OTT Platformlarındaki Canlı TV Kanalları İzlenmelerinin Zaman Serisi Tahmini

This study compares various time series models for forecasting live TV channel views on an OTT (Over-the-Top) streaming platform, based on volatile user behavior on digital platforms and increasing demand for live broadcasts. The study evaluates traditional statistical models (ARIMAX, SARIMAX, Exponential Smoothing), machine learning algorithms (Random Forest, Gradient Boosting, LightGBM, XGBoost), deep learning architectures (LSTM, GRU, RNN), and Transformer-based models such as FEDFormer, TimesNet, Autoformer. These models were tested on proprietary OTT platform data and benchmarked against publicly available datasets, including Wikipedia page views and weather data. A standardized sliding window approach was implemented to ensure consistent model training and evaluation across different input lengths and prediction horizons. Model performance was assessed using Mean Absolute Error (MAE), Mean Squared Error (MSE), and runtime. The findings reveal that Transformer-based models achieve the highest forecasting accuracy, particularly in dynamic and non-stationary environments, though at a significant computational cost. Tree-based models demonstrate an optimal balance between efficiency and accuracy, making them practical for real-time deployment. Traditional models remain highly effective for structured and seasonal datasets. These results can help in selecting forecasting models based on characteristics, system constraints, and application requirements, offering valuable insights for both researchers and industry professionals in the fields of time-series forecasting and OTT media services.