 1 (1).docx (349 kB)
Revision Manuscript with Highlight

Response to Editor Editing Comment ver 1.docx (42 kB)
Table of Response to Editor Editing Comment

Response to Editor Minor Revisi.docx (38 kB)
Table of Response to editor fo the minor revisions

Share

COinS