TY - GEN

T1 - Improving Time Series Forecasting by Applying the Sliding Window Approach

AU - Ндунги, Ребекка

AU - Блеканов, Иван Станиславович

PY - 2025

Y1 - 2025

N2 - The sliding window approach is a widely used machine learning method to improve forecasting accuracy by partitioning continuous time series into overlapping subsets. This segmentation helps models capture temporal patterns and adapt to data dynamics effectively. Particularly useful for nonstationary data, sliding windows highlight localized trends and reduce forecasting errors. This study explores sliding window applications, emphasizing their importance in time-based machine learning regression across fields as compared to traditional methods. This approach provides flexibility, enabling models to adjust to data changes. Selecting an appropriate window size is crucial for enhancing model accuracy and adaptability. LSTM, ARIMA, SARIMA, and Holt’s Model were employed on a dataset of 1,048,576 job-related records, with performance evaluated using MSE, RMSE, and MAE metrics. The window approach with LSTM demonstrated superior efficiency due to its ability to learn complex patterns and long-term dependencies, achieving an improvement of LSTM exhibited superior performance, with improvements reflected in the following error metrics: MAE: 0.248, 2.649 on MSE, and 0.162 on RMSE with a sliding window value of 30. Sliding window significantly contributes to developing reliable, adaptive forecasting models, particularly in deep learning for time-series predictions.

AB - The sliding window approach is a widely used machine learning method to improve forecasting accuracy by partitioning continuous time series into overlapping subsets. This segmentation helps models capture temporal patterns and adapt to data dynamics effectively. Particularly useful for nonstationary data, sliding windows highlight localized trends and reduce forecasting errors. This study explores sliding window applications, emphasizing their importance in time-based machine learning regression across fields as compared to traditional methods. This approach provides flexibility, enabling models to adjust to data changes. Selecting an appropriate window size is crucial for enhancing model accuracy and adaptability. LSTM, ARIMA, SARIMA, and Holt’s Model were employed on a dataset of 1,048,576 job-related records, with performance evaluated using MSE, RMSE, and MAE metrics. The window approach with LSTM demonstrated superior efficiency due to its ability to learn complex patterns and long-term dependencies, achieving an improvement of LSTM exhibited superior performance, with improvements reflected in the following error metrics: MAE: 0.248, 2.649 on MSE, and 0.162 on RMSE with a sliding window value of 30. Sliding window significantly contributes to developing reliable, adaptive forecasting models, particularly in deep learning for time-series predictions.

KW - IT Job Market

KW - Machine Learning

KW - Non-Stationary Data

KW - Sliding Window

KW - Time Series Forecasting

UR - https://link.springer.com/chapter/10.1007/978-3-031-98304-7_34

UR - https://www.mendeley.com/catalogue/1be1b293-9d60-3ce1-8da5-a77f98b5a4d8/

U2 - 10.1007/978-3-031-98304-7_34

DO - 10.1007/978-3-031-98304-7_34

M3 - Conference contribution

SN - 978-3-031-98303-0

T3 - Lecture Notes in Networks and Systems

SP - 294

EP - 302

BT - Intelligent and Fuzzy Systems (INFUS 2025)

PB - Springer Nature

Y2 - 29 July 2025 through 31 July 2025

ER -