Robust Customer Lifetime Value Estimates Under Censoring: A Conceptual Method for Domain-Aware, Mixed-Frequency, and Stability-Sensitive Prediction

1. Introduction and Research Question

Customer lifetime value is used to prioritize retention spending, allocate service resources, and compare customer segments by expected future revenue. In practice, however, CLV is often estimated from incomplete account histories. A customer may still be active when the observation window closes, or a contract may continue beyond the last available transaction. Under these conditions, a naïve estimator that replaces unseen future cash flows with zero or with a fixed-horizon average will distort value estimates. The central research question is therefore estimand-specific: how should one estimate horizon-specific future cumulative value, defined at a stated prediction origin, when some customer trajectories are right-censored, the available signals arrive at different frequencies, and some behavioral or textual features are unstable across domains? This formulation fixes the object of prediction before model choice and distinguishes administrative observability from the customer’s latent economic lifetime.

The evidence scope is intentionally narrow. Rather than claiming a systematic review, this manuscript offers a targeted narrative synthesis of methodological foundations that can support censored CLV estimation. The selected sources show that financial customer analytics benefits from domain-aware automated feature engineering, that cross-domain representations can be stabilized through adaptation, that multimodal and mixed-frequency forecasting can integrate heterogeneous signals, and that finite-length sequence estimators and robustness checks can reduce instability in small or noisy samples. These strands are not themselves censoring-specific CLV evidence, but together they justify a precise conceptual design for a censored CLV pipeline. [1] [2] [3] [4] [5] [6]

2. Prior Work Relevant to Censored CLV Estimation

The most direct application source in the selected set is the domain-aware automated framework for financial customer analytics. That work states that generic AutoML systems often miss marketing-specific constructs such as RFM relationships, CLV trajectories, and engagement sequences, and it proposes automatic data modality recognition, feature engineering, model selection, and pipeline assembly for customer analytics tasks. For the present topic, the relevant implication is not a reported superiority claim but a design principle: CLV estimation should be embedded in a pipeline that can generate marketing-relevant features from raw customer data without requiring manual reconstruction at every deployment site. [1]

Two additional methodological strands are especially transferable. First, multimodal forecasting work in power marketing shows that combining consumption data with external covariates through a dual-branch architecture can improve the modelling of short-term fluctuations and long-term trends, while a causal attention component increases interpretability. Second, unrestricted mixed-frequency regression demonstrates that when predictors are observed at different temporal resolutions, preserving their native frequency can outperform simple aggregation. For censored CLV, these findings motivate a model that retains event-level transaction history, daily or weekly service activity, and lower-frequency calendar or macro signals instead of forcing all inputs into a single coarse time grid. [3] [4]

A further methodological lesson comes from finite-length sequence correlation modelling. The housing-price spillover study argues that standard correlation or transfer-entropy estimators can become unstable when sequences are short, and that bias correction is needed for finite observations. This is relevant to CLV because many customers have short observed histories or few purchases, precisely the setting in which unstable dependence estimates can contaminate predicted future value. In parallel, cross-domain sentiment analysis and brand-polarity scoring show that transfer across heterogeneous sources benefits from domain adaptation and explicit stability assessment rather than reliance on one fixed representation. [5] [2] [6]

3. Problem Definition and Variables

Let i index customers and t index discrete time on a declared analysis grid, such as days, weeks, or months. For each customer, let Y_{it} denote realized net value at time t, defined consistently as revenue minus variable service cost, contribution margin, or another prespecified accounting measure. Let A_i be the administrative end of observable data for customer i, and let \tau_i be the unobserved terminal time at which value generation ceases. The observed follow-up time is U_i = min(\tau_i, A_i), with event indicator \Delta_i = 1(\tau_i \le A_i). Thus, \Delta_i = 0 denotes right censoring. For a prediction origin s \le U_i and horizon H, define the target quantity as CLV_i(s,H) = \sum_{h=1}^{H} Y_{i,s+h}. The outcome Y_{i,s+h} is treated as observed only when the record for period s+h is available and passes the study’s prespecified data-quality checks; if operational systems allow intermittent internal gaps before A_i, an additional measurement indicator M_{i,s+h} must be recorded and the analysis uses O_{i,s+h}=1(A_i \ge s+h)M_{i,s+h} as the observation indicator. In the default design, the manuscript assumes no intermittent internal gaps after preprocessing, so M_{i,s+h}=1 and O_{i,s+h}=1(A_i \ge s+h). The target is therefore partially observed whenever O_{i,s+h}=0 for any h within the horizon.

The feature set is grouped into four blocks. Transactional features include recency, frequency, monetary intensity, average basket value, recurrence intervals, and volatility of spending. Exposure-adjusted features include past-window observation coverage, missingness patterns and interruption indicators known at the prediction origin. Future-horizon observation coverage is used only for weighting and evaluation, never as an input feature. Behavioral features include service usage, digital engagement, complaint load, and response latency. Optional textual or categorical features include support tickets, review sentiment, and channel-specific interaction descriptors. To preserve interpretability, all derived features are computed on a declared window and with explicit censoring marks. [1] [2] [6]

4. Proposed Methodology

The proposed estimator is a censoring-aware, domain-adaptive CLV learner with three linked stages and one common notation. Stage 1 performs automatic feature generation from raw records. Transaction histories are mapped to recency-frequency-monetary summaries, cohort age, purchase intervals, and exposure-adjusted activity rates. Text or review inputs, if available, are converted into stable sentiment or topic scores using a domain-adapted encoder rather than a generic off-the-shelf representation. Stage 2 estimates the observability function for the future outcome path from origin s, not a separate generic uncensoring probability. Specifically, for customer i and horizon step h, the model estimates G_i(s+h \mid s, X_i)=P\{O_{i,s+h}=1 \mid X_i(s)\}, where X_i(s) contains only information available at origin s and O_{i,s+h} is the observation indicator defined in Section 3. Stage 3 predicts future value with a censoring-adjusted objective that uses \hat{G}_i(s+h \mid s, X_i) to reweight observed outcomes. This resolves the earlier ambiguity between E_{it} and G_i(s+h): the default implementation uses the horizon-specific observability probability G throughout. [1] [2] [6]

A default operational rule is to fit the value model with horizon-specific inverse-probability-of-observation weights and to treat any augmented estimator as a secondary sensitivity analysis rather than as automatically doubly robust. For customer i, origin s, and step h, let G_i(s+h \mid s, X_i)=P\{O_{i,s+h}=1 \mid X_i(s)\}. The primary training loss over observed outcomes is L = \sum_i \sum_{h=1}^{H} O_{i,s+h} w_{ih}(Y_{i,s+h}-\hat{m}_h(X_i(s)))^2, where \hat{m}_h(X_i(s)) approximates E[Y_{i,s+h}\mid X_i(s)] and w_{ih}=1/max(\hat{G}_i(s+h \mid s, X_i),\varepsilon). The default choice is \varepsilon=0.05, fixed before evaluation; sensitivity analysis should repeat estimation with \varepsilon in {0.02, 0.10}. If an augmented estimator is examined, it should be defined at the target-estimation stage by the horizon-specific estimating equation \hat{\mu}_h = n^{-1}\sum_i \left[\hat{m}_h(X_i(s)) + O_{i,s+h}\{Y_{i,s+h}-\hat{m}_h(X_i(s))\}/max(\hat{G}_i(s+h \mid s, X_i),\varepsilon)\right], or by an equivalent cross-fitted augmented inverse-probability procedure. That estimator may have double-robust properties only under the formal conditions required by the chosen estimating equation, sample-splitting scheme, and target parameter. This manuscript therefore does not claim that regressing pseudo-outcomes by squared error is sufficient to establish double robustness. Any augmented analysis should be reported separately from the primary weighted learner, with its identification assumptions and implementation details fixed before outcome comparison.

The learner itself can be implemented as a hybrid sequence model. One branch ingests transaction and event sequences with a recurrent, temporal convolutional, or state-space encoder; the second branch ingests exogenous covariates at their native frequency. Fusion occurs through gated concatenation or attention, followed by a value head that outputs the expected undiscounted future margin over the fixed horizon H. This structure is consistent with the logic of multimodal forecasting, but the present manuscript restricts its purpose to CLV under censoring and does not claim superiority of any specific neural backbone. A simpler comparator, such as gradient-boosted survival regression, should be retained as a transparent baseline. [3] [4]

5. Decision Sequence for Implementation

The implementation sequence is intended to be explicit enough for replication. First, define the value target, horizon, and censoring rule. Second, map all customer histories to a common event schema and record the last observable time. Third, detect data modalities and generate domain-specific features automatically, including CLV-related summaries and engagement scores when supported by the source fields. Fourth, estimate an exposure model for the probability that future value remains observable. Fifth, fit the CLV predictor using a censoring-adjusted objective. Sixth, calibrate predicted values by segment and horizon using held-out data. Seventh, inspect residuals, weight instability, and subgroup drift before deployment. [1] [2] [3]

The decision logic should also specify when not to use the full hybrid model. If the available data contain only sparse transactions and no auxiliary covariates, a simpler censored regression or parametric survival-value decomposition may be preferable because the added complexity of multimodal fusion will not be justified. If text data are present but are domain-shifted across product lines, the encoding layer should be adapted before joint training, since cross-domain instability can otherwise be mistaken for customer heterogeneity. If observation windows are very short, the model should emphasize interval-level summaries and avoid high-capacity sequence encoders that may overfit finite-length noise. [2] [5]

6. Comparator Models, Validation Design, and Outcome Measures

The comparator set should include at least four prespecified families. The first is a naïve uncensored CLV model that ignores right censoring and trains only on fully observed horizons or treats unobserved future value as zero. The second is a concrete survival-value decomposition baseline: estimate time to relationship termination with a Cox proportional-hazards model or, if proportional hazards diagnostics fail, with a prespecified piecewise-constant-hazard model as the fallback; switching to another proportional-hazards model would not remedy a failed proportional-hazards assumption, then combine the estimated survival curve with a separate regression for expected period value among active customers. The third is a mixed-frequency baseline that aggregates all inputs to one common interval before prediction. The fourth is a generic AutoML pipeline without domain-specific feature construction. The proposed method should be compared against these baselines on the same temporal splits. The main outcome is fixed-horizon future-value prediction error, measured by mean absolute error, root mean squared error, and a calibration score comparing predicted versus realized cumulative value. A secondary outcome is ranking quality, for example concordance between predicted CLV and realized post-window contribution. [1] [4]

Validation should be prospective in design, even if executed retrospectively on historical data. Training must end before the evaluation window begins to prevent leakage from future customer events. At least one temporal split should be retained as an untouched test period. Stability should be assessed by perturbing the observation window, removing a subset of modalities, and re-estimating the model after random text or feature noise. For each perturbation, the direction of change in predicted CLV should be inspected, not only aggregate metrics. This is consistent with the logic of stability assessment in customer sentiment scoring and with cross-domain validation principles in heterogeneous review modelling. [6] [2]

Because censoring can be informative, the evaluation protocol should include an explicit sensitivity analysis. One scenario assumes missingness is conditionally independent of future value after conditioning on observed customer state; a second scenario allows weak dependence on prior activity and account tenure. If predictions change sharply across these scenarios, the model should be considered unstable for policy use. The study should also report subgroup performance across acquisition cohorts, product lines, and tenure bands, because a pooled average may conceal systematic underestimation for short-tenure customers. [5] [1]

7. Discussion of Boundary Conditions and Conflicting Evidence

The selected sources point in a compatible direction, but they do not eliminate important tensions. Automation of feature generation can reduce manual effort, yet it can also conceal whether a derived feature is meaningful under censoring or merely predictive in a specific historical sample. Mixed-frequency modelling preserves information, but it increases the risk of misaligned lag structures when customer events are irregular and sparse. Domain adaptation can stabilize heterogeneous text inputs, but transferred representations may become less interpretable when the source and target product contexts differ sharply. These are not contradictions; they are boundary conditions that shape when the proposed CLV pipeline is justified. [1] [4] [2]

A second limitation concerns the finite-length nature of many customer histories. Methods developed for short sequences warn that dependence estimates can be biased when sample lengths are small, which is common in early-life customers and in low-frequency purchase categories. For that reason, the proposed method should not rely on long-horizon correlation estimates as core predictors unless enough exposure time is available. Similarly, a text-based sentiment module should be optional rather than mandatory, because not every institution has consistent review or service-text data and because transfer performance is not guaranteed when the textual domain shifts. [5] [2]

8. Prospective Evaluation Protocol

A suitable prospective evaluation would sample customers from at least two operational segments, such as subscription and transaction-based accounts, and reserve later-origin cohorts for testing. The protocol would first construct monthly or weekly CLV targets over a fixed horizon and record censoring times independently of model fitting. It would then train the proposed model and comparators on pre-specified historical windows, tune hyperparameters only on validation folds, and evaluate final performance on the held-out cohorts. Outcomes should be computed both overall and within tenure strata so that the model’s behavior for newly acquired customers can be separated from that for mature accounts.

To test robustness, three stress conditions should be preregistered. The first removes one modality at a time, such as text or calendar covariates, to measure dependence on any single input source. The second shortens the observation window to mimic early deployment. The third perturbs transaction timing on the native time grid. A reproducible default rule is to shift each eligible event independently by J \in {-1,0,+1} grid units with probabilities {0.25, 0.50, 0.25}, truncated at the customer-specific observation bounds and rejected if it would violate within-customer event ordering or move an event past the prediction origin used to construct features. If multiple events collapse onto the same timestamp after perturbation, their original order is preserved and the events are treated as co-occurring within that grid cell. After perturbation, every time-dependent feature, exposure measure, and sequence input must be recomputed from the perturbed event stream rather than copied from the unperturbed data; purely static attributes remain unchanged. Horizon outcomes and censoring indicators are not perturbed unless the protocol explicitly defines a separate outcome-noise analysis. Sensitivity analysis may widen the jitter to \pm 2 grid units, but the main stress test should remain the \pm 1 rule. A model that remains calibrated under these perturbations would be more credible for operational use than one that only performs well in a single static split. [6] [3] [2]

Generalization should be evaluated cautiously. A model trained on one institution, channel mix, or product portfolio may not transfer without recalibration, because CLV is shaped by pricing, retention policy, and customer lifecycle structure. The framework therefore treats external validation as necessary, not optional. Only if temporal calibration and subgroup ranking remain acceptable after transfer can the model be recommended for broader deployment. [1] [2]

9. Conclusion

This manuscript establishes a conceptual framework for estimating CLV under right censoring that is explicit about variables, censoring adjustment, feature generation, and validation. The main proposal is to combine domain-aware automated pipeline construction with censoring-aware weighting with an optional separately specified augmented target estimator, while retaining mixed-frequency inputs and optional text-based adaptation when available. The literature support is indirect but coherent: financial customer analytics motivates automatic marketing-feature engineering; mixed-frequency forecasting preserves information across time scales; finite-length correlation work warns against unstable dependence estimates; and cross-domain sentiment research supports explicit robustness and transfer checks. [1] [4] [5] [2] [6]

What remains untested is the empirical performance of the proposed estimator relative to conventional CLV baselines across real censored customer populations. In particular, it is not yet known whether the censoring-adjusted hybrid model will outperform simpler methods when data are sparse, whether domain-adapted textual features add value beyond transactional summaries, or how sensitive the estimator is to misspecification of the censoring mechanism. Those questions require prospective evaluation with pre-defined comparators and held-out cohorts before any operational recommendation can be made.

Contribution

The manuscript contributes a bounded conceptual method for customer lifetime value estimation when follow-up is incomplete and inputs are heterogeneous. Its main contribution is not a new empirical claim, but a precise integration rule: define CLV on an explicit prediction origin and horizon, separate administrative observability from economic activity, estimate censoring probabilities from information available at the prediction origin, and train the value model with a preregistered censoring-adjusted loss. Around that core, the manuscript organizes domain-aware feature generation, mixed-frequency alignment, optional domain-adapted text processing, and stability checks into one implementable sequence. The contribution is therefore methodological and operational: it turns several transferable ideas from customer analytics, mixed-frequency forecasting, short-sequence robustness, and sentiment-transfer research into a falsifiable design for censored CLV evaluation rather than an open-ended set of modelling options. [1] [4] [5] [6] [3] [2]

Restricted-horizon target and identification requirements

The target is the undiscounted sum of period-specific net margins over H, not infinite lifetime value. Fit one conditional mean m_h for each horizon step and sum its predictions. Inverse-observation weighting identifies these means only under conditional independence of administrative censoring and future outcomes given origin-time features, with positive observation probability in every target stratum. Weight truncation trades variance for bias and does not restore identification when observation probability is zero. Administrative end dates that deterministically remove all follow-up in a cohort require restricting the target population; they cannot be repaired by a probability floor. Intermittent missingness is a separate process and is not set to zero. Informative unmeasured censoring is addressed only through explicitly labeled sensitivity scenarios. The primary evaluation uses origins with complete H-step follow-up. An absorbing-churn survival baseline predicts each step as survival probability times expected margin conditional on remaining active. Timing perturbations preserve pre-origin membership, sort ties by immutable event ID and recompute features; no future event may cross into the input window. Housing forecasting, transfer-entropy and sentiment studies supply methodological analogies, not validation of this CLV estimator.

Validation Status

Conceptual research only; no empirical validation has been performed in this manuscript. All model structures, notation, loss functions, thresholds, comparator definitions, and stress tests are proposed prospectively and should be treated as preregistered design elements for later evaluation rather than as evidence of performance. In particular, the manuscript does not report fitted parameter values, benchmark comparisons, sample characteristics, or observed gains from domain-aware features, mixed-frequency inputs, or text adaptation. Validation remains to be carried out on temporally separated customer cohorts with explicit observation indicators, prespecified horizons, and sensitivity analyses for weight truncation, modality removal, short observation windows, and timing jitter. The default empirical protocol should report overall and subgroup results, calibration by horizon, the distribution of estimated inverse-probability weights, and failure cases in which observability probabilities approach zero. Any later study that departs from the default Cox-or-Weibull comparator rule, the primary weighting loss, or the stated truncation levels should record that departure before examining test outcomes.

References

1. Tian, Ye; Shao, Wenqian; Deng, Zihan (2026). Marketing-AutoM3L: domain-aware automated machine learning for financial customer analytics. Frontiers in Artificial Intelligence. 9, 1726900. doi:10.3389/frai.2026.1726900.
2. Kumar, Ajeet; Abhishek, Kumar; Shafeeq B M, Ahamed (2025). SentXFormer: a transformer-enhanced hybrid deep learning framework for cross-domain sentiment analysis of customer reviews. Scientific Reports. 16(1), 3528. doi:10.1038/s41598-025-33526-1.
3. Pang, Weilin; Wang, Xingbo; Li, Kehuan (2026). Mamba ECIS for power marketing customer behavior forecasting using multimodal deep learning. Scientific Reports. 16(1), 13063. doi:10.1038/s41598-026-43438-3.
4. Singh, Nutan; Shanmugam, Kavitha (2026). Exploring predictive models to improve the accuracy of Housing Price Index forecasts in India’s real estate sector. PLOS One. 21(1), e0341026. doi:10.1371/journal.pone.0341026.
5. Qiu, Lu; Jiao, Yanzhe; Dong, Gege; Cui, Guangcan (2026). Finite-Length Spatiotemporal Modelling for Housing Price Network Spillovers. Entropy. 28(5), 537. doi:10.3390/e28050537.
6. Mathew, Sujith Samuel; Hayawi, Kadhim; Venugopal, Neethu; El Barachi, May (2026). Quantifying customer sentiment for automobile brand perception analysis using machine learning on Twitter. Scientific Reports. 16(1), 5703. doi:10.1038/s41598-026-35637-9.