AI-Driven Employee Turnover Warning for Talent-Retention Decision Support in Finance and Taxation Education: Model Development and Comparison

1. Introduction

Employee turnover interrupts client relationships, removes tacit knowledge, and increases recruitment and training costs. The problem is especially uneven in finance and taxation education and training. Instructors and curriculum researchers depend on professional knowledge; sales consultants work under performance pressure and frequent customer contact; operations and back-office employees follow different career paths. A single generic risk rule can therefore hide meaningful occupational differences.

Artificial intelligence can rank employees for supportive review, but administrative HR data impose hard limits. The study enterprise supplied one de-identified administrative export rather than versioned monthly snapshots. Employees recur across reconstructed monthly records, departures are uncommon, organizational categories are sparse, and some field values cannot be confirmed at every historical prediction date. Random row splitting would allow the same employee profile to enter both development and evaluation. Accuracy alone would reward majority prediction, while a high score obtained after extensive tuning could reflect selection noise rather than generalizable information.

The analysis addresses five questions:

• How do widely used classifiers compare when they follow the same employee grouping and time rules?
• Can one ensemble use standard fields, organizational hierarchy, and category-aware learning together?
• How well do the models work for employees who took no part in model development?
• Which measures describe risk ranking, the usefulness of the alert list, and the accuracy of probability scores?
• How should uncertain field dates, small job-group counts, and responsible-use requirements limit the conclusions?

The proposed model is introduced at first mention as the Hierarchical Risk Encoding XGBoost–CatBoost Ensemble (HRE-XCB). HRE denotes hierarchical risk encoding, and XCB identifies its XGBoost and CatBoost foundations. It is a nonnegative weighted average of Raw-XGB, an XGBoost branch augmented with employee-excluding hierarchical risk encoding (HRE-XGB), and native CatBoost. The method is not assumed to be superior by design. It is evaluated alongside 22 general-purpose classifiers, with comparable tuning for the strongest baselines.

Four parts of the study are especially important. First, 278 employees are set aside before feature screening, encoding, tuning, weighting, or cutoff selection. Second, the method turns organizational hierarchy into smoothed summaries calculated without the employee being scored. It also keeps the original fields and CatBoost’s category-aware representation. Third, ranking, alert-list performance, probability error, component ablation, bootstrap intervals, and field-date checks are reported separately. Fourth, the model has one permitted use: confidential and supportive HR review. It cannot trigger dismissal, pay reduction, promotion restriction, or any other penalty. The remaining sections explain the theory, methods, results, and limits.

2. Related Work and Research Framework

2.1. Turnover Theory and the Observable-Data Boundary

Turnover theory evolved from linear withdrawal processes to heterogeneous pathways. Mobley linked dissatisfaction, thoughts of quitting, alternative search, and departure [8]. Meta-analytic work connected satisfaction and commitment to intention and realized turnover [9]. Job embeddedness explained why links, fit, and sacrifice can keep employees in place [7], while job demands–resources, perceived organizational support, and psychological-contract theories described workload, resources, reciprocity, and expectations [10,11,12]. Historical reviews further showed that shocks, opportunities, and local context can redirect the process [1,2,3,4,5,6]. These theories define plausible mechanisms, but the present administrative data do not measure them directly. Job, department, tenure, education, and age are predictive proxies, not validated measures of commitment, embeddedness, or intent.

2.1.1. Machine-Learning Development and Employee-Turnover Studies

The technical literature spans linear probability models, nearest neighbors, support vector machines, trees, bagging, boosting, and neural networks [13,14,15,16,17,18,19,20]. Attrition studies have applied overlapping subsets and reported different winners [50,51,52,53,54]. Their disagreement is expected: model ranking depends on the outcome definition, positive rate, feature history, split unit, and tuning protocol. Public benchmarks support replication but rarely reproduce sparse internal categories, uncertain information timing, or repeated employee records. Consequently, this study compares algorithms under one protocol instead of importing a winner from another dataset.

Representation is central in the study enterprise. One-hot encoding is transparent but unstable for rare categories. Direct target encoding is compact but can leak an employee’s outcome. CatBoost reduces this risk through ordered category statistics, whereas HRE uses employee-excluding, parent-shrunk organizational summaries. Research that converted spatial layouts and qualitative color concepts into explicit numerical structures illustrates the broader principle that domain knowledge should become inspectable features rather than unverifiable labels [57,58]. Resource-constrained embedded AI also reinforces the need to match model capacity to available information and computing resources [55].

2.1.2. Imbalance, Ranking, and Probability Reliability

Turnover warning is an imbalanced screening task. SMOTE can expand the minority class [22], but interpolation in sparse mixed-type space may create implausible employee combinations; resampling before grouped splitting can also leak information. Imbalance methods must therefore be evaluated with the classifier and metric rather than treated as automatic improvements [23]. Recall measures coverage of future leavers, precision measures the efficiency of the alert list, and F1 balances both at one cutoff. PR-AUC evaluates ranking when positives are uncommon [25], while ROC-AUC supplies a complementary global ranking measure [26].

Ranking and probability meaning are distinct. Platt scaling, isotonic regression, and related transformations can improve the agreement between scores and observed frequencies without improving rank order [24,27]. Enterprise early-warning research likewise shows that indicator definition, risk output, and management response must form one decision chain [56]. Work on negative indicators and ordering stability adds a related warning: direction and rank interpretation must remain explicit when evidence is combined [59]. This study therefore reports cutoff-based, ranking, and probability-reliability measures separately.

2.1.3. Limited-Data Validation and Model Explanation

Limited data make model selection part of the uncertainty. Prediction-model sample size depends on events, candidate parameters, expected fit, and overfitting [29]. Small-sample rankings can vary substantially [30], and cross-validation becomes optimistic when the same results drive both selection and error estimation [31,32]. Folds must also respect time and employee dependence [33]. A later month with overlapping employees is not equivalent to evaluation on employees absent from training. This distinction is essential in the present study.

LIME and SHAP explain model behavior, not employee psychology [21,35,36]. Prediction and mechanism explanation answer different questions [34]. TRIPOD’s reporting logic is therefore relevant even outside clinical research: readers need the outcome, prediction time, candidate variables, missing-data handling, selection process, and validation role to be explicit [28]. For the three-branch ensemble, grouped permutation importance provides one model-level dependence measure across heterogeneous component types; branch-specific explanations are not relabeled as ensemble explanations.

2.1.4. People Analytics and Algorithmic Governance

People analytics requires alignment among the question, data, analytical capability, and organizational action [37,38,39,40]. AI can support HR decisions, but it also redistributes authority and can reshape control and job design [42,43,47,48,49]. Privacy, proxy discrimination, incompatible fairness definitions, and high-stakes opacity remain material risks [41,44,45,46]. The legitimate use in this study is therefore narrow: restricted-access risk ranking, confidential human review, supportive contact, and monitored outcomes. The score cannot justify dismissal, pay reduction, promotion restriction, or any automatic adverse action.

2.2. Conceptual Framework

The framework links management theory to observable administrative fields without treating proxies as psychological constructs. Job embeddedness, job demands–resources, organizational support, and psychological contract theory guide plausible questions about attachment, workload, resources, and reciprocity [7]. The prediction layer then uses only information supportable at the prediction date. A three-branch representation produces risk scores, and the evaluation layer separates ranking, alert-list utility, probability error, uncertainty, and subgroup evidence. The action layer restricts output to confidential human review.

Figure 1. Employee -turnover warning framework for the study business setting.

Figure 1. Employee -turnover warning framework for the study business setting.

The framework begins with prediction-time information screening and employee separation before model development. Raw-XGB, HRE-XGB, and CatBoost form HRE-XCB with development-selected weights of 0.65, 0.15, and 0.20. Evaluation separates risk ranking, alert coverage, review burden, and probability error; model output enters confidential human review rather than automated employment action.

2.3. Research Gaps

Three gaps motivate the design. First, much attrition evidence comes from public benchmarks or row-random splits that do not reproduce repeated employees and sparse enterprise categories. Second, studies often blur prediction rank, one-cutoff classification, and calibrated probability, even though they answer different management questions. Third, domain-informed features are rarely tested under strict employee separation and explicit information-timing sensitivity. These gaps prevent a high internal score from establishing transfer to different employees or future operating conditions.

2.4. Methodological Contribution

HRE-XCB contributes a testable representation rather than a universal-best claim. Its hierarchical encoding excludes the evaluated employee and shrinks sparse categories toward parent groups. Its three branches preserve raw, hierarchical, and category-native representations. The empirical protocol compares the ensemble with tuned baselines on the same separated employees, quantifies uncertainty in model differences, and reports timing sensitivity and component ablation. The ablation can reject an architectural advantage; this distinction keeps novelty separate from empirical ranking.

2.5. From Management Theory to Measurable Evidence

Management theories and prediction algorithms occupy different levels of explanation. A turnover theory proposes why withdrawal may occur; a prediction model estimates which records resemble observed departures; an intervention study tests whether a management action changes outcomes. Conflating these levels creates two errors. The first is construct inflation: calling an administrative field such as tenure a direct measurement of commitment. The second is causal inflation: interpreting a predictive association as proof that changing the field would prevent departure. The present framework avoids both errors by stating what each theory contributes and what the available data cannot test.

Table 1. Management theories, observable evidence, and inference boundaries.

Theoretical perspective	Core mechanism in the literature	Administrative evidence available here	Valid use in this study	Inference not supported
Withdrawal process	dissatisfaction may lead to thoughts, search, and departure	realized voluntary departure and employment dates	define a 90-day prediction target	reconstruct an employee’s unobserved psychological sequence
Job embeddedness	links, fit, and sacrifice inhibit leaving	job family, organizational location, and tenure proxies	motivate structured organizational representation	claim that a category code measures embeddedness
Job demands–resources	demands deplete resources; resources buffer strain	no validated workload or resource scale	motivate supportive follow-up questions	estimate burnout or workload effects
Perceived organizational support	reciprocity follows perceived care and recognition	no direct perception measure	constrain alerts to supportive human review	infer perceived support from a risk score
Psychological contract	unmet expectations can trigger withdrawal	no expectation or breach measure	motivate confidential dialogue about expectations	diagnose contract breach from administrative data
Unfolding and shock models	events can initiate different departure paths	no event-history or exit-interview sequence	explain why one static rule may miss heterogeneous paths	identify the shock that caused an individual departure

Table 1. Management theories, observable evidence, and inference boundaries.

Theoretical perspective	Core mechanism in the literature	Administrative evidence available here	Valid use in this study	Inference not supported
Withdrawal process	dissatisfaction may lead to thoughts, search, and departure	realized voluntary departure and employment dates	define a 90-day prediction target	reconstruct an employee’s unobserved psychological sequence
Job embeddedness	links, fit, and sacrifice inhibit leaving	job family, organizational location, and tenure proxies	motivate structured organizational representation	claim that a category code measures embeddedness
Job demands–resources	demands deplete resources; resources buffer strain	no validated workload or resource scale	motivate supportive follow-up questions	estimate burnout or workload effects
Perceived organizational support	reciprocity follows perceived care and recognition	no direct perception measure	constrain alerts to supportive human review	infer perceived support from a risk score
Psychological contract	unmet expectations can trigger withdrawal	no expectation or breach measure	motivate confidential dialogue about expectations	diagnose contract breach from administrative data
Unfolding and shock models	events can initiate different departure paths	no event-history or exit-interview sequence	explain why one static rule may miss heterogeneous paths	identify the shock that caused an individual departure

This separation changes the interpretation of feature importance. A high contribution from job family can mean that the model uses stable differences among categories. It does not reveal whether those differences arise from supervision, compensation, labor-market opportunities, career fit, or data-recording practices. Conversely, a theoretically important construct may have no predictive contribution because it was never measured. The model is therefore a screening aid situated within management theory, not a computational replacement for that theory.

The framework also clarifies the research claim. The architectural question is whether employee-excluding hierarchical summaries add a useful representation to raw structured features and ordered categorical learning. The empirical question is whether that representation improves specified endpoints under a fixed evaluation boundary. The management question is whether any resulting ranking can support a proportionate, non-punitive response. A positive answer to one question does not automatically answer the others.

3. Research Methodology

3.1. Study Design, Prediction Target, and Notation

3.1.1. What the Model Predicts

This study uses de-identified operational employee data from an unnamed finance and taxation education and training enterprise. The practical question is simple: among employees who are still working on a given monthly observation date, who will voluntarily leave within the next 90 days? The task is binary because each monthly prediction record is labeled either as a voluntary departure within that window or as no observed voluntary departure within that window.

This task predicts behavior; it does not establish why the behavior occurs. Turnover theories discuss job satisfaction, organizational commitment, alternative opportunities, shocks, job embeddedness, job demands and resources, perceived organizational support, and psychological contracts [1,2,3,4,5,6,7,8,9,10,11,12]. The available administrative fields do not directly measure most of these concepts. Job family, tenure, department, and education may help prediction, but they are only observed characteristics or indirect signals. A statistical association with turnover is not evidence that the characteristic caused the departure.

The model is intended to order employees by risk for supportive review when HR staff cannot speak with everyone at once. It does not estimate the effect of a salary increase, reassignment, or other retention action. A feature that is important to prediction should not be treated as a lever that will necessarily change an outcome. Prediction and causal explanation answer different questions [34]. For this reason, the paper uses terms such as predictive association, risk signal, and information used by the model.

Let employees be indexed by $i=1,\dots ,N$, and let monthly observation points be indexed by $t\in \mathcal{T}$. A record is created only when the employee has joined and has not already left by month t. The available predictors for that employee and month are written as ${\mathbf{x}}_{it}$. The target is

$$y_{it}=\mathbb{I}\left\{T_i^{\left(V\right)}>t,T_i^{\left(V\right)}\le t+90\mathrm{days}\right\},$$

where $T_i^{\left(V\right)}$ is the recorded voluntary-departure date and $\mathbb{I}\{·\}$ is an indicator that equals one when the condition is true. Thus, $y_{it}=1$ means that employee i voluntarily leaves in the 90 days after monthly observation point t. A person who was dismissed or otherwise left involuntarily during the same period is excluded from that monthly comparison. Such a record should not be mislabeled as either a voluntary departure or a stable non-departure. An employee who had already left before the observation point is not eligible for that month’s prediction.

The predictive model f produces a score

$${\widehat{p}}_{it}=f\left({\mathbf{x}}_{it}\right)\in [0,1].$$

The score becomes an HR alert only after applying an alert cutoff $\tau$:

$${\widehat{y}}_{it}=\mathbb{I}\{{\widehat{p}}_{it}\ge \tau \}.$$

The score is bounded between zero and one, but it is first interpreted as a research risk score. A larger score should place an employee higher on the near-term risk list. Before the model is tested on genuinely future cohorts, the number should not be presented as a literal personal probability. The study therefore separates three questions: whether the model orders employees well, whether the chosen cutoff creates a useful alert list, and whether predicted values agree with observed group frequencies [24,25,27]. The last question is called probability reliability, or calibration, and is explained in Section 3.6.

3.2. General-Purpose Comparison Models

The experiment implements 22 general-purpose comparators under the same employee-group and time rules. They cover a no-information reference, linear and probabilistic models, neighborhood and margin methods, trees and ensembles, boosting, and a small neural network. The catalog is broad enough to test competing ways of learning from data, but it is not presented as every classifier ever published.

3.2.1. No-information, Linear, and Probabilistic Models

DummyPrior assigns the training prevalence to every record and exposes the majority-class accuracy trap. Logistic regression models a regularized linear function on the log-odds scale; SGDLogLoss optimizes the same loss incrementally and is more sensitive to scaling and step size. Shrinkage LDA estimates a linear discriminant while stabilizing the covariance matrix. GaussianNB models continuous predictors with class-specific Gaussian distributions, whereas BernoulliNB models binary indicators. Their low capacity is useful with limited data, although conditional-independence and distributional assumptions can miss HR interactions.

3.2.2. Neighborhood, Margin, and Neural Models

KNN predicts from nearby training records and therefore depends on scaling, distance, dimensionality, and local density [17]. Repeated monthly records make it especially vulnerable to recognizing employee profiles rather than learning patterns that transfer to different employees. LinearSVM estimates a maximum-margin linear boundary; RBFSVM adds a radial kernel and greater nonlinear flexibility [14]. The MLP learns nonlinear interactions through back-propagation [18], but its parameter count and flexible surface increase selection variance when independent employees are limited. These models are regularized and tuned only within the development protocol.

3.2.3. Trees, Bagging, and Imbalance-Aware Ensembles

A decision tree recursively selects splits that reduce node impurity. Its interpretability is offset by high variance. BaggingTree averages trees fitted to resampled data. RandomForest additionally randomizes candidate features [13], and ExtraTrees randomizes split selection more strongly. BalancedRandomForest trains trees on class-balanced samples; EasyEnsemble averages learners fitted to several balanced subsets. These two methods can raise minority recall, but changing the training class distribution may reduce precision and distort raw probability values.

3.2.4. Boosting Models

AdaBoost increases the influence of misclassified observations [16]. GradientBoosting fits successive trees to the loss gradient [15], and HistGradientBoosting bins continuous values for speed and regularization. XGBoost adds second-order loss information, explicit tree penalties, row and column sampling, and efficient construction [19]. LightGBM uses histogram-based, leaf-wise growth. CatBoost processes categories natively and uses ordered statistics to reduce target leakage. Comparative evidence shows that their relative performance remains data dependent [20]; library names do not determine the winner.

All 22 comparators use identical prediction-time feature restrictions, employee weighting, grouped folds, model-selection data, and alert-cutoff rules. Earlier attrition studies are used to motivate coverage, not to transfer their reported scores [50,51,52,53,54]. The unified protocol tests whether an apparent advantage survives the same business target and the same evaluation settings.

3.3. Proposed HRE-XCB Three-Branch Ensemble

3.3.1. Design Rationale

The proposed method is named the Hierarchical Risk Encoding XGBoost-CatBoost Ensemble (HRE-XCB). The formal name identifies how categories are summarized, which tree learners are used, and how their outputs are combined. It does not claim a newly invented boosting algorithm. The method combines established learners in a structure designed for the limited employee data of the study enterprise.

The design targets three sources of instability at once: sparse one-hot categories, noisy outcome summaries for small organizational groups, and optimistic category statistics computed with an employee’s own result. The branches address these problems differently, while the constrained weighted average limits the freedom of the final combination. This structure makes each claimed contribution testable through single-branch and partial-ensemble comparisons.

The central problem is how to express several types of HR information without fitting noise. The predictors include numeric values, stable labels, category fields with many distinct job or department names, and broad-to-specific business groups. One-hot encoding is transparent, but it treats every department or job as unrelated to every other one. An outcome-based category average, often called target encoding, can summarize risk but may give a rare category an extreme value after one departure. Calculating that average from all rows would also let an outcome help create its own predictor. CatBoost offers another category-handling method, but it does not guarantee the specific broad-to-narrow borrowing designed for this study. No single representation resolves all these problems.

HRE-XCB therefore uses three modeling branches built from the same permitted employee information:

• Raw-XGB learns from filled, standardized, and one-hot-encoded fields.
• HRE-XGB adds category risk summaries that exclude the employee’s own outcome and borrow strength from broader business groups.
• CatBoost keeps the original category labels and calculates category summaries without using the current row’s outcome.

The branches use overlapping information but learn patterns differently. Raw-XGB finds nonlinear splits in the standard prepared matrix. HRE-XGB adds the planned business hierarchy. CatBoost handles fields with many distinct labels without manually creating a column for each one. When the branches make partly different errors, averaging may make the final score more stable near the alert cutoff. This possible benefit is tested through fixed weight selection and direct component comparison; it is not assumed from the architecture alone.

3.3.2. Branch I: Raw-XGB

Raw-XGB receives the permitted HR fields after training-only median and mode filling, missing-value flags, rare-category handling, standardization, and one-hot encoding. Its loss uses the employee-equal recency weight $w_{it}$ and a positive-class scale computed from weighted class totals:

$$\rho ={\displaystyle \frac{{\sum }_{i,t:y_{it}=0}{w}_{it}}{{\sum }_{i,t:y_{it}=1}{w}_{it}}}.$$

The class scale changes the relative influence of departures and non-departures during learning. Employee weights prevent repeated monthly records from increasing one person’s total influence. Candidate Raw-XGB settings vary tree depth, learning rate, tree count, minimum child weight, row and column sampling, $L_1$ regularization, $L_2$ regularization, and recency decay. The search favors shallow or strongly regularized trees because unrestricted boosting can divide small employee groups too precisely.

This branch serves two roles. First, it is a strong boosted-tree learner that can be compared with the general XGBoost baseline under a richer search. Second, it anchors the ensemble in fields prepared without using departure outcomes. If an outcome-based category summary reacts too strongly to a rare job, Raw-XGB can contribute a less outcome-sensitive score.

3.3.3. Business-Derived Categorical Structure

Before constructing the outcome-informed HRE branch, a small number of fixed business fields are derived. Continuous tenure is grouped into less than six months, six to twelve months, one to two years, two to three years, three to five years, and more than five years. Age is grouped into at most 25, 26-30, 31-35, 36-40, 41-50, and more than 50 years. The outcome does not determine these bands. They express understandable career stages and let the model combine nearby values.

A combined category allows the model to distinguish, for example, early-tenure sales advisers from early-tenure instructors. The implemented combinations are job family with tenure stage, job family with center type, job level with center type, education with major group, entry year with job family, and residential region with center type. The list is fixed as part of the model design. It is not created by searching every possible pair against the outcomes. This limits chance findings and keeps each combination understandable. Even so, the combination remains a predictive signal rather than proof of a behavioral mechanism.

The HRE branch also defines where a rare or new category should borrow information. Department borrows from center type; job and job level borrow from job family. Job-family interactions return to job family, job-level-by-center returns to job level, education-by-major returns to education, and residential-region-by-center returns to center type. A category without a sensible parent uses the overall departure rate in the development data. This is a practical representation of the business hierarchy. It is not a map of causal relationships.

3.3.4. Hierarchical Risk Encoding

Consider a specific category c, such as one job title, and its broader group $g\left(c\right)$, such as a job family. Within training data, let $s_c$ be the departure count, $n_c$ the record count, and ${\mu }_{g\left(c\right)}$ the estimated risk of the broader group. The smoothing value $\alpha >0$ controls how much the specific category borrows from that group:

$$r_c={\displaystyle \frac{s_c+\alpha {\mu }_{g\left(c\right)}}{n_c+\alpha }}.$$

The formula combines the category’s own history with the broader group’s history. In statistical terms, it is a simple beta-binomial posterior mean when $\alpha {\mu }_{g\left(c\right)}$ and $\alpha [1-{\mu }_{g\left(c\right)}]$ are read as prior pseudo-counts. Here, prior means the rate supplied before the specific category has enough observations; it is learned from the development data, not imposed as a universal truth. A category without a defined broader group uses the overall training departure rate $\mu$. Broader groups are also pulled toward $\mu$. A rare job therefore first borrows from its job family, while an unseen family ultimately uses the overall rate.

The shrinkage has an intuitive sample-size effect:

$$r_c={\displaystyle \frac{n_c}{n_c+\alpha }}{\displaystyle \frac{s_c}{n_c}}+{\displaystyle \frac{\alpha }{n_c+\alpha }}{\mu }_{g\left(c\right)}.$$

When $n_c\ll \alpha$, the broader parent group has more influence. When $n_c\gg \alpha$, the category’s own history has more influence. Candidate smoothing values are 12, 35, and 70. They range from lighter to stronger borrowing and are selected only through grouped development data. Results from later historical evaluations do not choose the smoothing value.

Three outputs are generated for every encoded feature:

$$R_c=r_c,L_c=log{\displaystyle \frac{r_c}{1-r_c}},Q_c=log(1+n_c),$$

Values are kept slightly above zero and below one so that the log-odds is defined. Here, $n_c$ is the raw category support count learned from the employee-excluding encoding-fit data. It is fixed after fitting and does not depend on the number of employees scored together. $R_c$ is the borrowed risk estimate. $L_c$ expresses it on the log-odds scale, which spreads out values near zero and one. $Q_c$ tells the model how much training support underlies the estimate. Two categories can have the same estimated risk but very different evidence, and the downstream tree can learn that distinction.

For an unseen node at transformation time, $R_c$ and $L_c$ are computed from its available parent estimate, and its support count is set to zero. Unknown values do not receive an arbitrary zero risk. This behavior matters operationally because new job titles or reorganized departments can appear after training.

3.3.5. Cross-Fitted Category Summaries That Exclude the Employee

An outcome-based category summary must not use the result of the employee it is describing. Removing only the current row is not enough because another monthly record of the same person can contain a closely related result. HRE therefore uses cross-fitted encoding that excludes all records of the employee being encoded.

Let employees be partitioned into K deterministic folds ${\mathcal{G}}_1,\dots ,{\mathcal{G}}_K$. For every employee in fold k, the encoding state ${\Theta }^{(-k)}$ is fitted using employees outside ${\mathcal{G}}_k$:

$${\mathbf{e}}_{it}^{\mathrm{CF}}=E({\mathbf{x}}_{it};{\Theta }^{(-k)}),i\in {\mathcal{G}}_k.$$

No month of employee i is allowed into ${\Theta }^{(-k)}$. A stable hash of the employee identifier and random seed assigns the groups reproducibly. Five groups are used for the final encoding. Each training row first receives a summary built without that employee. A full training-only encoding table ${\Theta }^{\left(all\right)}$ is then created for the model-selection period, later evaluations, and deployment records:

$${\mathbf{e}}_{jt}^{\mathrm{new}}=E({\mathbf{x}}_{jt};{\Theta }^{\left(all\right)}).$$

This two-step procedure makes training resemble real use. During fitting, XGBoost sees category summaries that did not use the employee’s own outcome. During later scoring, it uses a encoding table learned from all permitted development employees, none of whom supplies the unknown future outcome of the person being scored. Excluding the whole employee is stronger than removing only one row and is central to preventing inflated results.

3.3.6. Branch II: HRE-XGB

HRE-XGB concatenates the conventional Raw-XGB matrix ${\mathbf{z}}_{it}$ and HRE vector ${\mathbf{e}}_{it}$:

$${\mathbf{u}}_{it}=[{\mathbf{z}}_{it};{\mathbf{e}}_{it}].$$

A regularized XGBoost learner then estimates $p_H=P(y=1\mid \mathbf{u})$. The same employee-equal, recency, and weighted class-ratio logic applies. Including the raw matrix prevents the risk encodings from becoming an information bottleneck. Numeric gradients and categories that do not benefit from supervised pooling remain available directly.

HRE-XGB is not assumed to be stronger on its own than Raw-XGB or CatBoost. Its purpose is to add a structured signal that may help only for certain employees or at a small ensemble weight. Direct component comparison must report weak performance if it occurs. The proposed contribution can lie in a tested combination and careful validation; every branch does not need to improve every measure by itself.

3.3.7. Branch III: CatBoost

The third branch keeps cleaned category labels and the fixed business combinations. Numeric columns remain numeric, while CatBoost is told which columns contain categories. Instead of calculating each category’s outcome average from the entire training set, CatBoost repeatedly places records in random orders. For the record at position k, its category summary can use only earlier positions in order $\pi$:

$${OTS}_{{\pi }_k,c}={\displaystyle \frac{{\sum }_{j<k}\mathbb{I}(x_{{\pi }_j}=c)y_{{\pi }_j}+a\mu }{{\sum }_{j<k}\mathbb{I}(x_{{\pi }_j}=c)+a}}.$$

Ordered boosting applies the same idea while fitting each new tree, so a row’s outcome has less opportunity to shape the prediction made for that row. Several random orders and CatBoost’s internal category combinations produce a representation different from the manually defined HRE hierarchy.

Candidate CatBoost settings vary ordered versus ordinary boosting, depth three to five, 360 to 520 iterations, learning rate 0.025 to 0.03, leaf regularization, random strength, and recency decay. The selected category branch uses CatBoost. Weighted class ratios and employee-equal sample weights are supplied. Shallow symmetric trees, $L_2$ leaf regularization, randomization, and slow learning limit complexity when employee data are limited.

CatBoost and HRE prevent self-use of outcomes in different ways. HRE builds an inspectable cross-fitted map that excludes the employee and borrows from broader business groups. CatBoost uses only earlier records within random orderings when it calculates category summaries. HRE follows the planned business hierarchy; CatBoost is more flexible when learning category combinations. The two branches therefore do not duplicate the same encoding method.

3.3.8. Nonnegative Probability Fusion

Let $p_R$, $p_H$, and $p_C$ be scores from Raw-XGB, HRE-XGB, and CatBoost. The final model takes a nonnegative weighted average whose weights sum to one. Mathematically, this is a weighted average:

$${\widehat{p}}_{\mathrm{HRE}-\mathrm{XCB}}=w_R{p}_R+w_H{p}_H+w_C{p}_C,w_R,w_H,w_C\ge 0,w_R+w_H+w_C=1.$$

The employee-separated development procedure selected the weights

$$(w_R,w_H,w_C)=(0.65,0.15,0.20).$$

Thus, Raw-XGB carries the largest development-selected weight, CatBoost provides a category-native view, and HRE-XGB contributes the explicit hierarchy experiment. Reporting the weights makes the role of each branch explicit; the ensemble is not an equal partnership.

The selected Raw-XGB uses 300 trees, learning rate 0.033333266, depth four, minimum child weight three, row subsampling 0.935770109, column subsampling 0.870363951, $L_1$ regularization 0.0659628, $L_2$ regularization 1.62984271, and recency decay 0.08. HRE-XGB uses 280 trees, learning rate 0.04, depth three, minimum child weight two, row and column subsampling 0.95, $L_1$ regularization 0.05, $L_2$ regularization 3.0, HRE smoothing 70, and recency decay 0.18. CatBoost uses 360 iterations, learning rate 0.03, depth five, leaf $L_2$ regularization 12, random strength 1.8, plain boosting, and recency decay 0.30. These settings are estimates for the development sample, not generic recommendations for another organization.

This simple weighted average is preferred to an unrestricted second-level model for three reasons. Its output remains within the range of the component scores. Only two weights are free after requiring the total to equal one, which limits instability. Every final score can also be rebuilt from three stored component scores. A more complex stacking model could learn nonlinear relationships among them, but limited employee data would make that extra layer easier to overfit. The formula reflects this study’s development evidence and is not presented as a universal weighting rule.

3.4. Two-Stage Model Search and Alert-Cutoff Selection

3.4.1. Candidate Space Defined before Later Evaluation

The search defines 152 base configurations before any later evaluation. Raw-XGB contributes 32 candidates: eight regularization and tree-shape profiles crossed with four recency-decay values. HRE-XGB contributes 96 candidates: the same eight profiles and four recency values crossed with three levels of borrowing from broader groups. Native CatBoost contributes 24 candidates: six CatBoost profiles crossed with four recency values. The saved experiment plan records this arithmetic, so the search can be checked and repeated.

Stage 1 uses three employee-based folds and seed 20260731 to screen all 152 candidates. Stage 2 retains six candidates from each modeling branch, producing 18 finalists. They are checked again with four employee-based folds and three seeds: 20260731, 20261728, and 20262725. Using several seeds reveals whether a result depends on a favorable split or initialization. Seven candidates are then fixed for the final comparison: selected single-branch models and a small set of nonnegative weighted averages. Only the model-selection period may rank them.

The two stages balance breadth and stability. Repeating the most expensive procedure for every configuration would increase computation without adding independent employees. Selecting directly from one three-fold run, however, could choose a lucky candidate. The second stage spends repeated evaluation only on a limited finalist set. Cross-validation can still look optimistic after many models have been compared [31,32]. Candidate selection therefore uses grouped performance across folds and seeds, while the records from employees kept out of model development remain outside every development decision.

3.4.2. Choosing an Alert Cutoff That Protects Recall

The alert cutoff is an HR decision setting, not an intrinsic property of XGBoost or CatBoost. For model-selection labels $y_i$ and scores $p_i$, candidate cutoffs are taken from the precision-recall curve and limited to a prespecified range. Let $R\left(\tau \right)$, $F_1\left(\tau \right)$, and $A\left(\tau \right)$ denote recall, F1, and accuracy. Given minimum acceptable values $R_{min}$ and $F_{min}$, the rule is

$${\tau }^{*}=arg\underset{\tau }{max}\left\{A\left(\tau \right):R\left(\tau \right)\ge R_{min},F_1\left(\tau \right)\ge F_{min}\right\}.$$

Ties are resolved by F1, precision, and then cutoff. If no cutoff satisfies these conditions, the development procedure falls back to a recall-oriented rule. In plain terms, the model must first find enough future leavers and maintain a minimum balance between missed and false alerts. It may then reduce total errors and HR workload. This order prevents accuracy from selecting an almost all-negative alert list.

The employee-separated search selected HRE-XCB weights of 0.65/0.15/0.20 for Raw-XGB, HRE-XGB, and CatBoost, respectively. After refitting the selected definitions with the fixed audit seeds, the May alert cutoff was 0.317443. The common development objective uses PR-AUC for ranking and requires recall of at least 0.80 and F1 of at least 0.55 before secondary criteria are considered.

3.5. Evaluation Metrics and Probability Interpretation

3.5.1. Measures for the Final Alert List

Let $TP$, $FP$, $TN$, and $FN$ denote correct alerts, false alerts, correct non-alerts, and missed departures at the final chosen cutoff. The reported classification metrics are

$$Accuracy={\displaystyle \frac{TP+TN}{TP+FP+TN+FN}},$$

$$Precision={\displaystyle \frac{TP}{TP+FP}},Recall={\displaystyle \frac{TP}{TP+FN}},$$

and

$$F_1=2{\displaystyle \frac{Precision·Recall}{Precision+Recall}}.$$

Balanced accuracy, the average of sensitivity and specificity, is used as a supplementary diagnostic during broad screening. Accuracy estimates the overall proportion of correct decisions, but it is not the primary indicator of attrition-warning utility. With a low event rate, a large number of correct negatives can conceal poor retrieval of future departures. Recall asks what proportion of employees who later leave were placed on the review list. Precision asks what proportion of alerts correspond to an observed departure and therefore describes review workload. F1 balances recall and precision and penalizes a model that improves one by sacrificing the other.

The business priority is stated before reading later evaluation results. Recall is the first alert-list safety measure because an employee missed by the model cannot receive a timely retention conversation. Precision and alert rate represent HR workload and the possible harm of unnecessary scrutiny. F1 summarizes this tension but does not assign a monetary value to every outcome. Accuracy is retained for completeness and comparison with earlier attrition studies. It is not used alone as proof that the model recognizes future leavers.

3.5.2. Ranking Measures Across Possible Cutoffs

Receiver operating characteristic area under the curve (ROC-AUC) is the probability that a randomly selected future leaver receives a larger score than a randomly selected non-leaver, with ties handled conventionally [26]. It evaluates pairwise ranking across all cutoffs. However, its false-alert rate is divided by the usually large number of non-leavers. ROC-AUC can therefore look favorable even when many employees on the alert list do not leave.

Precision-recall area under the curve (PR-AUC), implemented as average precision, focuses on the less common departure class. Its reference level depends on the departure rate, so values should be interpreted within each evaluation set. In an imbalanced problem, the precision-recall plot can be more informative than the ROC plot [25]. PR-AUC is therefore the main score for ranking quality across possible cutoffs, while ROC-AUC remains a useful supplement. Neither selects the final alert cutoff, and neither shows that a numeric score has a reliable probability meaning.

3.5.3. Reliability of the Numeric Risk Score

The Brier score gives a direct penalty when a predicted number is far from the observed result. It is the mean squared probability error:

$$Brier={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{({\widehat{p}}_i-y_i)}^2.$$

It is a proper scoring rule, meaning that an honestly estimated conditional probability minimizes expected loss. The score reflects both reliability and the ability to separate risk levels, and it is affected by the event rate. Lower is better, but two evaluation sets with different departure rates are not directly interchangeable.

Expected calibration error (ECE) provides an easier group-level view. It places scores into bands $B_1,\dots ,B_M$ and compares the mean prediction with the observed departure frequency:

$$ECE={\displaystyle \sum _{m=1}^M}{\displaystyle \frac{|B_m|}{n}}\left|acc\left(B_m\right)-conf\left(B_m\right)\right|.$$

Here, $acc\left(B_m\right)$ is the observed event share and $conf\left(B_m\right)$ is the mean score in band m. ECE is intuitive but changes with the band definitions and is not itself a proper scoring rule. It is therefore reported with Brier score and reliability plots. Numeric probability quality is a known weak point of predictive analytics, especially after flexible model and cutoff selection [27].

These distinctions determine how individual scores are described. A high PR-AUC indicates useful ordering; it does not mean that a score of 0.70 is a proven 70% personal chance of leaving. A low ECE in one historical month may not persist after changes in management, labor markets, or data recording. Until a genuinely future group accumulates complete outcomes, HRE-XCB scores should support ranking, percentiles, and broad risk bands. Literal personal probabilities require future-data adjustment and continued monitoring.

3.6. Uncertainty, Model Explanation, and Job-Group Checks

3.6.1. Bootstrap Uncertainty

A single metric can exaggerate a small difference when few employees leave. The final model settings are therefore assessed with 5,000 bootstrap resamples in each evaluation. Bootstrap resampling means repeatedly drawing complete employee clusters with replacement to show how much a metric changes under similar samples. Departures and non-departures are drawn separately so that both outcomes remain present. Each draw uses the already fixed scores and cutoff. The 2.5th and 97.5th percentiles form a descriptive 95% interval.

For a metric M, replicate b gives $M^{\left(b\right)}$, and the interval is

$$\left[Q_{0.025}{\left\{M^{\left(b\right)}\right\}}_{b=1}^{5000},Q_{0.975}{\left\{M^{\left(b\right)}\right\}}_{b=1}^{5000}\right].$$

When comparing an alternative setting with the selected research model, both are calculated on the same resampled employees. This paired comparison summarizes $\Delta M^{\left(b\right)}$ and removes variation caused merely by drawing different records. A positive difference favors the alternative for higher-is-better measures. A negative difference favors it for Brier score or ECE. Overlap between two separate intervals is not treated as a formal paired comparison.

The intervals describe repeated sampling within the observed evaluation data. They do not include uncertainty from the single administrative snapshot, missing historical predictors, the number of models tried, a future economic change, or use in another enterprise. Grouping employees during development addresses only part of the dependence among repeated monthly records. A difference of one or two thousandths is therefore reported as a point-estimate difference unless its practical size and paired interval support a stronger statement. Guidance for limited prediction datasets likewise cautions against fine-grained claims [29,30].

3.6.2. Which Variables the Model Uses

The final HRE-XCB model is explained by asking how much ranking quality falls when one source variable is randomly rearranged. This method is called grouped permutation importance. Let S be the original ROC-AUC and let $S_{j,r}^{perm}$ be ROC-AUC after predictor j is rearranged in repetition r. Importance is

$$I_j={\displaystyle \frac{1}{R}}{\displaystyle \sum _{r=1}^R}(S-S_{j,r}^{perm}).$$

Rearranging a source variable breaks its observed association with departure while leaving the fitted model unchanged. All derived columns from one source concept are treated together, so dozens of one-hot or HRE columns do not receive misleadingly separate credit. A larger decrease means that the fixed model relied more strongly on that variable to order employees in the current evaluation. It does not establish that the variable caused turnover.

Correlated predictors can share or mask permutation importance. For example, job, job family, department, and center overlap structurally. Permuting one while retaining the others may understate its total information, while an unrealistic permutation can create combinations absent from the organization. Importance is therefore a model-behavior diagnostic, not an estimate of a feature’s causal effect or managerial responsibility.

LIME approximates one prediction with a simple local model [21]. SHAP assigns additive contribution values based on Shapley-value theory, and tree-specific SHAP methods can summarize these local explanations [35]. Both can be useful, but each explanation must be linked to the model that produced it. A component-level SHAP value is not an explanation of the three-branch HRE-XCB score, particularly when categorical tree splits require a different explainer. The paper therefore uses grouped permutation importance for the ensemble and confines LIME or SHAP interpretation to the component actually analyzed. Reviews of explainable AI likewise stress that an explanation should match its user and decision context [36].

3.6.3. Job-Group and Sensitive-Feature Checks

Metrics are disaggregated by financially and operationally meaningful job families, including sales advisers, instructors and curriculum research, operations and service, comprehensive management, and back-office functions. Each cell reports its number of observations, positives, and negatives alongside recall, precision, and F1. A minimum of 20 observations, three positives, and three negatives is required for a metric cell to be shown as calculable. This is only a descriptive threshold; it does not guarantee adequate power or stable ranking. Groups with very few departures are explicitly marked as insufficient for comparative conclusions.

Protected or potentially sensitive attributes, such as sex, marital status, household registration, or residential region, require a separate governance review. The fact that a field predicts an outcome does not make its use legitimate. It may reflect social structure or historical organizational practice. The system must compare missed and false alerts across meaningful groups, document whether each sensitive field is necessary, and test a restricted model when legal or ethical review requires exclusion. No group result may become an automatic adverse-employment rule.

3.7. Data, Employee Separation, and Experimental Protocol

The source contained 1,397 de-identified employee records. Seven reference dates produced 6,935 reconstructed monthly records for voluntary turnover within 90 days. Reconstruction expresses how one administrative export would be queried at different dates; it does not turn that export into genuine historical snapshots. Missing values were retained through training-learned numeric medians and explicit categorical missing levels. No record was removed merely because a noncritical field was absent. Numeric transformations, category vocabularies, hierarchy summaries, and model parameters were learned inside the development data.

The employee, not the row, was the independent unit. Each source row had a unique internal identifier; names were never used as employee keys. Among 928 employees evaluable in June 2022, 278 were selected with seed 20261640 without reading outcomes. Every December–May record for those employees was excluded before feature screening, hierarchy encoding, hyperparameter tuning, ensemble weighting, and cutoff selection. Development contained 3,735 training rows from December 2021 through April 2022 and 710 May selection rows. The retrospective June evaluation contained 278 separated employees and 24 departures; employee overlap was zero. June outcomes were accessible during the research, so this evaluation is not described as blinded, prospective, external, or untouched.

The full catalog contained 22 general-purpose classifiers. Five strong baselines–logistic regression, random forest, Extra Trees, histogram gradient boosting, and XGBoost–received 40 randomized configurations each under four employee-group folds. The HRE-XCB search evaluated 152 base structures, then reevaluated 18 finalists across four employee-group folds and three seeds. Candidate selection used May data only. The selected weights were 0.65 for Raw-XGB, 0.15 for HRE-XGB, and 0.20 for CatBoost; after the selected component definitions were refitted with the fixed audit seeds, the May alert cutoff was 0.317443. All evaluation predictions were generated without retuning.

PR-AUC was the primary ranking endpoint because the evaluation positive rate was 8.63%. Recall, precision, F1, accuracy, and alert rate described the selected alert list. Brier score, expected calibration error, calibration intercept, calibration slope, and reliability curves assessed whether numeric scores resembled probabilities. Model differences were evaluated with 5,000 paired employee-cluster bootstrap resamples, stratified by outcome. Field-timing sensitivity held the fitted HRE-XCB structure fixed while masking organization fields whose historical values could not be confirmed. Forty-two voluntary departures without a departure date were excluded from monthly 90-day outcome reconstruction and are reported in the sample flow.

Figure 2. Time periods and employee-level separation.

Figure 2. Time periods and employee-level separation.

Development uses December 2021 through April 2022 for training and May 2022 for model selection. The 278 evaluation employees contribute no December–May records to any development step. June 2022 contains 278 separated employees and 24 departures. June outcomes were available during the study, so this is a retrospective evaluation rather than a blinded, prospective, or external test.

3.8. Models Compared and Their Purpose

The broad screen covers different inductive biases. Its purpose is not to claim that every algorithm received an unlimited search budget. It establishes whether a linear boundary, local neighborhood, probabilistic independence assumption, margin, single tree, randomized ensemble, sequential booster, native-category learner, or neural interaction surface is plausible for the same target. The strongest families then enter the equal-budget comparison. This two-level design preserves breadth without allowing weak default settings to determine the final ranking.

Table 2. Classifiers compared and the purpose of each comparison.

Implemented classifier	Main learning assumption	Limited-data control	Role in the study evidence
DummyPrior	no feature information	none required	exposes majority-class accuracy
Logistic regression	regularized linear log-odds	L2 shrinkage and standardized numeric inputs	tuned linear reference
SGDLogLoss	linear log-odds learned incrementally	regularization and controlled learning schedule	optimization-sensitive linear check
Shrinkage LDA	shared class covariance and linear separation	covariance shrinkage	low-variance discriminant check
GaussianNB	conditionally independent Gaussian numeric variables	small parameter count	distributional probabilistic check
BernoulliNB	conditionally independent binary indicators	small parameter count	sparse-indicator check
KNN	nearby records share outcomes	scaling and bounded neighborhood size	tests local similarity and identity-recurrence risk
Linear SVM	maximum-margin linear boundary	margin penalty	linear boundary without calibrated probability output
RBF SVM	smooth nonlinear similarity	penalty and kernel-width tuning	flexible margin check
Decision tree	recursive axis-aligned partitions	depth and leaf-size limits	interpretable high-variance reference
Bagged trees	averaging reduces tree variance	resampling and ensemble averaging	isolates the value of bagging
Random forest	bagging plus feature randomization	minimum leaf size and feature subsampling	tuned employee-separated-sample ranking reference
Extra Trees	strongly randomized tree partitions	depth, leaf size, and averaging	alternative variance-reduction reference
Balanced Random Forest	balanced subsamples per tree	ensemble averaging	imbalance-aware recall check
EasyEnsemble	multiple balanced majority subsets	aggregation across subsets	under-sampling ensemble check
AdaBoost	sequential emphasis on errors	shallow weak learners and learning rate	reweighting-based boosting check
Gradient boosting	additive correction of loss gradients	shallow trees and shrinkage	classical boosting reference
Histogram gradient boosting	binned additive trees	leaf and L2 constraints	tuned probability-error reference
XGBoost	regularized second-order boosting	depth, sampling, and L1/L2 penalties	tuned general boosting reference
LightGBM	histogram-based leaf-wise growth	leaf-count and minimum-data limits	efficient boosting check
CatBoost	ordered categorical statistics and symmetric trees	ordered boosting and leaf regularization	category-native high-recall reference
MLP	nonlinear interactions in hidden layers	small network, weight decay, and early stopping	flexible neural reference

Table 2. Classifiers compared and the purpose of each comparison.

Implemented classifier	Main learning assumption	Limited-data control	Role in the study evidence
DummyPrior	no feature information	none required	exposes majority-class accuracy
Logistic regression	regularized linear log-odds	L2 shrinkage and standardized numeric inputs	tuned linear reference
SGDLogLoss	linear log-odds learned incrementally	regularization and controlled learning schedule	optimization-sensitive linear check
Shrinkage LDA	shared class covariance and linear separation	covariance shrinkage	low-variance discriminant check
GaussianNB	conditionally independent Gaussian numeric variables	small parameter count	distributional probabilistic check
BernoulliNB	conditionally independent binary indicators	small parameter count	sparse-indicator check
KNN	nearby records share outcomes	scaling and bounded neighborhood size	tests local similarity and identity-recurrence risk
Linear SVM	maximum-margin linear boundary	margin penalty	linear boundary without calibrated probability output
RBF SVM	smooth nonlinear similarity	penalty and kernel-width tuning	flexible margin check
Decision tree	recursive axis-aligned partitions	depth and leaf-size limits	interpretable high-variance reference
Bagged trees	averaging reduces tree variance	resampling and ensemble averaging	isolates the value of bagging
Random forest	bagging plus feature randomization	minimum leaf size and feature subsampling	tuned employee-separated-sample ranking reference
Extra Trees	strongly randomized tree partitions	depth, leaf size, and averaging	alternative variance-reduction reference
Balanced Random Forest	balanced subsamples per tree	ensemble averaging	imbalance-aware recall check
EasyEnsemble	multiple balanced majority subsets	aggregation across subsets	under-sampling ensemble check
AdaBoost	sequential emphasis on errors	shallow weak learners and learning rate	reweighting-based boosting check
Gradient boosting	additive correction of loss gradients	shallow trees and shrinkage	classical boosting reference
Histogram gradient boosting	binned additive trees	leaf and L2 constraints	tuned probability-error reference
XGBoost	regularized second-order boosting	depth, sampling, and L1/L2 penalties	tuned general boosting reference
LightGBM	histogram-based leaf-wise growth	leaf-count and minimum-data limits	efficient boosting check
CatBoost	ordered categorical statistics and symmetric trees	ordered boosting and leaf regularization	category-native high-recall reference
MLP	nonlinear interactions in hidden layers	small network, weight decay, and early stopping	flexible neural reference

Only logistic regression, random forest, Extra Trees, histogram gradient boosting, and XGBoost received the final equal budget of 40 randomized configurations because they represented strong distinct families under the common encoded pipeline. CatBoost was tuned in the category-native branch, and HRE-XCB was developed through the three-branch search. The main table separates these comparably developed candidates from the predefined 22-model screen. The full screen is reported in the supplement with its lower tuning status made explicit.

3.9. How Hierarchical Risk Encoding Stabilizes Rare Categories

For category c nested in parent group g, let $n_c$ be the number of development employees in the category, ${\overline{y}}_c$ their observed event rate, ${\mu }_g$ the parent estimate, and $\alpha$ the smoothing strength. HRE uses

$$r_c={\displaystyle \frac{n_c{\overline{y}}_c+\alpha {\mu }_g}{n_c+\alpha }}={\lambda }_c{\overline{y}}_c+(1-{\lambda }_c){\mu }_g,{\lambda }_c={\displaystyle \frac{n_c}{n_c+\alpha }}.$$

The estimator has three useful properties. It remains between the category and parent estimates; it approaches the parent when evidence is sparse; and it approaches the category rate as evidence grows. Unknown jobs first fall back to job family, unknown departments to center type, and insufficient parent groups continue toward the global prior. The parent-based default therefore degrades gradually rather than replacing every unseen category with an arbitrary zero or extreme code.

HRE outputs the smoothed rate, its log-odds transform, and $log(1+n_c)$, where $n_c$ is the raw support count learned from the employee-excluding encoding-fit data. This support value is fixed after fitting and is independent of scoring-batch size. It distinguishes a well-observed rate from the same numerical value obtained mainly from a broader group. The downstream shallow XGBoost branch can learn whether to trust the rate conditionally on its support.

The essential leakage control operates at employee level. For grouped fold k, the encoding table ${\Theta }^{(-k)}$ is estimated without any monthly record from employees in that fold. All records of employee i are then transformed with the same outside-employee encoding table. Row-wise leave-one-out encoding is insufficient because another month from the same employee can contain the same stable categories and related outcome. During evaluation, the encoding table is fitted only from development employees and applied unchanged to the separated employee. Smoothing controls variance; employee exclusion controls self-information. They solve different problems.

Employee-equal weighting addresses repeated exposure. If employee i contributes $m_i$ monthly records, the base weight of each record is proportional to $1/m_i$, so one employee’s total baseline contribution remains approximately one. Class weighting changes the cost of positive and negative errors, while recency decay changes the relative contribution of older months. These factors are multiplied only inside development and must not be estimated from strict outcomes.

3.10. Model Search, Selection, and Rejection Rules

The innovation search begins with 152 base structures: 32 Raw-XGB structures, 96 HRE-XGB structures, and 24 native CatBoost structures. The first stage uses employee-group folds to screen the full set. Six candidates per branch then form 18 finalists, which are reevaluated with four employee-group folds and three seeds. Nonnegative weighted averages preserve component-score bounds and add only two free weights. The selected combination assigns 0.65 to Raw-XGB, 0.15 to HRE-XGB, and 0.20 to CatBoost.

Candidate rejection is part of the method. A model is not retained merely because one metric rises. It must first meet the May recall requirement of 0.80 and F1 requirement of 0.55. PR-AUC then evaluates ranking, while accuracy, precision, alert rate, Brier score, fold dispersion, and model complexity provide secondary evidence. A change that raises recall by alerting almost everyone, improves Brier while weakening ranking, or wins only under one seed is recorded as a trade-off for one objective rather than a general improvement.

The same principle governs the evaluation comparison. Each candidate uses its development-selected cutoff on the same 278 employees. Paired resampling draws complete employee clusters within outcome class so that an interval describes the difference between two models on matched employees. This is more informative than comparing two independent confidence intervals. With 24 positive outcomes, however, intervals remain too wide to establish a permanent population order.

The implementation records the employee split, candidate identifier, random seed, model settings, and evaluation results. Figures and manuscript tables come from the same saved results. Reproducibility means that the stated procedure can reproduce the reported results for this sample. It does not turn a retrospective study into a prospective validation.

4. Results

4.1. Model Development Under Employee Separation

The protocol separated employees before candidate screening. Of 928 employees evaluable in June, 278 were sampled with a fixed seed without reading outcomes. Their December–May records were excluded from encoding, fitting, search, weighting, and cutoff selection. Development then used 3,735 training records and 710 May model-selection records. The June evaluation contained 278 separated employees and 24 departures, with zero employee overlap. June outcomes were available during the study, so the evidence is retrospective rather than prospective or external.

The HRE-XCB search covered 152 base structures and reevaluated 18 finalists with employee-group folds and three random seeds. After the selected definitions were refitted with the fixed audit seeds, HRE-XCB achieved May recall 0.8120, F1 0.6859, and PR-AUC 0.7576. Extra Trees achieved F1 0.7917 and PR-AUC 0.8445; native CatBoost achieved 0.7041 and 0.8069. The May period selected settings only and was not used as final evidence.

Figure 3. Model selection under common employee-separated development rules.

Figure 3. Model selection under common employee-separated development rules.

Points compare F1 and PR-AUC in the May model-selection period. The highlighted ensemble is HRE-XCB. The chart describes development-stage selection and is not evidence of future superiority.

4.2. Results for Employees Kept Out of Model Development

Table 3 and Figure 4 report results for the separated employees. Among the equally tuned shortlist, random forest produced the highest PR-AUC (0.3772). Extra Trees produced the highest F1 (0.4324), precision (0.3200), and accuracy (0.8489). Histogram gradient boosting achieved recall 0.7917, F1 0.4270, and the lowest Brier score (0.0752). HRE-XCB achieved recall 0.7500, F1 0.3750, PR-AUC 0.3477, and Brier 0.1023. Native CatBoost reached recall 0.7917 but alerted 39.21% of employees. DummyPrior obtained 91.37% accuracy while detecting no departures, directly showing why accuracy cannot be the primary endpoint.

Table 3. Retrospective employee-separated performance under comparable development rules.

Model	Accuracy	Precision	Recall	F1	PR-AUC	Brier
RF	.8094	.2542	.6250	.3614	.3772	.1134
HistGB	.8165	.2923	.7917	.4270	.3756	.0752
Extra Trees	.8489	.3200	.6667	.4324	.3622	.1074
XGBoost	.8022	.2615	.7083	.3820	.3611	.1305
HRE-XCB	.7842	.2500	.7500	.3750	.3477	.1023
LR	.7806	.2394	.7083	.3579	.3319	.1306
CatBoost	.6583	.1743	.7917	.2857	.3127	.1684
Prior-only	.9137	.0000	.0000	.0000	.0863	.1318

Table 3. Retrospective employee-separated performance under comparable development rules.

Model	Accuracy	Precision	Recall	F1	PR-AUC	Brier
RF	.8094	.2542	.6250	.3614	.3772	.1134
HistGB	.8165	.2923	.7917	.4270	.3756	.0752
Extra Trees	.8489	.3200	.6667	.4324	.3622	.1074
XGBoost	.8022	.2615	.7083	.3820	.3611	.1305
HRE-XCB	.7842	.2500	.7500	.3750	.3477	.1023
LR	.7806	.2394	.7083	.3579	.3319	.1306
CatBoost	.6583	.1743	.7917	.2857	.3127	.1684
Prior-only	.9137	.0000	.0000	.0000	.0863	.1318

RF, HistGB, and LR denote tuned random forest, histogram gradient boosting, and logistic regression, respectively.

Figure 4. Retrospective employee-separated comparison under comparable tuning.

Figure 4. Retrospective employee-separated comparison under comparable tuning.

Each cell prints the raw metric value. Color represents the within-metric rank, with Brier ordered in the lower-is-better direction. The different darkest cells show why the data do not support a universal winner.

4.3. HRE-XCB Component Ablation

The development procedure selected weights of 0.65 for Raw-XGB, 0.15 for HRE-XGB, and 0.20 for CatBoost. Raw-XGB alone achieved PR-AUC 0.3481 and F1 0.4091 on separated employees. Raw-XGB plus HRE-XGB achieved 0.3492 and 0.3871. The complete HRE-XCB achieved 0.3477 and 0.3750. These differences are too small and uncertain to establish a ranking advantage, and the full ensemble did not outperform Raw-XGB on F1 or PR-AUC. HRE remains a defined and reproducible representation, but the data do not establish incremental predictive value. A runtime audit also scored all 278 evaluation employees both together and one at a time; the maximum absolute probability difference was 0.0, confirming that HRE-XCB predictions no longer depend on scoring-batch size.

Figure 5. HRE-XCB structure selected within employee-separated development.

Figure 5. HRE-XCB structure selected within employee-separated development.

The component experiment reports separated-employee PR-AUC and F1 for all single branches, pairwise combinations, and the complete ensemble. It rejects a claim that every added branch improved performance.

4.4. Decision-Specific Model Roles

The empirical results support several roles rather than one permanent ranking. Random forest is the ranking reference within the equally tuned shortlist. Extra Trees offers the highest observed F1 and the smallest review list among the higher-performing candidates. Histogram gradient boosting offers high recall and the lowest probability error. HRE-XCB remains a representation experiment, not the operational winner. These roles are provisional because their PR-AUC intervals overlap. The separate preset-model screen is reported independently and is not merged into this ranking.

Figure 6. Model roles under different human-resource decision objectives.

Figure 6. Model roles under different human-resource decision objectives.

The four roles are conditional on the observed employee-separated sample. They do not constitute permanent rankings or authorization for automated employment action.

4.5. Paired Uncertainty

Five thousand paired employee-cluster bootstrap samples compared each candidate with HRE-XCB. Random forest differed by -0.0136 in F1 (95% interval -0.1078–0.0719), 0.0295 in PR-AUC (-0.0377–0.1061), and 0.0111 in Brier (0.0033–0.0187). Extra Trees differed by 0.0574 in F1 (-0.0375–0.1444) and 0.0145 in PR-AUC (-0.0709–0.0939). Histogram gradient boosting differed by 0.0520 in F1 (-0.0140–0.1235), 0.0278 in PR-AUC (-0.0250–0.0688), and -0.0271 in Brier (-0.0357 to -0.0187). No tuned candidate had a PR-AUC difference interval excluding zero. The Brier intervals show objective-specific differences: HRE-XCB had lower probability error than random forest, whereas histogram gradient boosting had lower error than HRE-XCB.

Figure 7. Paired bootstrap differences relative to HRE-XCB.

Figure 7. Paired bootstrap differences relative to HRE-XCB.

Points are candidate-minus-HRE-XCB differences and bars are 95% employee-cluster bootstrap intervals. Positive favors the candidate for F1 and PR-AUC; negative favors the candidate for Brier score.

4.6. Feature Timing and Probability Calibration

The one-time administrative export could not confirm every field at each historical reference date. The uncertain fields were center, department, job, job family, job level, household registration, marital status, political affiliation, and residence. The sensitivity experiment fixed the architecture, hyperparameters, and ensemble weights, refitted the model after masking these fields in the development data, and selected a new cutoff on masked May data (0.314602). On the separated employees, PR-AUC fell from 0.3477 to 0.3120, recall from 0.7500 to 0.5417, and F1 from 0.3750 to 0.3095. Brier improved from 0.1023 to 0.0938. Lower probability error therefore did not imply better ranking or coverage.

Calibration was imperfect. HRE-XCB’s mean score was 0.2333 against an observed rate of 0.0863, with ECE 0.1470, calibration intercept -1.5976, and slope 0.8047. HistGB was closer, with mean score 0.1386, ECE 0.0627, intercept -0.7672, and slope 0.7344. Reliability curves support ranking or broad risk bands, not exact individual probabilities.

Figure 8. Feature-timing sensitivity and probability reliability.

Figure 8. Feature-timing sensitivity and probability reliability.

The left panel compares the full-feature model with the restricted-feature refit under fixed architecture, hyperparameters, and weights; each uses a May-selected cutoff. The right panel compares predicted scores with observed rates in equal-count bins. Neither panel supports causal claims or exact personal probabilities.

4.7. Integrated Empirical Interpretation

Figure 9 integrates ranking with alert-list behavior. HRE-XCB detected 18 of 24 departures with 72 alerts. HistGB detected 19 with 65 alerts, while Extra Trees detected 16 with 50 alerts. The appropriate model depends on ranking, coverage, review capacity, and probability error; current intervals do not justify a permanent winner.

Figure 9. Integrated retrospective results for employees kept out of model development.

Figure 9. Integrated retrospective results for employees kept out of model development.

The left panel orders models by PR-AUC. The right panel shows the recall–precision trade-off. The evaluation contains 278 separated employees and 24 departures and does not establish future, cross-job, or cross-enterprise superiority.

4.8. Full Algorithm Screen and Review Workload

All 22 predefined classifiers were also fitted under the same employee split and cutoff-selection rule. Gradient boosting had the highest screening PR-AUC (0.4229), followed by bagged trees (0.4015). These screening values are reported for coverage, not as a final ranking, because the models did not receive the same 40-trial search as the tuned shortlist. DummyPrior produced PR-AUC equal to prevalence (0.0863) and zero recall. The complete screen, including all threshold, ranking, and calibration measures, is supplied with the reproducibility files.

Table 4. Equal-budget tuning results for the five strong baselines.

Tuned baseline	Search trials	Best grouped-CV PR-AUC	Selected complexity control
Logistic regression	40	.5770	L1 penalty, C=.1473
Random forest	40	.6199	depth 8, minimum leaf size 4, square-root feature sampling
Extra Trees	40	.6195	unrestricted depth, minimum leaf size 5, square-root feature sampling
Histogram gradient boosting	40	.6296	17 leaf nodes, minimum leaf size 8, learning rate .0239, L2=.2206
XGBoost	40	.6142	depth 5, minimum child weight 3, row/column sampling, L1/L2 penalties

Table 4. Equal-budget tuning results for the five strong baselines.

Tuned baseline	Search trials	Best grouped-CV PR-AUC	Selected complexity control
Logistic regression	40	.5770	L1 penalty, C=.1473
Random forest	40	.6199	depth 8, minimum leaf size 4, square-root feature sampling
Extra Trees	40	.6195	unrestricted depth, minimum leaf size 5, square-root feature sampling
Histogram gradient boosting	40	.6296	17 leaf nodes, minimum leaf size 8, learning rate .0239, L2=.2206
XGBoost	40	.6142	depth 5, minimum child weight 3, row/column sampling, L1/L2 penalties

Alert counts make the management trade-off concrete. HRE-XCB detected 18 of 24 departures and issued 72 alerts. HistGB detected 19 with 65 alerts, while Extra Trees detected 16 with 50. Random forest detected 15 with 59 alerts. XGBoost and logistic regression each detected 17, with 65 and 71 alerts, respectively. CatBoost detected 19 with 109 alerts. No monetary cost score was calculated because the dataset contains no defensible cost for a missed departure or an unnecessary review.

Table 5. Departure coverage and review workload for separated employees.

Model	True positives	False negatives	Total alerts	False-positive alerts	Alert rate
Random forest	15	9	59	44	21.22%
HistGB	19	5	65	46	23.38%
Extra Trees	16	8	50	34	17.99%
XGBoost	17	7	65	48	23.38%
HRE-XCB	18	6	72	54	25.90%
Logistic regression	17	7	71	54	25.54%
CatBoost	19	5	109	90	39.21%

Table 5. Departure coverage and review workload for separated employees.

Model	True positives	False negatives	Total alerts	False-positive alerts	Alert rate
Random forest	15	9	59	44	21.22%
HistGB	19	5	65	46	23.38%
Extra Trees	16	8	50	34	17.99%
XGBoost	17	7	65	48	23.38%
HRE-XCB	18	6	72	54	25.90%
Logistic regression	17	7	71	54	25.54%
CatBoost	19	5	109	90	39.21%

The development-to-evaluation change was material. May prevalence was 16.48%, whereas the separated-employee prevalence was 8.63%. Precision, F1, alert volume, and probability calibration therefore changed even when cutoffs remained fixed. This result supports monitoring prevalence and workload together with predictive measures. It does not justify changing a cutoff after viewing employee identities or outcomes.

5. Discussion

5.1. Interpretation of the Empirical Evidence

The evidence supports decision-specific model choice. HRE-XCB was competitive in development, but its complete structure did not outperform Raw-XGB on separated employees. Random forest ranked best by tuned PR-AUC, Extra Trees by F1, and histogram gradient boosting by Brier score. Their PR-AUC intervals overlapped. The paper therefore treats the evaluation design, not universal ensemble superiority, as the primary contribution.

For risk ranking, PR-AUC is the primary endpoint because departures are uncommon. For an actionable alert list, recall measures missed high-risk employees, while precision and alert rate measure review burden. F1 is a secondary summary at one cutoff. Brier score and calibration diagnostics are required only when numeric scores are given probability meaning. Under the current evidence, scores should be reported as ranks, percentiles, or broad risk bands rather than exact personal departure probabilities.

The management contribution is similarly bounded. Administrative job and tenure fields do not measure job satisfaction, commitment, embeddedness, organizational support, psychological contract, or job demands and resources. Those theories identify plausible mechanisms and responsible interventions; they are not empirically tested by proxy fields in this study. The model supports confidential, non-punitive review. It cannot establish why an employee will leave or authorize dismissal, pay reduction, promotion restriction, or other adverse action.

The 24 positive outcomes remain insufficient for stable job-specific assessment. The study does not establish job-group fairness or justify job-specific cutoffs. Random forest, Extra Trees, and HistGB are the stronger candidates for ranking, alert efficiency, and probability error, respectively. HRE-XCB remains useful only as a reproducible test of organizational representation.

The study covers one enterprise, a short period, and one administrative export. Dates are missing for 42 voluntary departures, which were excluded from monthly outcome reconstruction, and the employee-separated evaluation contains only 24 positive outcomes. June outcomes were available during the study, and no external or prospective cohort was available. Further evidence requires genuine month-end snapshots, a protocol fixed before outcomes mature, an institutional ethics determination where required, and evaluation of fixed models in a future quarter or another enterprise.

5.2. Credibility, Scope, and Management Implications

The study distinguishes four limits. Employee separation and employee-excluding encoding strengthen internal predictive credibility. The timing-masked analysis shows that historical field availability is only partly established. Administrative fields do not directly measure satisfaction, commitment, embeddedness, support, psychological contract, or job demands and resources. One enterprise and one retrospective period cannot establish transport to another time, job mix, or organization.

The model comparison has a practical consequence: selection must follow the decision. PR-AUC is the primary ranking measure; recall, precision, and alert rate define the review list; Brier and calibration diagnostics govern probability language. The present scores are suitable for ranking, percentiles, or broad risk bands. They are not validated personal departure probabilities. No score may trigger dismissal, pay reduction, promotion restriction, or other adverse action.

Before operational use, the organization should freeze the data dictionary, model, cutoff, and review capacity; score a genuine future cohort; and wait for the 90-day outcomes. Monitoring should report prevalence, alert rate, recall, precision, PR-AUC, Brier, calibration, missingness drift, new categories, and subgroup event counts. The model should be paused when data definitions change, review capacity is exceeded, probability error deteriorates, or subgroup evidence shows a material coverage failure.

Future work should preserve genuine month-end snapshots and add validated measures of job demands, resources, support, embeddedness, and turnover intention. A prospective design can then separate prediction from administrative data, the incremental value of psychological measures, and the causal effect of retention actions. Those questions cannot be answered by the current retrospective data.

6. Conclusions

The study delivers a reproducible evaluation framework for 90-day voluntary-turnover warning in finance and taxation education. Employee separation, comparable tuning, full algorithm screening, component ablation, employee-cluster uncertainty, calibration, and field-time sensitivity provide the basis for interpretation. HRE-XCB is a testable representation, not an established improvement. Random forest, Extra Trees, and histogram gradient boosting led different observed endpoints.

These findings support risk ranking and supportive HR review, not automated employment action or exact individual probabilities. They also show why a later-month score or a large grid search cannot substitute for independent employees and genuinely historical features. Prospective and external evidence is still required before any model is treated as a validated organizational decision system.