Performance of Deep Learning Models on Imputed Time Series Data: A Simulation Study and Application to Leading Airline Companies' Stock Price

Abstract

In this study, the validity of imputation techniques for deep learning methods in time series analysis is investigated using datasets based on daily closing data in the stock market. Datasets of daily closing stock prices for Turkish Airlines, Deutsche Lufthansa AG, and Delta Airlines, as well as a simulated dataset, are used. LSTM, GRU, RNN, and Transformer models, which are deep learning models, are employed. The original dataset and datasets with 5%, 15% and 25% missing data are analyzed imputing Linear, Spline, Stineman, Mean and Random imputation techniques. The results show that model performance varies depending on the imputation technique and the rate of missing data. GRU and Transformer models are favored for their robustness and excellent performance. For handling missing data, using spline and Stineman imputations is advisable to maintain high model accuracy. This study emphasizes the usability of various imputation techniques and deep learning models in time series analysis. It assesses model performance using both MAPE and RMSE to gain a comprehensive understanding of predictive accuracy and reliability, aiming to guide future research by comparing these methods.