RESEARCH DOSSIER
Penalized Regression and Variable Selection in Insurance
Abstract. Insurance prediction uses heterogeneous observations, intermittent losses, correlated risk classifications, and numerous potential predictors. Penalized regression controls estimation instability by combining a statistical loss with an explicitly defined penalty on coefficients. This monograph investigates ridge regression, the lasso, elastic net, grouped and adaptive penalties in insurance frequency, severity, and aggregate-cost settings. It develops the estimand and exposure foundations, derives central statistical properties, examines penalty tuning and validation, and distinguishes selected variables from reliable future insurance predictions. An exact two-predictor comparison illustrates the coefficient allocation induced by competing penalties in a highly correlated design. A second synthetic Poisson study demonstrates how shrinking a low-exposure segment contrast changes risk relativities while respecting portfolio-wide event totals. The findings integrate primary statistical literature, actuarial monographs, and relevant U.S. professional data guidance. The original contribution is a critical actuarial interpretation of established methods, assumptions, and reproducible mathematical examples. It introduces no new named statistical theory and claims no empirical performance result from an insurer.
Keywords: Penalized regression; insurance statistics; actuarial science; generalized linear models; ridge regression; lasso; elastic net; variable selection; shrinkage; claim frequency; claim severity; exposure offsets; correlated predictors; high-dimensional data; credibility; cross-validation; stability selection; insurance pricing.
Introduction
An insurance regression model answers a defined question about an insured event, population, observation period and exposure base. Its response may be the count of claims during a policy year, severity conditional on a claim, aggregate cost per insured exposure, the occurrence of a rare event, or duration until termination. These targets have different distributions and different units of observation. Penalization concerns the parameters of a model only after the response, exposure, and prospective prediction problem have been specified.
The established CAS insurance GLM guide situates rating factors inside the insurance modeling workflow. The newer comprehensive pricing monograph examines techniques and challenges encountered with contemporary insurance data. Correlated policy descriptors, numerous classes, interactions, sparse outcomes, and engineering of external risk signals make conventional parameter estimates unstable. Independent insured entities and claim events, rather than the raw number of electronic records, govern the strength of statistical evidence.
The historical statistical methods are grounded in Hoerl and Kennard's ridge analysis, Tibshirani's lasso research, and the elastic-net development. Ridge regularizes nonorthogonal least squares. Lasso permits a genuinely sparse coefficient vector. Elastic net supplies an intermediate construction useful when columns overlap. Their distinct coefficient geometries have consequences for insurance decisions, including whether a class is assigned a measurable relativity, pooled with a complement or retained inside a correlated group.
The scientific problem is to determine when shrinkage improves estimation and prediction, what a selected subset legitimately communicates, and how a fitted model should be evaluated before use in underwriting or pricing. This requires disentangling five relationships: the insured outcome and its observation process; the risk signals and their encoding; the estimator and penalty; the validation sample and deployment portfolio; and the prediction and actual insurance decision.
The research method comprises a comparative review of original statistical articles, current CAS monographs, and specified professional guidance, supplemented by explicitly formulated algebra and synthetic numerical calculations. All numerical examples are hypothetical. Claims about the accuracy of commercial insurance models require insurer data and independently defined prospective test sets, which are outside this research record. The analysis makes substantive interpretive arguments but does not create a new statistical estimator, named doctrine or conceptual system.
The main boundary is regularization and feature selection in regression. Broader credibility theory, reserving methods, model governance, causal risk intervention, jurisdiction-specific insurance law and market conduct are addressed where they establish necessary assumptions or decision consequences; they remain independent research subjects.
High-Dimensional Insurance Regression
A regression model begins with an estimand. For a claim-frequency GLM, let be the count of claims observed for policy-exposure unit , earned exposure, and the covariate row available before prediction. A Poisson model uses and . The coefficient of the exposure offset is fixed at one. Frequency per earned vehicle year differs substantively from frequency per active contract or per report; the construction of must follow the target.
Severity analysis models positive claim amounts conditional on the relevant occurrence and observation rule. A log-link gamma GLM may be a useful starting hypothesis, subject to checking mean–variance relationships, claim truncation, limits and reporting status. A two-part frequency–severity construction predicts expected loss as predicted frequency times expected conditional severity under a compatible target and conditioning structure. Alternatively, compound Poisson–gamma insurance analysis addresses aggregate cost with a point mass at zero and positive amounts. These are different likelihoods, so a common penalty value does not imply comparable optimization.
Write the general penalized estimator as
Here is a declared negative log-likelihood or predictive loss, are prespecified case weights, a fixed loss normalizer, an offset, a dispersion parameter when relevant and the regularization strength. The model intercept and known exposure offset normally remain unpenalized. Whether software averages the loss over policies, total exposure or events changes the numerical scaling of the fitted . Every published comparison should disclose its exact objective.
In Gaussian reference notation , with , the covariance of ordinary least squares is where this inverse exists. Very small eigenvalues of the Gram matrix create large coefficient variability. The original ridge solution moderates this behavior by adding positive diagonal curvature to the estimation problem. Other generalized linear models replace the least-squares geometry with the curvature of their likelihood, but the fundamental issue remains the statistical identifiability of nearly redundant columns.
Insurance data create multiple routes to near redundancy. Vehicle market value and vehicle class may encode related risk information. Postal classifications, geographic hazard maps and regional socioeconomic scores can capture overlapping places. A telematics trip count can be associated with distance driven. Group-level variables may correlate with the underwriting channel. Factor expansions and interactions may create dozens of columns for a relatively small number of independent claims.
Dimensionality must be understood relative to independent evidence and the effective rank of the design. Ten million transactions arising from a smaller number of claims are not ten million independent events. Renewal records may repeatedly observe the same vehicles and drivers. Multi-policy catastrophe losses share event shocks. If these dependencies are ignored, a superficially large dataset can support spuriously stable coefficient estimates.
Risk selection further changes the observation mechanism. Quotes converted into policies, policies renewed, claims reported and claims closed have different conditional distributions. A frequency model fitted only to reported claims lacks the denominator for insured events unless it explicitly reconstructs relevant exposure. The ASOP No. 23 data-quality guidance is relevant to understanding the fitness, disclosure and limitations of actuarial data within its U.S. professional scope. No coefficient penalty reconstructs unavailable outcomes or removes target leakage.
For high-dimensional insurance modeling, the recommended order of scientific specification is outcome and valuation date; observation and exposure units; the temporal eligibility and provenance of candidate fields; dependency among observations; training population and intended deployment population; and then statistical likelihood and penalty. The penalty is a decision about parameter estimation inside this already defined research object.
Lasso, Ridge, and Elastic Net
The standard penalties can be expressed in a common coefficient notation:
The displayed elastic-net convention uses for lasso and for ridge. The original Zou–Hastie elastic-net formulation uses other scaling notation in places. Comparing implementations by a bare number called lambda without first reconciling the loss and penalty conventions can give misleading results.
For centered Gaussian data, minimize . Writing and gives the ridge solution . In the eigenbasis of , the fitted coefficients are shrunk by the relation between each eigenvalue and . In particular, small singular directions are attenuated, while the positive ridge penalty makes the penalized Gram matrix invertible. Ridge ordinarily produces a dense coefficient vector, so it is a technique for stable estimation rather than direct feature deletion.
For an orthogonal standardized design , lasso solves . The coordinate solution is the soft threshold . If and , then ; the ridge estimate under its corresponding quadratic penalty is , rounded. The lasso's exact zero arises from the non-differentiability of at the origin, as developed in the original lasso paper.
The zero is conditional on the fitted design, sample, penalty and preprocessing. It expresses a sparse numerical solution. It does not establish the scientific absence of a risk mechanism or a causal effect. Correlated variables can substitute for one another, producing similar fitted losses with different selected names. Shrinkage bias may also be important when translating coefficients into insurance rate relativities.
Elastic net combines L1 sparsity with quadratic stabilization, and the elastic-net grouping analysis shows why correlated predictors can be retained together. A strictly positive quadratic component generally stabilizes the solution of convex regression loss even where the original design is rank deficient. Elastic net may be preferable when a substantive risk mechanism is represented by multiple related measures, although the penalty cannot establish which measures are ethically, legally or actuarially appropriate.
Categorical rating variables raise a separate feature-unit problem. If a vehicle-type factor becomes many dummy columns, conventional coefficientwise lasso can retain an arbitrary-looking collection of contrast columns, partly sensitive to the chosen reference category. Yuan and Lin's grouped-selection study considers selection at the group level, with a penalty of the form . Group definition and weights must be specified, particularly when different factors have different numbers of levels. Interactions may also require explicit hierarchy constraints if interpretability demands retaining a parent term whenever an interaction is fitted.
The adaptive-lasso study describes coefficient-specific penalty weights and conditional asymptotic recovery properties. Nonconcave penalized-likelihood research examines different penalty geometries and oracle properties. Such theoretical terms refer to assumptions about sparsity, signal strength, error structure and sample growth; they are not unconditional assurances that a selected pricing variable has an objectively correct real-world meaning. Nonconvex penalties also introduce optimization and local-solution concerns.
For actuarial models using nonlinear link functions, GLM coordinate-descent regularization paths give a practical computational approach for several important convex problems. Similar logic extends to duration models; lasso selection in the Cox model demonstrates the idea for partial likelihood. The objective and score must follow the response: counts, costs, binary events and event times cannot be treated as interchangeable.
Scale and encoding are integral parts of the penalty. A continuous insured-value field in raw currency units and a binary territory indicator can have radically different effective L1 penalties. Standardization should be learned on training data only and applied consistently to new observations. Factor coding, reference classes, missingness indicators and permitted-variable screening are substantive modeling choices. Their decisions and version history belong in the research record.
Penalty Choice and Cross-Validation
Regularization strength is selected for a specified prediction target and validation experiment. A decrease in in-sample deviance after adding factors provides limited evidence for future insurance performance, because the same dataset informed both the factor list and the coefficients. A suitable design separates construction of the candidate set, optimization of the penalty, and an independent evaluation of the finally selected procedure.
Nested validation provides one general approach. Inner folds choose , the elastic-net mixing parameter, transformations and any allowed model complexity. Outer held-out observations estimate the performance of that whole selection algorithm. Preprocessing steps that use future or held-out outcomes must be contained within the training fold. Leakage may occur through claim settlement status, future loss runs, current premium, renewal outcome, geographic target encodings, or an external database revised after the historical prediction date.
An insurer often has several plausible split units. A random policy-period split may place the same insured and vehicle into both sides. An insured-level split tests new entities. A future-calendar split tests new periods and changing exposures. An event-level split avoids shared catastrophe effects across training and testing. The target deployment question determines which split supports the claim. More than one split may be appropriate, but their performances must retain separate labels.
The score must match the model's statistical target. A Poisson negative log-likelihood or deviance measures conditional count fit with the declared exposure offset. Gamma deviance or log score may be appropriate for specified positive-claim severity, subject to the distributional hypothesis. For prospective total loss, frequency and severity forecasts must be combined and evaluated with coherent exposure weights. Claim-level squared errors, policy-level per-exposure errors and annual portfolio-total errors answer different questions.
An evaluation report should identify the training years, data extraction and valuation dates, unit of withholding, model-selection grid, random seed, exact outcome functional, metric weighting and confidence interval or variability calculation. The familiar one-standard-error heuristic may choose a more regularized model near the minimum validation loss, but estimates from repeated correlated folds do not support an automatic independent standard-error interpretation. Every choice on the regularization path is a conditional decision rule.
The relevant CAS rating-model development material situates this work within an insurance modeling process, while recent pricing case studies examine practical challenges beyond simple GLMs. Their role is to clarify applied conditions; they do not imply that a penalty has a universal optimal value. Numerical lambda depends on the response scale, feature scaling, penalty parameterization, and normalization of the likelihood.
Performance can also conflict with organizational constraints. A path point with slightly lower average deviance may generate unstable rate changes in a sparse class. A highly regularized model may stabilize small groups while miscalibrating a persistently high-risk segment. Reporting only the minimum aggregate score hides the trade-off. Useful diagnostics include segment-level observed versus predicted loss, premium-weighted calibration, deviance by cohort, model residual structure, large-loss sensitivity, stability of selected factor blocks and rates under plausible sample changes.
Reserving, pricing and underwriting may also use the same field with different informational legitimacy. A model intended for quotation must use characteristics known when a quote is produced. A claims model may legitimately use post-accident information. Validation must reproduce the data-availability boundary of the intended application.
A scientifically reproducible optimization record states the exact mathematical objective; training-only feature transformations; outcome and exposure definitions; excluded leakage variables; nested selection procedure; candidate penalties; numerical convergence checks; independent holdout results; and business-meaningful implications. The optimal penalty is therefore defined relative to a documented forecasting problem.
Variable Stability and Selection
Selection identifies a subset of columns under a particular fitting and tuning rule. Prediction concerns the resulting behavior of the fitted function on future observations. Neither objective entails the other. A high-dimensional insurance model can exchange geographic scores or vehicle descriptors across repeated samples while producing similar total-loss forecasts. A stable subset can also lose its predictive relevance after changes in the population, the claims process or underwriting practice.
The high-dimensional estimation literature investigates the roles of sparsity, design structure and signal assumptions. The mathematical ability of lasso to choose a true support depends on conditions stronger than a low training deviance or a high coefficient magnitude. Such conclusions must not be treated as guarantees for insurance data with duplicated policy histories, clustered catastrophes or nonstationary reporting.
An empirical selection-stability protocol draws repeated admissible training samples, reruns the entire preprocessing and tuning process, and records each candidate variable's selection frequency. The stability-selection research studies a formal subsampling method and qualified error-control properties. The guarantees require the source study's assumptions, and they do not automatically transfer to data in which policies share claim events, related insureds, or leakage-bearing variables. Each reported frequency should specify whether the resampling units were policy rows, households, insured vehicles, catastrophe events or whole periods.
Selection frequency is one dimension of reliability. Coefficient stability measures sensitivity of fitted magnitudes and signs. Prediction stability measures the variation in policy and portfolio estimates. Time stability concerns subsequent periods. Implementation stability concerns how proposed prices change under reasonable alternative encodings or group definitions. These dimensions should be reported separately rather than collapsed into a single unexamined measure.
Inference after data-driven selection requires care. lasso degrees-of-freedom research characterizes properties of the active-set fit, including geometrical subtleties in rank-deficient designs. The ordinary standard errors from a conventional regression fitted to a selected subset do not automatically account for the variable-search process. Unpenalized post-selection refitting can reduce shrinkage bias but may restore much of the original estimation variance. Selection bias and uncertainty in a policy premium prediction must be distinguished.
Adaptive-lasso asymptotic results and nonconcave likelihood results establish certain oracle properties under formal conditions. The terminology does not imply that the algorithm knows which real-world risk variable is ethically or causally appropriate. In an actuarial model, a recorded field may proxy exposure, reporting behavior, claims-handling intensity or economic conditions. A selected coefficient is a predictive conditional association under the specified data-generating and observation mechanisms.
Interaction selection introduces interpretability constraints. If geographic region, vehicle type and their interaction are independently penalized, the interaction may remain while one of its parent effects is omitted. If a hierarchical form is required for rating communication, it should be imposed by a structured method before selection and tested on prospective observations.
A complete variable-selection audit should therefore contain the original candidate-field dictionary; temporal data availability; factor coding and missingness; group and interaction definitions; selected path by penalty; plausible alternative correlated inputs; repeated resampling stability; prospective prediction stability; data leakage exclusions; and a statement of which coefficient inferences are purely model-conditional. This is a substantive research record, not simply a list of nonzero parameters.
Pricing Model Implications
Insurance pricing translates a conditional statistical prediction into a financially and legally situated rate decision. The insurer first defines covered claims, contractual periods and the exposure denominator. Claims development, valuation date, losses excluded by contract, inflation and expense assumptions then affect the target. A predictive rate relativity may influence a tariff, but the tariff is also shaped by coverage, limits, deductibles, expenses, capital requirements and jurisdiction-specific rules.
Consider a second synthetic experiment with two independent policy groups. Group A records 10 claims on exposure 100; group B records 2 claims on exposure 10. Define expected event counts and , and penalize only their relative log-rate contrast :
This is Poisson negative log-likelihood up to additive constants plus a stated ridge penalty. The intercept remains unpenalized. At zero penalty, the group frequencies are and , so the B/A rate relativity is . Solving the two stationary equations for additional penalties gives the following deterministic outcomes to six decimal places.
| Penalty | Fitted frequency A | Fitted frequency B | B/A rate relativity |
|---|---|---|---|
| 0 | 0.100000 | 0.200000 | 2.000000 |
| 1 | 0.104178 | 0.158215 | 1.518692 |
| 10 | 0.108245 | 0.117548 | 1.085946 |
| 100 | 0.109001 | 0.109987 | 1.009042 |
The first-order equations are $100e^a+10e^{a+\delta}=12$ and $10e^{a+\delta}+\lambda\delta=2\lambda\to\infty, the fitted contrast tends to zero and the rates approach the common portfolio frequency \12/110\approx0.109091$. The monotonic convergence here is a property of this particular two-group convex objective and this dataset, not a general statement about arbitrarily constructed insurance rating models.
This example clarifies one meaning of statistical pooling. Less-supported subgroup experience can be drawn toward shared exposure experience as penalty strength increases. But the example is ridge penalization of a single log-rate contrast, and it is distinct from the specifically constructed lasso credibility methodology described by Holmes and Casotto. If a complement comes from a different population or external experience base, the analyst must examine transfer, overlap and possible double use of information.
For a log-link rate model, increasing a coefficient by one unit changes the conditional predicted mean by the multiplicative factor where the other model terms and offsets are fixed. A penalty that shrinks that coefficient toward zero generally pulls this conditional relativity toward one. It does not prove that equal premium treatment of two insured categories is actuarially justified. Raw claim history, exposure distributions and underwriting constraints may still distinguish them in ways that the chosen model omits.
The CAS insurance rating framework supplies context for the rest of tariff development, and the recent insurance pricing synthesis discusses further modeling challenges. A robust actuarial report should distinguish indicated expected loss cost, rate relativity, selected segmentation, expense and risk loading, actual approved premium, and prospective observed performance. Each is a different quantity, and the relationship among them must be explicit.
Data fitness is fundamental. The U.S. ASOP No. 23 supplies professional guidance on data selection, review and disclosure in its designated actuarial scope. It should not be presented as a single governing insurance statute across all jurisdictions. The legal treatment of a particular pricing factor, its proxy properties and the level of permissible segmentation must be assessed against the correct dated jurisdictional instruments before any deployment. A statistical optimization guarantee is not a regulatory approval.
A variable may predict claims because it captures road hazards, geographic claims-handling behavior, a distribution channel, or a coverage mix. These mechanisms support different underwriting narratives even if their predictive correlations look similar. The use of a variable to forecast costs is not evidence that changing the variable would change losses. Causal attribution, fairness evaluation and benefit design must be researched in their own right.
The operational documentation accompanying a penalized insurance model should include observation windows and data valuation; covered and excluded outcomes; independent policy entities and event clusters; available-at-quotation fields; candidate-factor constraints; exposure reconciliation; model family and loss scale; candidate penalty grid; coefficient path; validation and calibration results; insurer-specific segment sensitivity; monitoring triggers; and versions of all source data. Monitoring should consider risk inflation, legal change, claim reporting delay, shifts in portfolio composition and changes in the features' availability.
The international empirical agenda is substantial. Comparative studies could measure when stable predicted costs coexist with unstable selected factors, how group penalties alter permissible segmentation, how different credibility complements affect low-exposure groups, and how temporal drift changes tuning decisions. These are testable research questions grounded in existing methods, with no new named theory or doctrine asserted.
Conclusion
Penalized regression is a defined family of estimation methods that can improve the stability of regression coefficients and the control of complexity in insurance prediction. Ridge attenuates ill-conditioned directions and generally retains nonzero parameters. Lasso produces sparse solutions through its absolute-value penalty. Elastic net combines stabilization and selection, often retaining correlated features together. Group and adaptive penalties further express structural choices about the units of selection and relative penalty strength.
The first central finding is that correct insurance estimands, exposure units and valuation conventions are prerequisites to valid regularization. The second is that selection of model variables and generalization of insured-loss predictions must be separately tested. The third is that translating a fitted coefficient into a rating factor involves statistical evidence, credibility assumptions, contractual definitions and jurisdiction-specific implementation decisions.
The exact correlated-predictor calculation demonstrates distinct ridge, lasso and elastic-net coefficient allocations under the same stipulated data geometry. The independent two-group Poisson calculation shows how penalizing a log-frequency contrast pulls a small-exposure rate relativity toward a common portfolio estimate without changing the fitted total number of events. These are reproducible synthetic mathematical results and make no empirical claim about real insurers.
The literature demonstrates rigorous statistical methods for controlling complexity and relevant actuarial work on lasso credibility. Original analytical synthesis here organizes those methods around insurance data provenance, selection stability, rate interpretability and evaluation design. Scientific progress depends on explicit assumptions, source-specific verification and reproducibility of the exact objective rather than on formula names alone.
Future research should compare these methods on independently governed insurance datasets using properly separated fitting and deployment periods, with portfolio-level calibration, selected-factor sensitivity and subgroup consequences reported alongside conventional predictive scores. The present monograph establishes a complete methodological account of the defined research topic; it assigns no first-publication date and reports no external peer review or unverified scientific priority.
References
Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality (effective April 30, 2017). Actuarial Standards Board. https://www.actuarialstandardsboard.org/asops/data-quality/
Bühlmann, P., & van de Geer, S. (2011). Statistics for High-Dimensional Data: Methods, Theory and Applications. Springer. https://doi.org/10.1007/978-3-642-20192-9
Chalk, A., Deacon, D., Guillen, M., & Martinelli, M. (2026). From GLMs to Comprehensive Insurance Pricing: Techniques and Challenges (CAS Monograph No. 16). Casualty Actuarial Society. https://www.casact.org/publications-research/publications/flagship-publications/cas-monographs/monograph-no-16
Fan, J., & Li, R. (2001). Variable selection via nonconcave penalized likelihood and its oracle properties. Journal of the American Statistical Association, 96(456), 1348–1360. https://doi.org/10.1198/016214501753382273
Friedman, J., Hastie, T., & Tibshirani, R. (2010). Regularization paths for generalized linear models via coordinate descent. Journal of Statistical Software, 33(1), 1–22. https://doi.org/10.18637/jss.v033.i01 https://www.jstatsoft.org/article/view/v033i01
Goldburd, M., Khare, A., Tevet, D., & Guller, D. (2025). Generalized Linear Models for Insurance Rating (2025 revision; CAS Monograph No. 5). Casualty Actuarial Society. https://www.casact.org/monograph/cas-monograph-no-5
Hoerl, A. E., & Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1), 55–67. https://doi.org/10.1080/00401706.1970.10488634
Holmes, T., & Casotto, M. (2025). Penalized Regression and Lasso Credibility (2025 revision; CAS Monograph No. 13). Casualty Actuarial Society. https://www.casact.org/publications-research/publications/flagship-publications/cas-monographs/monograph-no-13
Jørgensen, B., & Paes de Souza, M. C. (1994). Fitting Tweedie's compound Poisson model to insurance claims data. Scandinavian Actuarial Journal, 1994(1), 69–93. https://doi.org/10.1080/03461238.1994.10413930
Meinshausen, N., & Bühlmann, P. (2010). Stability selection. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 72(4), 417–473. https://doi.org/10.1111/j.1467-9868.2010.00740.x
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological), 58(1), 267–288. https://doi.org/10.1111/j.2517-6161.1996.tb02080.x
Tibshirani, R. (1997). The lasso method for variable selection in the Cox model. Statistics in Medicine, 16(4), 385–395. https://doi.org/10.1002/(SICI)1097-0258(19970228)16:4%3C385::AID-SIM380%3E3.0.CO;2-3 https://pubmed.ncbi.nlm.nih.gov/9044528/
Tibshirani, R. J., & Taylor, J. (2012). Degrees of freedom in lasso problems. The Annals of Statistics, 40(2), 1198–1232. https://doi.org/10.1214/12-AOS1003
Yuan, M., & Lin, Y. (2006). Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1), 49–67. https://doi.org/10.1111/j.1467-9868.2005.00532.x
Zou, H. (2006). The adaptive lasso and its oracle properties. Journal of the American Statistical Association, 101(476), 1418–1429. https://doi.org/10.1198/016214506000000735
Zou, H., & Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 67(2), 301–320. https://doi.org/10.1111/j.1467-9868.2005.00503.x