Jul 2026· Big Data and Cognitive Computing· 0 citations· 56 references
TL;DR
This work proposes a tenure-stratified hybrid framework that couples an online weight-of-evidence logistic regression (WOE-LR) scorecard with an offline self-attention stacked ensemble whose calibrated PD is quantile-binned, WOE-encoded, and re-injected into the online scorecard as a single auditable predictor.
Abstract
Real-time retail credit scoring is a data-intensive cognitive computing task. Each decision must fuse heterogeneous signals, execute a non-linear model, return a calibrated probability of default (PD), and emit a regulator-compliant local explanation within milliseconds. We address the most demanding segment of unsecured lending in Kazakhstan—Salary-Project-Independent (SPI) borrowers, whose principal income stream is not observable by the lender—and frame scoring as a constrained optimisation problem where we maximise discrimination subject to interpretability, latency, and calibration constraints. We propose a tenure-stratified hybrid framework that couples (i) an online weight-of-evidence logistic regression (WOE-LR) scorecard with (ii) an offline self-attention stacked ensemble (LightGBM, CatBoost, and a tabular self-attention network) whose calibrated PD is quantile-binned, WOE-encoded, and re-injected into the online scorecard as a single auditable predictor. On 551,962 production contracts that originated in 2022–2024, the repeat-client hybrid attains an area under the receiver operating characteristic curve (AUROC) of 0.826, a Gini coefficient of 0.65, and a Kolmogorov–Smirnov (KS) statistic of 0.495, preserving roughly half of the offline ensemble’s lift over the linear baseline (AUROC 0.79→0.897) while retaining a fully auditable twelve-coefficient scorecard in production. The new-client scorecard attains an AUROC of 0.741. Non-parametric isotonic recalibration reduces the expected calibration error from 0.27 to below 0.01 and raises the Hosmer–Lemeshow p-value above 0.99 without altering discrimination. The framework complies with the model risk standards of the Agency of the Republic of Kazakhstan for Regulation and Development of the Financial Market and is delivered as a Spark/MLOps reference architecture, illustrating how big data engineering, attention-based representation learning, and post hoc explanations can be co-designed for a high-stakes, high-throughput, regulated AI application.
Payroll-anchored retail borrowers—individuals whose monthly remuneration is routed into an account at the lending institution through a salary-project arrangement—constitute the volume backbone of unsecured consumer lending in Kazakhstan, generating the largest origination flow, the lowest realized default rate, and the majority of the systemic regulatory and capital sensitivities of second-tier banks. Payroll anchoring also changes the lender’s information set, which motivates a study of how that advantage translates into model performance and borrower outcomes. We design and internally validate an explainable hybrid artificial-intelligence framework stratified by client tenure into two production models: a Weight-of-Evidence (WOE) logistic-regression scorecard for new salary-project applicants, and a hybrid scorecard for repeat applicants, in which a stacked ensemble of LightGBM, CatBoost and a multi-head self-attention neural network contributes a single WOE-encoded predictor to a second-stage L2-regularized logistic regression. The hybrid recovers a substantial share of the ensemble’s discriminatory lift while preserving an auditable, monotone scorecard at the point of decision, and isotonic recalibration restores the predicted probabilities of default to the empirical bad-rate scale required for IFRS 9 expected-credit-loss accrual and risk-based pricing. We report discrimination, calibration and stability evidence under a strict anti-leakage protocol and set out the structural preconditions under which the architecture transfers to other emerging-market payroll-anchored portfolios. We are explicit about scope: a true out-of-time validation and a full group-conditional fairness audit are identified as required next steps rather than claimed here. The contribution is a reproducible, interpretable scoring design that exploits payroll visibility while retaining full coefficient interpretability inside the production decision engine.
Gulnaz Zakariya, A. Moldagulova, Nor’ashikin Ali· Applied Informatics· 0 citations
CreditR1 delivers calibrated PDs with evidence-grounded reasoning that supports internal model validation and human review that supports transferability beyond the Chinese A-share market remains an open empirical question.
SME credit files arrive incomplete, and the gaps leave room for anchoring, confirmation bias and loss aversion. We compare human, algorithmic and hybrid credit decisions using a benchmark credit dataset alongside a vignette experiment with credit analysts working in Morocco's Souss-Massa region. The modelling arm pairs L2-regularised logistic regression with gradient-boosted trees, adding stratified validation, calibration analysis, SHAP and LIME. In the human arm, matched cases vary the requested amount while everything else is held constant. Analysts were least stable on borderline files, and their decisions moved with the anchor. The boosted model held steadier but leaned harder on indicators that track how thick a file is. AI-first assistance improved consistency and deepened deference to the model; human-first assistance preserved contextual overrides; explanation-gating struck the best balance, though only where SHAP and LIME agreed. We assess distribution through demographic-parity difference, disparate-impact ratio, equal-opportunity difference and false-positive-rate difference. What the results support is a governed hybrid: weak explanations withheld, overrides auditable, human review genuinely available. A regional sample and benchmark data bound how far any of these travels.
Hassan Ennaqui, Mohamed El Bourki, Abdellah Bakrim et al.· EPJ Web of Conferences· 0 citations
Financial prediction typically relies on task-specific regression, ranking, or policy heads, separating the language model from the numerical object ultimately evaluated. We investigate whether a causal language model can instead represent forecasts and decisions directly through constrained token generation. FinATOM introduces a unified, head-free interface for three-step stock-return forecasting and dynamic five-ETF allocation. The forecasting model autoregressively emits volatility-standardized return tokens and is trained with ordinal and ranking supervision followed by a one-epoch token-level policy stage. The allocation model generates normalized long-only weights; supervised fine-tuning imitates a causal mean--variance anchor, and DAPO-augmented GRPO optimizes realized 21-day Sharpe subject to anchor consistency. In 2023--2025 ETF tests, the allocation policy improves pooled gross Sharpe from 1.428 to 1.529 and net Sharpe under a 5-bp transaction-cost model from 1.394 to 1.494. The multimodal allocation input attains the highest three-period mean Sharpe of 1.540, with its clearest advantage in 2025. On FinTexTS, the SFT and policy strategies achieve 73.52\%/2.68 and 73.72\%/2.69 cumulative-return/Sharpe, respectively. These results support the feasibility of direct language-model token generation for financial numerical prediction and decision-making, while motivating broader tests across assets, regimes, and random seeds.
The credit risk assessment remains a critical challenge for financial institutions as manual and semi-automated loan evaluation processes often produce inconsistent decisions, high default rates, and operational inefficiencies. This paper discusses an Integrated Data-Driven Loan Management Framework that comprises an Artificial Neural Network (ANN) for credit default prediction, Structured Query Language (SQL) to systematically extract and transform data, and Interactive Visual Analytics Dashboards to aid in providing transparency within the decision-making process. For the study, a public loan dataset was used to pre-process the target dataset containing a total of 38,577 records and 24 attributes with pre-processing techniques such as feature engineering, one-hot encoding and class re-balancing with the use of Synthetic Minority Oversampling Technique (SMOTE) yielding 43 model-ready inputs.
The final ANN architecture was developed with 2 hidden layers (43 and 21 neurons with ReLU activation) and a sigmoid output neuron for binary classification. SHapley Additive exPlanations (SHAP) were computed to provide interpretability for each prediction made by the model. The overall accuracy of the model is 87% and the weighted precision, recall, and F1-score are 0.87 for the held-out test set (n = 12,858). The framework is implemented as a web-based decision support tool using Flask that enables users to receive risk scores in real-time along with explainable outputs and visual dashboards. The results from this experimentation indicate that an integrated pipeline (including data querying, predictive modeling, interpretability, and visualization) provides better decision-making and stakeholder transparency than using a model in isolation; therefore, it serves as a proof-of-concept prototype for intelligent loan management in banks and other financial institutions.
Ponsak .S. Bande, Chimuzuoroke E. Ugwuja, Blessing .C. Uzo et al.· International Journal of Lat...· 0 citations
Standard credit-risk scorecards rely on linear ratio thresholds that break down when feature interactions are nonlinear and observations carry temporal dependencies. Qualitative signals embedded in corporate disclosures—tone shifts, forward-looking hedges, and sector-specific terminology—remain largely ignored by numeric-only models, even though such signals often precede ratio deterioration. This paper introduces a tri-modal deep learning framework that jointly trains three complementary branches: a Convolutional Neural Network (CNN) for cross-sectional ratio-pattern detection, a Long Short-Term Memory (LSTM) network for multi-quarter trend modelling, and a Natural Language Processing (NLP) branch for disclosure-text encoding. Prior to deep-model training, LASSO regularisation removes collinear financial indicators and SMOTE oversampling corrects the severe class imbalance characteristic of distress datasets. A feature-concatenation fusion layer integrates all three branch outputs; the resulting vector feeds a sigmoid classifier that produces a calibrated distress probability. Benchmarked against five baselines on four financial datasets, the model reaches 94.8% accuracy and 91.3% minority-class recall, with a 4.1-point F1 advantage over the strongest single-modality competitor.
Paiinti Meenakshi, Muthaluru Bhuvaneshwari, Jalla Ganesh et al.· International Conference Com...· 0 citations