Execution Time Optimization Through Feature and Temporal Reduction in Asset Pricing
CONCURRENCY COMPUTATION PRACTICE AND EXPERIENCE, cilt.38, sa.15, ss.1-26, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 38 Sayı: 15
- Basım Tarihi: 2026
- Doi Numarası: 10.1002/cpe.70881
- Dergi Adı: CONCURRENCY COMPUTATION PRACTICE AND EXPERIENCE
- Derginin Tarandığı İndeksler: Applied Science & Technology Source, Scopus, Technology Collection (ProQuest), Aerospace Database, Science Citation Index Expanded (SCI-EXPANDED), Compendex, INSPEC, zbMATH
- Sayfa Sayıları: ss.1-26
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Ankara Üniversitesi Adresli: Evet
Özet
High-dimensional financial machine learning (ML) pipelines are computational workloads as much as predictive models: theirpractical value depends on runtime, memory footprint, scalability, and the ability to retrain under resource constraints. Thispaper treats empirical asset pricing as a demanding real-world workload and proposes Cost-Aware Adaptive Optimization, usingXGBoost (extreme gradient boosting), as a systems-aware framework to improve the computational efficiency of gradient-boostedtree pipelines while preserving application-level predictive utility. The framework uses computational feedback; optimization isdriven by training time, prediction time, memory footprint, feature dimensionality, temporal training-window depth, model com-plexity, and marginal utility per unit of computation. We first profile 10 ML algorithms in a controlled high-performance computingsetting and decompose execution cost into training and prediction components. XGBoost is then selected as the optimization tar-get because it offers a strong cost – utility baseline on the asset-pricing workload. The proposed framework adaptively controlsfeature-dimensionality and temporal-depth choices, evaluates memory- and execution-aware trade-offs, and analyzes scalabilitythrough runtime growth, parallel speed-up, and parallel efficiency. Application-level utility measures, including out-of-sampleR2 , mean squared error (MSE), and the Sharpe ratio of decile-sorted long – short portfolios, are retained as application-level val-idation metrics rather than as the sole objective. The results show that meaningful computational savings can be obtained bytreating dimensionality, temporal depth, and model complexity as controllable workload variables. The study therefore contributesa computational feedback-driven optimization framework for scalable financial ML execution, framing workload adaptation andconfiguration selection as runtime-, memory-, and utility-aware system problems.