Zaman serisi analizini etkili motif keşfi ile geliştirme ve motif keşfinin tahmin uygulamalarına entegre edilmesi

Time series motif discovery is a powerful technique for identifying recurring patterns in sequential data, offering valuable insights into diverse applications such as finance, healthcare, and climate science. This thesis presents an innovative framework for motif discovery in time series data, addressing key challenges in identifying recurring patterns of varying lengths and instances. The proposed methodology improves sensitivity and accuracy of motif detection by using Binary Integer Programming (BIP) for optimal motif selection and employing a heuristic algorithm for efficient pattern discovery, while integrating a dynamic feedback mechanism to refine parameters. Additionally, the study demonstrates how the discovered motifs can be applied to forecasting tasks by integrating them into a Transformer model pipeline augmented with motif embeddings, assessing whether motif-augmented predictions outperform simpler baselines. Extensive experiments on real-world financial datasets reveal that while the identified motifs can effectively capture recurring structures, incorporating them into advanced forecasting models does not necessarily lead to improved predictive performance. In particular, Transformers with or without motif embeddings underperform relative to basic econometric techniques. These findings underscore the complexity of leveraging motif information within deep learning frameworks for time series forecasting, suggesting that although motif discovery is valuable for understanding temporal patterns, more targeted modeling strategies or alternative machine learning approaches may be necessary to realize its full potential in predictive tasks.