TY  - JOUR
AU  - Contreras, Edimer Mahecha
PY  - 2026
DA  - 2026/06/11
TI  - Adaptive Risk Evaluation in FinTech Systems via Reinforcement-Based Continuous Policy Optimization
JO  - ICCK Journal of Software Engineering
T2  - ICCK Journal of Software Engineering
JF  - ICCK Journal of Software Engineering
VL  - 2
IS  - 2
SP  - 156
EP  - 168
DO  - 10.62762/JSE.2026.605759
UR  - https://www.icck.org/article/abs/JSE.2026.605759
KW  - reinforcement learning
KW  - continuous policy optimization
KW  - adaptive risk evaluation
KW  - FinTech software systems
KW  - sequential decision learning
KW  - credit risk modeling
AB  - The key feature of FinTech software systems is the ability to accurately assess risk in real time, making decisions on high-volume streams of information that are associated with very low latency and are robust to concept drift, and able to be updated without disrupting services. This paper addresses the problem of adaptive risk scoring using a reinforcement learning approach by modeling the risk evaluation problem as a continuous-action Markov Decision Process and continuously optimizing the policy via streaming transactional, behavioral events and outcome driven reward feedback. In addition to the learning algorithm, we also view ARL-CPO as a deployable software architecture that separates online learning from inference serving to enable a modular approach to integrating ARL into production risk pipelines, such as an inference microservice, which is wrapped around an asynchronous update loop, updating ARL models continuously without periodic batch retraining — a capability not available in the Random Forest, Gradient Boosting, or Transformer baselines. We assess the approach on the prediction of credit default and adaptive asset allocation in a big data dataset of 8.5 million credit records, generated in a custom FinTech environment simulator. The performance based on precision and F1 score, of ARL-CPO is compared with the baselines, and it outperforms them with 97.4% classification accuracy, 98.8% trend adaptation rate (responsiveness to distributional shifts), and 96.1% cumulative long-term performance index (normalized long-horizon reward). The findings show that reinforcement learning-based continuous policy updates is an achievable, adaptive element for real-time risk systems in FinTech under the evolution of market and user conditions.
SN  - 3069-1834
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Contreras2026Adaptive,
  author = {Edimer Mahecha Contreras},
  title = {Adaptive Risk Evaluation in FinTech Systems via Reinforcement-Based Continuous Policy Optimization},
  journal = {ICCK Journal of Software Engineering},
  year = {2026},
  volume = {2},
  number = {2},
  pages = {156-168},
  doi = {10.62762/JSE.2026.605759},
  url = {https://www.icck.org/article/abs/JSE.2026.605759},
  abstract = {The key feature of FinTech software systems is the ability to accurately assess risk in real time, making decisions on high-volume streams of information that are associated with very low latency and are robust to concept drift, and able to be updated without disrupting services. This paper addresses the problem of adaptive risk scoring using a reinforcement learning approach by modeling the risk evaluation problem as a continuous-action Markov Decision Process and continuously optimizing the policy via streaming transactional, behavioral events and outcome driven reward feedback. In addition to the learning algorithm, we also view ARL-CPO as a deployable software architecture that separates online learning from inference serving to enable a modular approach to integrating ARL into production risk pipelines, such as an inference microservice, which is wrapped around an asynchronous update loop, updating ARL models continuously without periodic batch retraining — a capability not available in the Random Forest, Gradient Boosting, or Transformer baselines. We assess the approach on the prediction of credit default and adaptive asset allocation in a big data dataset of 8.5 million credit records, generated in a custom FinTech environment simulator. The performance based on precision and F1 score, of ARL-CPO is compared with the baselines, and it outperforms them with 97.4\% classification accuracy, 98.8\% trend adaptation rate (responsiveness to distributional shifts), and 96.1\% cumulative long-term performance index (normalized long-horizon reward). The findings show that reinforcement learning-based continuous policy updates is an achievable, adaptive element for real-time risk systems in FinTech under the evolution of market and user conditions.},
  keywords = {reinforcement learning, continuous policy optimization, adaptive risk evaluation, FinTech software systems, sequential decision learning, credit risk modeling},
  issn = {3069-1834},
  publisher = {Institute of Central Computation and Knowledge}
}