Abstract Modeling the dynamics of global debt markets requires frameworks capable of capturing nonlinearity, nonstationarity, and structural regime shifts inherent in macro-financial time series. This study presents a rigorous comparative analysis of linear econometric and nonlinear deep learning models, specifically ARIMA, SARIMA, and Long Short-Term Memory (LSTM) networks, to evaluate their effectiveness in forecasting global debt issuance. Using a quarterly dataset from the Bank for... Modeling the dynamics of global debt markets requires frameworks capable of capturing nonlinearity, nonstationarity, and structural regime shifts inherent in macro-financial time series. This study presents a rigorous comparative analysis of linear econometric and nonlinear deep learning models, specifically ARIMA, SARIMA, and Long Short-Term Memory (LSTM) networks, to evaluate their effectiveness in forecasting global debt issuance. Using a quarterly dataset from the Bank for International Settlements spanning 1962- 2025, the analysis integrates comprehensive preprocessing, including stationarity transformations, seasonal decomposition, and feature scaling, to ensure methodological robustness. Empirical results reveal that global debt issuance exhibits strong persistence, pronounced quarterly seasonality, and nonlinear responses to macroeconomic shocks, particularly during the Global Financial Crisis and the COVID-19 period. While ARIMA adequately captures linear dependencies under stationarity assumptions, it fails to model seasonal and nonlinear structures. SARIMA improves performance by explicitly incorporating seasonal autoregressive and moving average components, yielding statistically robust residual diagnostics consistent with white-noise processes. However, LSTM significantly outperforms both models by learning complex temporal dependencies and nonlinear interactions without requiring explicit model specification. Model evaluation using MAE, RMSE, MSE, and MAPE consistently demonstrates the superior predictive accuracy and adaptability of LSTM, particularly under structural instability and regime shifts. The findings underscore the limitations of purely linear frameworks and highlight the efficacy of deep learning architectures in financial time-series forecasting. Furthermore, the results support the adoption of hybrid modeling strategies that integrate econometric interpretability with machine learning flexibility, offering a robust paradigm for analyzing and forecasting complex global debt dynamics. Read More Read Less

Abstract Code duplication is a pervasive phenomenon in software repositories that poses significant risks for both software quality and machine learning evaluation. This study presents a comprehensive analysis of code duplication, data leakage, and clone detection within large-scale Python corpora, focusing on the widely used py150 benchmark and its declbodies splits. Using a duplication index, we identify 7,336 duplicate groups comprising 17,033 entries, with an... Code duplication is a pervasive phenomenon in software repositories that poses significant risks for both software quality and machine learning evaluation. This study presents a comprehensive analysis of code duplication, data leakage, and clone detection within large-scale Python corpora, focusing on the widely used py150 benchmark and its declbodies splits. Using a duplication index, we identify 7,336 duplicate groups comprising 17,033 entries, with an average cluster size of 2.3. Critically, we detect 575 leakage groups (7.9%) spanning training, validation, and test splits, including 280 test samples and 303 validation samples that appear verbatim in the training data a violation of dataset independence that can inflate performance metrics through memorization. To address this, we implement a deduplication strategy that eliminates all cross-split leakage. We then evaluate binary and multi-class clone detection using traditional models (SVM, Random Forest) and transformer-based architectures (CodeBERT). Results show that transformer models achieve strong performance (AUC up to 0.97) on syntactic clones (Types I and II) but exhibit systematic degradation on semantic clones (Type IV, F1-score 0.72). Statistical tests confirm that observed differences are significant (p < 0.01). Our findings underscore that uncontrolled redundancy compromises experimental validity, and we advocate for routine duplication indexing and leakage quantification as essential preprocessing steps for reliable benchmarking of code intelligence systems. Read More Read Less

Abstract Real-world datasets increasingly encompass heterogeneous modalities, yet conventional fusion techniques often struggle with varying data reliability, missing inputs, and dynamic cross-modal interactions. To address these limitations, this study proposes an integrated multi-model learning framework that synergistically combines transformer, tree, and sequential based architectures. Central to this approach is the AdaptiveFusion module, which dynamically estimates modality reliability, facilitates cross modal attention, and aggregates features adaptively via... Real-world datasets increasingly encompass heterogeneous modalities, yet conventional fusion techniques often struggle with varying data reliability, missing inputs, and dynamic cross-modal interactions. To address these limitations, this study proposes an integrated multi-model learning framework that synergistically combines transformer, tree, and sequential based architectures. Central to this approach is the AdaptiveFusion module, which dynamically estimates modality reliability, facilitates cross modal attention, and aggregates features adaptively via a learnable gating mechanism. The framework operationalizes this architecture by integrating BERT for deep textual semantics, XGBoost for structured numerical temporal prediction, and LSTM for sequential dependency modeling. Empirical validation is conducted on a curated dataset tracking the chronological evolution of conversational AI systems, comprising categorical, temporal, numerical, and textual attributes. Experimental results demonstrate that each component effectively captures its designated data dimension: BERT achieves rapid semantic convergence, XGBoost delivers robust structured classification with high F1-scores (>0.80), and LSTM successfully identifies latent temporal trajectories. When integrated, these paradigms overcome individual limitations, yielding a holistic analytical system that enhances both predictive accuracy and interpretability. Despite constraints related to dataset scale and class imbalance, the proposed framework establishes a scalable, context-aware methodology for multimodal data fusion. This work underscores the critical value of hybrid modeling strategies in extracting comprehensive insights from complex, real-world information ecosystems. Read More Read Less