The research, conducted by Feng Shen and Jia Kuang of the School of Finance at Southwestern University of Finance and Economics in Chengdu, together with Dao Lan of the university’s School of Business Administration and Run Wang of Southwest University of Science and Technology, introduces an enhanced cost-sensitive Boosting model that replaces binary credit approval with a three-way decision framework powered by incremental learning. Rather than classifying every applicant as either good or bad credit risk, the model assigns cases to three categories: accept, reject, or defer for further evaluation. The middle category—delayed commitment—is not a failure of the algorithm but a deliberate design choice, one rooted in a decision theory known as three-way decision, or 3WD.

Three-way decision theory, first articulated by Yiyu Yao in the context of rough set theory in 2009, formalizes a simple human instinct: when the evidence is insufficient, postpone the decision rather than guess. In the credit context, this means that applicants whose predicted default probability falls within an uncertain boundary region are neither approved nor rejected outright. Instead, they are flagged for additional scrutiny—perhaps more documentation, a phone call, a manual review, or a second look once new information arrives. The economics are compelling. Approving a customer who later defaults typically costs far more than rejecting one who would have repaid, and rejecting a good customer carries lost-revenue costs of its own. A model that forces a premature commitment on borderline cases can rack up enormous expected losses precisely where uncertainty is greatest.

What distinguishes the new work from earlier applications of three-way decisions to credit risk is how the model handles change. Existing 3WD approaches, the authors note, suffer from three structural limitations: they rely on static datasets, they optimize their decision thresholds offline in a one-shot process, and their threshold learning is driven by classification accuracy rather than by money. A threshold pair—the upper and lower probability boundaries that define the positive, negative, and boundary regions—is typically calibrated once, on historical data, and then frozen. But credit markets are anything but frozen. Borrower populations shift, macroeconomic conditions evolve, and the distribution of defaults drifts over time. A threshold that was optimal last year may be quietly bleeding money today.

The proposed model confronts this problem by making threshold adaptation an ongoing, endogenous process. As credit data arrive sequentially—new applications, newly observed repayment outcomes, updated bureau records—the model dynamically updates the boundary thresholds that separate the three decision regions. At the heart of this mechanism is a dynamic cost matrix that incorporates not only the familiar misclassification costs, such as the loss from lending to a defaulter, but also evaluation costs: the expense of gathering additional information on deferred cases. This means threshold learning is guided by expected economic loss rather than raw predictive accuracy. In effect, the model asks not “How often is this boundary right?” but “What does this boundary cost the lender?”—a reframing that aligns the machine learning objective with the financial reality of lending.

The second pillar of the approach is incremental learning, embedded not as an add-on but as a core mechanism for boundary-sample re-evaluation. Boosting models, which build strong classifiers from ensembles of weak learners by iteratively focusing on hard-to-classify examples, are natural workhorses of credit scoring. The new model enhances the classic Boosting framework by continuously integrating three streams of samples: newly arrived cases, previously misclassified cases, and cases drawn from the boundary region where the model expressed the least confidence. This creates a feedback loop with real conceptual elegance. The three-way decision layer identifies uncertain cases; those uncertain cases, along with their eventual outcomes, feed back into the incremental learner; and the improved learner, in turn, sharpens the thresholds that define the boundary itself. The model, in other words, learns best exactly where it is least sure.

This is more than an academic nicety. Boundary-region samples are, in a statistical sense, the most informative examples a classifier can encounter—they sit near the decision surface where the model’s knowledge is weakest, and each new outcome for such a case carries maximal learning value. By prioritizing these samples for re-evaluation and retraining, the approach concentrates computational and analytical attention where it yields the greatest marginal improvement. It also mirrors how experienced credit officers behave in practice: they do not re-examine every approved loan with equal vigor, but they do watch the marginal applicants closely, updating their intuitions as evidence accumulates.

The technical architecture draws on a rich lineage. Boosting itself traces back to Jerome Friedman’s gradient boosting machine, published in the Annals of Statistics in 2001, and to scalable implementations such as XGBoost, introduced by Tianqi Chen and Carlos Guestrin in 2016. The cost-sensitive strand of the new model builds on a substantial literature showing that credit and fraud detection models should weigh errors asymmetrically—for instance, work on cost-sensitive support vector machines and on thresholding for fraud detection. The three-way decision strand draws on Yiyu Yao’s foundational formulation and on subsequent work on sequential three-way decisions and multilevel incremental processing. What the new study contributes is the synthesis: a Boosting ensemble whose decision layer is ternary, whose thresholds are learned dynamically under a full economic cost model, and whose training regime is incremental by design.

The authors also address a concern that has become central to the deployment of machine learning in regulated financial settings: interpretability. A model that defers judgment on some applicants must be able to explain why, both to comply with fair-lending regulations and to give human reviewers something to work with. The study includes an interpretability analysis that reveals the key drivers of the model’s decisions, exposing which features push an applicant into the accept region, the reject region, or the uncertain middle. This transparency is doubly important for deferred cases, because the entire premise of the boundary region is that something specific—some combination of borderline income, thin credit history, or mixed repayment signals—makes the case genuinely ambiguous. Being able to articulate that ambiguity turns the third decision from a black-box shrug into an actionable instruction.

The practical implications for lenders could be substantial. Traditional two-stage scoring pipelines and reject-inference frameworks have long struggled with the problem of unobserved outcomes: institutions never learn what would have happened to rejected applicants, so the data used to retrain scoring models is systematically censored. Three-way decision structures attack this problem directly by shrinking the rejected pool and channeling uncertain applicants into a monitored middle stream, where their eventual outcomes become training data. Combined with incremental learning, this means the model’s understanding of the risky middle of the distribution improves continuously rather than waiting for periodic offline recalibration. For lenders operating in fast-moving environments—digital consumer lending, small and medium enterprise credit, fintech platforms where applicant volumes are high and conditions change quickly—that agility translates directly into reduced expected loss.

The work also fits into a broader shift in how the machine learning community thinks about uncertainty. For decades, the dominant paradigm measured success by accuracy, area under the curve, and related statistical metrics. But in decision-critical domains, the true currency is expected cost, and a model that is slightly less accurate overall but dramatically better at handling the uncertain middle can be far more valuable. The new study’s dynamic cost matrix, which prices both misclassification and the cost of further evaluation, is a concrete embodiment of that philosophy. It suggests a future in which credit models do not merely predict, but negotiate—the algorithm deciding not just what it believes, but whether it is worth spending resources to find out.

The research was supported by the National Natural Science Foundation of China, the National Social Science Fund of China, and institutional programs at Southwestern University of Finance and Economics, including its Fintech Innovation Center and the Engineering Research Center of Intelligent Finance under the Ministry of Education. The authors report no competing interests, and data supporting the findings are available upon reasonable request.

As artificial intelligence systems are entrusted with ever more consequential judgments—loans, insurance, hiring, medical triage—the humble act of saying “not yet” may prove to be one of the most powerful tools in the algorithmic toolbox. This study shows that in credit scoring, at least, teaching machines to recognize the limits of their own confidence, to defer where deferral is cheap and commit where commitment is warranted, and to keep learning from every uncertain case they encounter, can turn the messiest part of the decision space into a source of competitive advantage. The gray zone, it turns out, is not a problem to be eliminated. It is an asset to be managed.