PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Churn: Predicting Financial Fragmentation in Retail Banking with Temporal Machine Learning

Ananyaa Chopra, Brandon Xu, Brendan Yuen, Lauren Zung, Sarabroop Aulakh

arXiv:2608.30364v1cs.LG

TL;DR

Retail banking churn models often miss earlier partial movements of assets and activity while accounts remain open. This paper defines that state as financial fragmentation and develops a four-stage temporal XGBoost system using client-month data to predict and contextualize external outflows. The system achieves strong rare-event discrimination and concentrated ranking performance, while the authors retain explicit boundaries around data constraints and causal interpretation.

  • Problem

    Existing churn formulations typically treat attrition as terminal, limiting attention to partial external movements that can occur while a multi-product banking relationship remains open.

  • Method

    A four-stage temporal XGBoost cascade predicts external fragmentation within 90 days, expected outflow amount, originating product, and destination institution from client-month data.

  • Results

    The system achieves strong rare-event discrimination and concentrated top-of-list performance across event, amount, source-product, and destination predictions.

  • Takeaways & Limitations

    The framework provides a practical decision-intelligence layer for investigating emerging fragmentation earlier in the relationship lifecycle while preserving human review.

  • Takeaways & Limitations

    Results are based on a 10% sample, leaving rare products and small destination classes data constrained and potentially limiting minority-class performance.

Abstract

from arXiv · show

Retail banking attrition is usually represented as a terminal binary event, even though client relationships often weaken earlier through partial movements of deposits, investments, and recurring activity to external financial institutions. This paper defines that preceding state as financial fragmentation and presents an end-to-end temporal machine-learning system for predicting it before complete disengagement. Using anonymized multi-source data from a large retail bank, the framework predicts whether a valid external transfer or investment event will occur within 90 days. The study uses 595,220 client-month observations, with 346 engineered features combining monthly client profiles, balances, product relationships, prior flow-of-funds behavior, macroeconomic conditions, and competitor activity. A four-stage XGBoost cascade estimates (1) whether an external outflow will occur within 90 days, (2) the expected amount, (3) the originating product, and (4) the destination financial institution. The primary classifier achieved a test precision-recall area under the curve of 0.823. At the validation-selected threshold, it produced 86.4% precision, 75.1% recall, and an F1 score of 0.803. Ranking test observations in descending Stage 1 fragmentation score, the top 1% of clients yielded 95.3% precision, while the top 5% captured 78.7% of observed outflow cases. The amount model placed 94.9% of predictions within an adjacent amount bucket. Destination prediction reached a macro-F1 of 0.81 across 27 classes; source-product prediction achieved a weighted F1 of 0.92. By moving the analytical focus from terminal churn to earlier fund migration, the proposed approach provides a practical foundation for proactive, explainable, and economically informed client-retention decision support.

1 Introduction

The paper reframes retail banking attrition as financial fragmentation: observable external movements can precede complete disengagement. It proposes a four-stage predictive system to identify, quantify, and contextualize these movements for human-reviewed decision support.

  • Motivation and framing: Conventional churn models focus on terminal outcomes such as closure, prolonged inactivity, or formal defection, while multi-product customers may partially defect.This creates a gap between formal retention and the economic weakening of a relationship.
  • Motivation and framing: Financial fragmentation is the redistribution of assets or transactional activity across external institutions while the primary banking relationship remains open.The definition is behavioural and targets observable movements rather than inferred dissatisfaction.
  • Proposed system: Stage 1 predicts at least one qualifying external flow-of-funds event within 90 days, followed by predictions of amount, originating product, and destination institution.The outputs are combined into an expected-outflow score for prioritizing investigation.
  • Proposed system: The system joins client profiles, balances, product relationships, prior flow-of-funds behaviour, macroeconomic conditions, and competitor activity in a temporal client-month framework.The setting includes rare positive events, repeated monthly observations, and potential temporal leakage risks.
  • Operational contribution: The paper presents an explainable, human-reviewed workflow using expected-outflow ranking, dashboards, and conversational analysis.The intended users include portfolio, analytics, and relationship teams with limited outreach capacity.
  • Scope and boundaries: The results are promising but bounded by difficult outflows, weak recovery of minority source products, and data limitations.These constraints shape interpretation and future development.

2 Related Work

Related work motivates moving beyond static, terminal churn targets toward temporal prediction of partial relationship erosion. It also supports value-aware ranking, tree-based modeling, rare-event evaluation, temporal validation, and careful interpretation of feature importance.

  • Churn and partial defection: Prior churn research shows that target definitions, included observations, and performance measures materially affect predictive conclusions.Partial defection research treats relationship erosion as a process rather than a single endpoint.
  • Churn and partial defection: Financial fragmentation differs from conventional churn because it captures observed external movement within a defined future window from a monthly snapshot.A client may be nonfragmenting in one month and fragmenting in the next, enabling within-client temporal learning.
  • Value-aware retention: Value-aware retention separates fragmentation probability from potential outflow magnitude, allowing rankings to reflect unequal economic consequences and outreach costs.The implemented score multiplies predicted probability by predicted amount without assuming treatment response or intervention cost.
  • Modeling approach: XGBoost is selected for nonlinear interactions, sparse and missing data handling, multiclass objectives, class weighting, and enterprise-feasible training.These properties match financial features such as balances, ratios, lags, categorical encodings, and macroeconomic variables.
  • Evaluation: Precision-recall and capacity-constrained ranking are emphasized because rare-event workflows must identify concentrated positives within review limits.The evaluation uses PR-AUC, precision, recall, F1, capture, and lift.
  • Evaluation: Temporal separation and robustness to changing conditions are necessary because future leakage and random splits can overstate deployed performance.The project uses a three-month buffer between training and held-out evaluation data.
  • Interpretability and risk: Feature-importance scores describe how the model uses variables but do not establish that changing a feature would prevent fragmentation.Interpretation should therefore remain distinct from causal intervention claims.
  • Interpretability and risk: Production use also requires attention to unstable dependencies, drift, pipeline coupling, ownership, documentation, and governance.These concerns extend beyond predictive error.

3 Methodology

The methodology models financial fragmentation at the client-month level using joined data sources and a four-stage prediction pipeline. Temporal buffering and held-out validation and test periods support model selection and evaluation.

  • Research setting and data: The dataset combines four source families joined at a monthly analytical grain.The supplied methodology identifies the source families and governance table as part of the data construction.
  • Research setting and data: The analytical unit is a client-month snapshot indexed by client i and observation month t.Features contain information available by the snapshot date.
  • Targets and pipeline: The dataset records a 90-day fragmentation label, future outflow amount, originating-product class, and destination-institution class.These targets correspond to the pipeline's conditional prediction stages.
  • Targets and pipeline: The four-stage model converts each monthly snapshot into a risk probability, amount estimate, source-product prediction, destination prediction, and human-reviewed priority list.This links event detection with downstream contextualization and prioritization.
  • Temporal evaluation: 595,220 client-month observations are used, with January-March 2025 for training and July-September 2025 for evaluation after a three-month buffer.The evaluation pool is randomly divided approximately 50/50 into validation and test subsets.
  • Temporal evaluation: Validation supports hyperparameter and threshold selection, while the test set is not used for those choices.Both validation and test are held out from training.

3.3 Fragmentation label and conditional targets

The paper defines fragmentation as qualifying external financial activity within a forward horizon and derives conditional targets for amount, originating product, and destination institution.

  • Fragmentation label: A qualifying fragmentation event is an external OFI Investments or OFI Transfers transaction with a valid, non-null external destination institution.Internal destinations, unknown values, and not-applicable labels are excluded.
  • Fragmentation label: The primary label has no minimum dollar threshold, allowing detection of small early external movements.
  • Conditional targets: The originating product is consolidated from 16 raw labels into six business-relevant groups, while normalized destination institutions form 27 classes.The six source groups are chequing, savings, mutual funds, wealth management, GICs, and other.
  • Conditional targets: Stages 2–4 are trained only on observations with actual positive outflow events because their targets are defined conditionally on fragmentation.In deployment, Stage 1 gates the downstream outputs shown for analyst review.
  • Feature construction: The feature matrix contains 346 predictors built only from information available at or before the snapshot date.Feature families include client profiles, balances, product relationships, flow-of-funds behavior, and external context.

3.5 Stage 1: 90-day fragmentation classifier

Stage 1 uses a balance-weighted XGBoost classifier to predict 90-day external fragmentation, with validation-based threshold selection and regularized tuning.

  • Model: Stage 1 uses an XGBoost binary classifier with logistic loss to estimate fragmentation risk.
  • Weighting: Total balance is used as a sample weight so higher-AUM clients receive greater influence without allowing extreme outliers to dominate training.
  • Weighting: The normalized training weights have mean 1.000 and range from 0.133 to 2.146.
  • Tuning: A 40-trial randomized search tuned class weighting, tree depth, and regularization using validation PR-AUC and early stopping.No negative-class downsampling was applied, and the selected scale_pos_weight was 1.
  • Threshold selection: The validation-selected threshold was τ*=0.3321, yielding precision 0.8657, recall 0.7567, and F1 0.8075.The threshold was selected to maximize validation F1.

3.6 Stage 2: conditional amount model

Stage 2 models future outflow amounts conditionally on fragmentation, addressing strong right skew while balancing overall fit with detection of larger events.

  • Amount modeling: Future outflow amounts are highly right skewed, so Stage 2 uses an XGBoost regressor with transformed predictions returned to the dollar scale.The continuous target preserves ordering and provides a dollar estimate.
  • Amount modeling: Bucket-based weights increase attention to larger outflows without assigning them unbounded influence.Weights are 1.0 below $500, 1.5 from $500 to $2,500, 2.5 from $2,500 to $12,500, and 4.0 at $12,500 or more.
  • Configuration: The final amount model uses max_depth=5, learning_rate=0.02, min_child_weight=50, subsample=0.7, and colsample_bytree=0.7.
  • Configuration: Model selection combines validation log-scale R² with recall for outflows above $12,500.This balances average fit against high-value outflow detection.

3.7 Stages 3 and 4: originating product and destination

Stages 3 and 4 predict the originating product and destination institution, while the operational score combines fragmentation likelihood with estimated outflow magnitude for portfolio ranking.

  • Stages 3 and 4: The originating-product and destination models use XGBoost’s multiclass soft-probability objective.
  • Stages 3 and 4: Both models combine total-balance weighting with square-root-dampened inverse-frequency class weighting.This reduces majority-class dominance without the extreme weights of full inverse-frequency weighting.
  • Configuration: Both stages use n_estimators=500, max_depth=4, learning_rate=0.05, and early_stopping_rounds=25.
  • Decision score: The operational ranking score is the product of Stage 1 likelihood and predicted outflow amount.A 90% risk score with a predicted $50,000 outflow receives a $45,000 expected-outflow score.
  • Decision score: A retention-value extension could include intervention response and cost, but it is not evaluated because controlled treatment-response data are unavailable.
  • System integration: The dashboard and conversational agent provide advisory access to predictions, without denying services, blocking transactions, or initiating automated client action.

3.9 Evaluation protocol

The evaluation protocol uses task-specific metrics across the cascade, emphasizing rare-event ranking, monetary accuracy, and multiclass performance. Stage 1 scores lack separate probability calibration and should therefore be interpreted mainly as ranking and relative-risk measures.

  • Stage 1 uses PR-AUC, threshold precision, recall, F1, and capacity-constrained lift to evaluate rare-event prediction and portfolio review performance.At review fraction k, lift is defined relative to the observed positive rate; capture measures the fraction of positives in the top-ranked k observations.
  • Stage 2 evaluates log-target RMSE and R6, original-scale MAE, amount-bucket accuracy, within-one-bucket accuracy, and threshold-specific performance for large outflows.
  • Stage 3 and Stage 4 use accuracy, weighted F1, macro-F1, per-class precision and recall, and confusion matrices.Macro-F1 receives particular emphasis because it gives equal weight to minority classes.
  • No separate probability-calibration study was completed, so Stage 1 scores should be interpreted primarily as ranking and relative-risk measures.

4 Results

The four-stage cascade shows strong Stage 1 discrimination and portfolio-ranking performance, while downstream models vary by task: destination prediction is comparatively balanced, whereas amount and rare source-product predictions remain limited.

  • Stage 1 discrimination and threshold performance: 0.823 test PR-AUC was achieved by the Stage 1 classifier.Training and validation PR-AUC were 0.828 and 0.823, respectively.
  • Stage 1 discrimination and threshold performance: 86.4% precision and 75.1% recall were obtained at the validation-selected threshold, with 0.66% false-positive rate.The reported interpretation emphasizes that the model retrieves three quarters of observed fragmentation cases while 86% of positive flags are correct.
  • Capacity-constrained ranking and lift: 95.3% precision in the top 1% and 78.7% of observed positives captured in the top 5% support capacity-constrained review.The top 5% contains nearly four fifths of all positive cases, concentrating event prevalence for relationship-management teams.
  • Amount estimation: The amount model achieved log-scale RMSE 1.3613, R^2 of 0.4998, and $11,098 original-scale MAE on 7,893 positive test observations.Bucket performance was strongest for the smallest two ranges, with weaker recall for the largest range.
  • Amount estimation: Amount predictions provide broad segmentation but lose recall for very large outflows, limiting their use as precise extreme-tail estimates.The reported high-value predicted-to-actual ratio was 0.64, consistent with underprediction.
  • Conditional classification stages: The destination model reached 85% accuracy, 0.81 macro-F1, and 0.85 weighted F1 across 27 normalized institutions, while source-product weighted F1 was 0.92 but macro-F1 0.49.Source-product performance is strong for dominant chequing examples but materially weaker for minority products; destination performance is comparatively better balanced.

5 Discussion

The results support external flow behavior as a concentrated pre-churn signal and translate predictions into capacity-aware, human-reviewed retention decision support. The framework remains bounded by data, observability, tail-estimation, and causal-interpretation limitations.

  • Predictive signal and workflow: 0.823 test PR-AUC, 95.3% precision in the top 1%, and 78.7% case capture in the top 5% demonstrate concentrated Stage 1 signal.These rankings support investigation workflows when staff capacity is constrained.
  • Predictive signal and workflow: Expected-outflow scoring prioritizes clients by combining fragmentation likelihood with predicted financial magnitude.The score distinguishes high-probability small movements from lower-probability large movements.
  • Predictive signal and workflow: Destination and source-product predictions contextualize exposure by identifying likely receiving institutions and originating products for human review.The dashboard and conversational layer make these distinctions accessible to non-modeling users.
  • Behavioral interpretation: Destination diversity, recent acceleration, and source-product history provide behavioral signals consistent with provider exploration, repeated external activity, and increasing momentum.The interpretation is diagnostic rather than causal: destination count is a marker of relationship diversification, not an intervention target.
  • Limitations and next steps: The 10% sample constrains rare-product and small-destination performance, while direct-flow labels exclude observable multi-hop movements and depend on transaction normalization.Full-scale retraining may improve minority-class performance but could introduce computational and feature-pruning requirements.
  • Limitations and next steps: Amount predictions underrepresent the extreme tail, and feature importance identifies predictive associations rather than interventions that would change outcomes.The paper proposes tail-focused models and eventual causal or uplift evaluation as next steps.

6 Conclusion

The paper frames financial fragmentation as an actionable pre-churn prediction problem and demonstrates a temporal XGBoost framework for identifying, quantifying, and contextualizing external fund migration. Its practical contribution is an earlier, human-reviewed decision-intelligence layer rather than automated explanation or intervention selection.

  • Conclusion: The framework identifies, quantifies, and contextualizes external fund migration as an actionable pre-churn prediction problem.It uses a four-stage temporal XGBoost system across client-month observations.
  • Conclusion: Strong rare-event discrimination and concentrated top-of-list performance support moving organizational attention from completed churn toward emerging fragmentation.The conclusion emphasizes investigation earlier in the relationship lifecycle.
Loading 2608.30364v1…