FORECASTING METHODS IN MACHINE LEARNING: REVIEW, COMPARISON, AND FUTURE PERSPECTIVES

SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 12 DECEMBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 92 FORECASTING METHODS IN MACHINE LEARNING: REVIEW, COMPARISON, AND FUTURE PERSPECTIVES S.U. Abdunabiev Tashkent International University of Education https://doi.org/10.5281/zenodo.17950454 Abstract. The given paper presents an analysis of modern forecasting methods in machine learning. The study aims to systematize existing approaches and identify their key characteristics. Four main categories of forecasting techniques are distinguished: traditional statistical approaches, classical machine learning algorithms, ensemble models, and deep learning methods. For each category, the most commonly used algorithms are examined, highlighting their respective strengths and limitations. Special attention is given to the comparative evaluation of methods according to criteria such as prediction accuracy, interpretability, computational complexity, and robustness to data noise. In addition, the paper discusses current trends in the development of forecasting technologies, including the growing adoption of explainable artificial intelligence (XAI), automation of model training processes (AutoML), and the integration of machine learning algorithms into real-time data processing systems. Keywords: forecasting, machine learning, deep learning, ensemble methods, explainable artificial intelligence, AutoML, time series, comparative analysis. 1 Introduction It is important to note that forecasting is one of the fundamental tasks of machine learning, playing a critically important role in a wide range of human activities. Its applications span economics, finance, technical systems, healthcare, transportation, energy, and many other areas where timely and well-founded decision-making is strategically significant [1][2][3]. Effective forecasting models not only allow for predicting future events and trends but also help optimize resource usage, minimize financial and operational risks, and increase the overall resilience of systems to uncertainty in the external environment. Certainly, modern forecasting methods represent a broad spectrum of approaches, which can be classified depending on the mathematical nature of the model, the structure of the data, the sample size, and the specific objectives of the analysis [4]. Traditional statistical methods, such as linear regression, ARIMA, and exponential smoothing, remain relevant due to their interpretability and mathematical rigor [5][6]. However, in recent years there has been a noticeable shift towards advanced machine learning and deep learning algorithms. These methods have a high capacity to model nonlinear dependencies, handle high-dimensional data, work with time series and spatiotemporal dependencies, and efficiently process large volumes of information with diverse structures [7][8]. The aim of this study is to systematize knowledge about modern forecasting methods, conduct a comparative analysis of their effectiveness, and identify promising directions for further development. The research considers the classification of existing forecasting methods and the analysis of their theoretical foundations, compares methods based on key practical characteristics SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 12 DECEMBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 93 such as accuracy, training speed, interpretability, and robustness to noise, and reviews current trends in forecasting methodology to identify prospective areas for future investigation. 2 Analysis and Results This study is based on a systematic approach, including classification, comparative analysis, and the examination of information from contemporary scientific and practical sources. During the analysis, modern forecasting methods are conventionally divided into four main categories: 3 Traditional Statistical Methods Definitely, traditional forecasting methods are based on classical statistical principles and mathematical models, which ensures high interpretability and relative ease of application. Among the most common approaches are autoregressive integrated moving average models (ARIMA), linear regression, and moving average (MA) models [9]. ARIMA is one of the well-known tools for time series forecasting. It combines autoregression, integration, and moving averages, which allows for effective handling of data with stationary properties, as well as accounting for short-term dependencies and seasonal fluctuations [10]. Linear regression is applied when there is a direct linear relationship between explanatory and dependent variables and is well-suited for forecasting when the data structure is relatively simple and predictable. Moving average models allow for smoothing out noise in the data and identifying general trends, making them convenient for short-term forecasting [11]. Although these methods require relatively small amounts of data and are easy to interpret, their capabilities are limited when working with high-dimensional data, complex nonlinear dependencies, and long-term patterns, which are characteristic of modern practical tasks in economics and technical systems. 2. Traditional Machine Learning Methods In turn, traditional machine learning algorithms allow modeling more complex and nonlinear dependencies, providing higher forecasting accuracy compared to classical statistical models, while remaining relatively interpretable and controllable. Among the most common methods, the following can be highlighted:  Decision Trees – a simple and visual model that builds predictions by sequentially splitting the data according to features. It is effective with small datasets and allows identifying the most significant variables influencing the outcome [12].  k-Nearest Neighbors (kNN) - an algorithm based on finding the closest objects in the feature space [13]. It is well-suited for forecasting based on the similarity of observations and is intuitive for interpreting results.  Support Vector Machines (SVM) allows modeling nonlinear dependencies using kernel functions. Applicable to both regression and classification tasks, especially when working with high-dimensional data due to its ability to identify complex structural relationships between features.  Naive Bayes - a probabilistic machine learning model based on Bayes’ theorem with the assumption of feature independence. It is characterized by high speed and simplicity of implementation, particularly effective for classification tasks involving categorical variables. These algorithms provide a reasonable balance between prediction accuracy and model interpretability, demonstrating relatively good performance when working with more heterogeneous and moderately large datasets. However, their application and effectiveness are mainly realized in addressing fundamental tasks and problems. SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 12 DECEMBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 94 4 Ensemble Machine Learning Methods The next logical step in the development of forecasting methods has been combined approaches, such as bagging, stacking, and tree ensembles. These methods aim to combine the strengths of different models to increase the accuracy and robustness of predictions. Combined approaches help minimize errors of individual models, providing more reliable forecasts through the aggregation of results from multiple algorithms to form a final output [14][15]. Thus, combined machine learning models open up broad opportunities for modeling complex nonlinear dependencies and analyzing large multidimensional datasets. Among the commonly used methods are Random Forest, which, thanks to its ensemble structure of decision trees, demonstrates high robustness to noise and outliers; gradient boosting (e.g., XGBoost and CatBoost), which provides high prediction accuracy through the stepwise correction of errors from previous models [16]; and stacking, which allows combining predictions from different models to create a meta-model for generating the final output [17]. The use of these methods allows for combining the advantages of individual algorithms; however, their effectiveness heavily depends on the correct selection and tuning of training hyperparameters, as well as the size and quality of the training data. In addition, using ensemble models increases computational complexity, which can complicate the development and deployment process. 5 Deep Learning Methods At the modern stage of forecasting, deep learning methods occupy a special place because they allow modeling complex nonlinear dependencies and analyzing large volumes of data with sequential or spatiotemporal structure. These methods are capable of detecting complex patterns that are inaccessible to traditional statistical and classical machine learning algorithms [18]. Among the most studied and well-known deep learning models are the following:  Recurrent Neural Networks (RNN, LSTM, GRU) – designed for working with time series and sequential data. LSTM and GRU modifications allow for effective long-term dependency memory and overcome the vanishing gradient problem characteristic of classical RNNs [19].  Convolutional Neural Networks (CNN) – capable of detecting local patterns and spatial dependencies in data. Their application is widespread not only in image processing but also in time series analysis, where local patterns are important [20]. Despite the obvious advantages of these methods, they require significant computational resources and large volumes of training data. Based on the comparative analysis of modern forecasting methods, the following patterns can be highlighted:  Traditional statistical methods demonstrate high interpretability and provide good results when working with relatively small datasets with clearly defined patterns; however, their capabilities are limited when modeling complex nonlinear dependencies and long-term correlations.  Machine learning algorithms are characterized by robustness to heterogeneous data and minimal preprocessing requirements, but they are sensitive to noise and require careful hyperparameter tuning to achieve optimal accuracy.  Deep learning methods, such as recurrent and convolutional neural networks, show high efficiency in analyzing large datasets with long-term and complex dependencies; however, SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 12 DECEMBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 95 their use is associated with high computational complexity and reduced interpretability of results, requiring professional expertise in model construction and application. 6 Modern Trends in Development Analysis of modern research allows identifying several key trends in the development of forecasting methods using machine learning algorithms. One of the most important trends is the development of explainable artificial intelligence (XAI), which increases the transparency of complex models and allows understanding which factors have the greatest influence on the forecast. The use of XAI tools, such as SHAP and LIME, makes model results more reliable and increases trust in them, which is especially important in high-risk areas such as finance and healthcare [21]. Another significant trend is the automation of model creation through AutoML platforms. These tools aim to simplify the selection, training, and tuning of models, making machine learning accessible to companies and professionals without deep expertise in data science [22]. AutoML also accelerates the development and deployment of forecasting models, allowing organizations to respond faster to changes and make data-driven decisions. There is also a trend toward the development of real-time forecasting methods and big data processing. As the volume of streaming information from IoT devices and other sources increases, there is growing interest in models capable of analyzing data and generating forecasts in real time. The use of big data from heterogeneous sources is considered a factor that contributes to improving model reliability and adaptability, enhancing prediction accuracy, and supporting decision-making in dynamic environments [23]. 7 Conclusion The analysis allows us to trace the diversity and development of forecasting methods in machine learning from traditional statistical approaches to modern deep learning methods. The analysis shows that different methods have their own advantages and limitations, which determine their areas of most effective application. Modern trends, such as the development of explainable AI, automated model building, and real-time data processing, reflect the movement toward more adaptive and transparent forecasting systems. REFERENCES 1. Lukashevich, M. N., & Kovalyov, M. Y. (2022). Machine learning models and methods for solving optimization and forecasting problems of the work of seaports. Informatics, 19(4), 94–110. https://doi.org/10.37661/1816-0301-2022-19-4-94-110. 2. Agafonov, A., & Myasnikov, V. (2014). An algorithm for city transport arrival time estimation using adaptive elementary predictions composition. Computer Optics, 38(2), 356–368. https://doi.org/10.18287/0134-2452-2014-38-2-356-368. 3. Strielkowski, W., Vlasov, A., Selivanov, K., Muraviev, K., & Shakhnov, V. (2023). Prospects and Challenges of the Machine Learning and Data-Driven Methods for the Predictive Analysis of Power Systems: A Review. Energies, 16(10), 4025. https://doi.org/10.3390/en16104025. 4. Klyuev, R. V., Morgoev, I. D., Morgoeva, A. D., Gavrina, O. A., Martyushev, N. V., Efremenkov, E. A., & Mengxu, Q. (2022). Methods of Forecasting Electric Energy Consumption: A Literature Review. Energies, 15(23), 8919. https://doi.org/10.3390/en15238919. 5. Kulikova, M. Kh., Khalidov, A. A., & Matasheva, Kh. P.. (2024). THE USE OF ECONOMIC AND MATHEMATICAL METHODS IN THE ANALYSIS OF THE ACTIVITIES OF SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 12 DECEMBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 96 ENTERPRISES. Economics and Entrepreneurship 1(162), 1469–1472. https://doi.org/10.34925/eip.2024.162.1.285. 6. Gorelik, A. Yu., & Korolyova, E. V. (2025). Empirical comparison of time series forecasting models: The case of common shares of PJSC “Sberbank.” Economics and Entrepreneurship, 7(180), 535–543. https://doi.org/10.34925/eip.2025.180.7.092. 7. Sotnikov, А.I. (2021). Modern technologies of deep learning for forecasting time series. Information-technological journal, 3(29), 95–105. https://doi.org/10.21499/2409-1650-29-395-105. 8. Karimova, Kh. I., Belozyor, A. O., & Dragulenko, V. V. (2023). APPLICATION OF DEEP LEARNING ALGORITHMS IN MICROECONOMIC ANALYSIS. Economics and Entrepreneurship, 10(159), 937–940. https://doi.org/10.34925/eip.2023.159.10.191. 9. Utlaut, T. L. (2008). Introduction to Time Series Analysis and Forecasting. Journal of Quality Technology, 40(4), 476–478. https://doi.org/10.1080/00224065.2008.11917751. 10. Le, L.-H. (2024). Time series analysis and applications in data analysis, forecasting and prediction. HPU2 Journal of Science: Natural Sciences and Technology, 3(1), 20–29. https://doi.org/10.56764/hpu2.jos.2024.3.1.20-29. 11. Svetunkov, I., & Petropoulos, F. (2017). Old dog, new tricks: a modelling view of simple moving averages. International Journal of Production Research, 56(18), 6034–6047. https://doi.org/10.1080/00207543.2017.1380326. 12. Dobra, A. (2009). Decision Trees. In Encyclopedia of Database Systems (pp. 769–769). Springer US. https://doi.org/10.1007/978-0-387-39940-9_553. 13. Friedman, J. H., Baskett, F., & Shustek, L. J. (1975). An Algorithm for Finding Nearest Neighbors. IEEE Transactions on Computers, C–24(10), 1000–1006. https://doi.org/10.1109/t-c.1975.224110. 14. Wu, H., & Levinson, D. (2021). The ensemble approach to forecasting: A review and synthesis. Transportation Research Part C: Emerging Technologies, 132, 103357. https://doi.org/10.1016/j.trc.2021.103357. 15. Mienye, I. D., & Sun, Y. (2022). A Survey of Ensemble Learning: Concepts, Algorithms, Applications, and Prospects. IEEE Access, 10, 99129–99149. https://doi.org/10.1109/access.2022.3207287. 16. Bentéjac, C., Csörgő, A., & Martínez-Muñoz, G. (2020). A comparative analysis of gradient boosting algorithms. Artificial Intelligence Review, 54(3), 1937–1967. https://doi.org/10.1007/s10462-020-09896-5. 17. Chatzimparmpas, A., Martins, R. M., Kucher, K., & Kerren, A. (2021). StackGenVis: Alignment of Data, Algorithms, and Models for Stacking Ensemble Learning Using Performance Metrics. IEEE Transactions on Visualization and Computer Graphics, 27(2), 1547–1557. https://doi.org/10.1109/tvcg.2020.3030352. 18. Wang, S., Cao, J., & Yu, P. S. (2022). Deep Learning for Spatio-Temporal Data Mining: A Survey. IEEE Transactions on Knowledge and Data Engineering, 34(8), 3681–3700. https://doi.org/10.1109/tkde.2020.3025580. 19. Waqas, M., & Humphries, U. W. (2024). A critical review of RNN and LSTM variants in hydrological time series predictions. MethodsX, 13, 102946. https://doi.org/10.1016/j.mex.2024.102946. 20. Zhao, B., Lu, H., Chen, S., Liu, J., & Wu, D. (2017). Convolutional neural networks for time series classification. Journal of Systems Engineering and Electronics, 28(1), 162–169. SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 12 DECEMBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 97 https://doi.org/10.21629/jsee.2017.01.18. 21. Kalasampath, K., Spoorthi, K. N., Sajeev, S., Kuppa, S. S., Ajay, K., & Maruthamuthu, A. (2025). A Literature Review on Applications of Explainable Artificial Intelligence (XAI). IEEE Access, 13, 41111–41140. https://doi.org/10.1109/access.2025.3546681. 22. Singh, V. K., & Joshi, K. (2022). Automated Machine Learning (AutoML): an overview of opportunities for application and research. Journal of Information Technology Case and Application Research, 24(2), 75–85. https://doi.org/10.1080/15228053.2022.2074585. 23. Almeida, A., Brás, S., Sargento, S., & Pinto, F. C. (2023). Time series big data: a survey on data stream frameworks, analysis and algorithms. Journal of Big Data, 10(1). https://doi.org/10.1186/s40537-023-00760-1.