The DOI matches its Crossref record; 63 of the 64 references it cites were found in a registry or an index; 18 of the 30 sources its claims cite were checked for support.
Modelling Motor Insurance Claim Frequency and Severity Using Gradient Boosting
Open accessAbstract
Modelling claim frequency and claim severity are topics of great interest in property-casualty insurance for supporting underwriting, ratemaking, and reserving actuarial decisions. Standard Generalized Linear Models (GLM) frequency–severity models assume a linear relationship between a function of the response variable and the predictors, independence between the claim frequency and severity, and assign full credibility to the data. To overcome some of these restrictions, this paper investigates the predictive performance of Gradient Boosting with decision trees as base learners to model the claim frequency and the claim severity distributions of an auto insurance big dataset and compare it with that obtained using a standard GLM model. The out-of-sample performance measure results show that the predictive performance of the Gradient Boosting Model (GBM) is superior to the standard GLM model in the Poisson claim frequency model. Differently, in the claim severity model, the classical GLM outperformed the Gradient Boosting Model. The findings suggest that gradient boost models can capture the non-linear relation between the response variable and feature variables and their complex interactions and thus are a valuable tool for the insurer in feature engineering and the development of a data-driven approach to risk management and insurance.
Overview
CoChat Summary
TL;DR
This paper compares gradient boosting machines (GBM) with standard GLMs for motor insurance claim frequency and severity using a large proprietary TPL auto insurance dataset. GBM outperforms GLM for claim frequency, but GLM performs better for claim severity.
Key findings
- The dataset contains 799,587 TPL motor insurance policies in force from 1 January 2016 to 31 December 2019, with 2,464,181 exposure observations, 78,264 claims, and €97.9 million total incurred cost.
- For claim frequency, GBM achieved lower deviance than GLM: training 428,621 vs 432,456, testing 106,773 vs 107,914, and overall 535,685 vs 540,406 total Poisson deviance.
- For claim severity, GLM outperformed GBM: training 8,787 vs 10,545, testing 2,162 vs 2,624, and overall 10,954 vs 13,209 total Gamma deviance.
- The strongest GBM predictors for claim frequency were the policyholder’s district of residence, Bonus-Malus level, vehicle brand, payment frequency, and policy seniority; for claim severity, they were district of residence, vehicle brand, driver’s age, years of driving experience, and policy seniority.
Methods
- The study used a proprietary private automobile insurance database supplied by a European insurer, then applied extensive preprocessing: data cleaning, feature selection/engineering, outlier treatment, and dimensionality reduction.
- The final calibration dataset used 2,464,181 observation durations and 21 feature variables after preprocessing; the original data included 36 characteristic variables.
- Models were trained and evaluated with 80% training / 20% testing and nested K-fold cross-validation with K in [1, 6]; GBM hyperparameters were tuned by grid search over tree number, learning rate, interaction depth, and bag fraction, using Poisson deviance for frequency and Gamma deviance for severity.
Limitations
- The paper states that the claim severity GBM likely performed worse because of the smaller sample size for severity, “circa 18,000 observations,” and the higher volatility of claim amounts.
- The dataset is proprietary and not publicly available, which limits reproducibility and external validation.
- Evident from the text: the analysis is based on one insurer’s TPL portfolio, so generalizability beyond this context is not established in the paper.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
15 Valid · 0 Not valid · 6 Not checked · 70 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Empirical studies show that they can provide superior predictive accuracy relative to single learners in several domain-specific areas (Ashofteh et al. 2022; Ayuso et al. 2021; Bravo 2021; Kim and Baek 2022).”
What the sources say
- Supportedcites Ashofteh et al. 2022 source
The Introduction describes superior predictive performance of ensemble methods and lists forecasting applications in population, epidemiology, meteorology, and finance. The Empirical experiments > Results section also reports that this paper’s ensemble outperformed two single forecasting methods in its excess-mortality application. Page numbers are not provided in our copy. (LLM, conf=0.92)
- Partly supportedcites Ayuso et al. 2021 sourceExperimental — accuracy not yet measured
The Introduction states that ensemble learning has improved traditional and machine-learning forecasts, supporting the possibility of better predictive accuracy. The Results discuss mortality-forecasting model weights but do not compare the ensemble’s accuracy with single models; this paper also does not establish the claim across several domain-specific areas. The supplied copy gives section labels but no page numbers. (LLM, conf=0.81)
- Supportedcites Bravo 2021 source
In Materials and methods, under “Mortality forecasting: a Bayesian Model Ensemble approach” (page not available in our copy), the paper says ensemble learning has improved forecasting results compared with a single model. It also describes its use beyond mortality forecasting, supporting the citing sentence’s broad, qualified claim. (LLM, conf=0.88)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“A similar conclusion was obtained by Henckaerts et al. (2021) using a portfolio of motor Third Party Liability from a Belgian insurer in 1997.”
What the sources say
- Supportedcites Henckaerts et al. (2021) source
Section 4.3 (p. 16) explicitly says that the most important risk factors in the gradient boosting machines are those selected in the GLMs. Section 4 (p. 13) identifies the data as a motor third-party liability portfolio from a Belgian insurer in 1997. (LLM, conf=0.95)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The out-of-sample performance measure results show that the predictive performance of the Gradient Boosting Model (GBM) is superior to the standard GLM model in the Poisson claim frequency model.”
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Differently, in the claim severity model, the classical GLM outperformed the Gradient Boosting Model.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The average optimal number of trees is significantly smaller in the claim severity model compared to the claim frequency model.”
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
37 more claims
Create a free account to see all 91 claims and every source verdict.
- Probably not a substantial claim. 34
- It may state two separate claims as one. 21
Citations
Cites
Breiman, L. (2001) · Machine Learning · doi:10.1023/a:1010933404324 · Cited by 133,538 · cited 3×
“To overcome the black-box nature of GBM, (Breiman 2001; Friedman and Popescu 2008; Hastie et al. 2009) introduced two important tools to interpret the GBM model: (i) variable importance measures and (ii) partial dependence plots.”
Greedy function approximation: A gradient boosting machine
Friedman, J. H. (2001) · The Annals of Statistics · doi:10.1214/aos/1013203451 · Cited by 30,731 · cited 2×
“Following Friedman (2001), in this paper, we focus on a class of additive expansions of the form Given M iteration steps, the parameter estimates can be written in the incremental form.”
The Elements of Statistical Learning
Hastie, T., Tibshirani, R., & Friedman, J. (2009) · Springer Series in Statistics · doi:10.1007/978-0-387-84858-7 · Cited by 25,970 · cited 7×
“Partial dependence plots exhibit the average effect of a feature on the predictions obtained from a model, see Hastie et al. (2009), whereas ICE plots disaggregate the averaged data, providing a method of inspecting how the instance’s prediction changes when a feature varies.”
InfluentialWolpert, D. H. (1992) · Neural Networks · doi:10.1016/s0893-6080(05)80023-1 · Cited by 7,951
Adaptive Mixtures of Local Experts
Jacobs, R. A., Jordan, M. I., Nowlan, S. J., & Hinton, G. E. (1991) · Neural Computation · doi:10.1162/neco.1991.3.1.79 · Cited by 5,384
Hansen, L. K., & Salamon, P. (1990) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/34.58871 · Cited by 4,260 · cited 1×
“…applications of machine-learning ensemble techniques in different domains include random forests (Breiman 2001), artificial neural network ensembles (Hansen and Salamon 1990; Shu and Burn 2004), Bayesian model ensembles (Bravo and Ayuso 2021; Raftery et al. 1997), bootstrap aggregating (bagging),…”
Goldstein, A., Kapelner, A., Bleich, J., & Pitkin, E. (2015) · Journal of Computational and Graphical Statistics · doi:10.1080/10618600.2014.907095 · Cited by 1,861 · cited 2×
“ICE plots allow us to capture heterogeneity in the relationship between the feature variable and the response created by variable interaction, see Goldstein et al. (2015).”
Bayesian Model Averaging for Linear Regression Models
Raftery, A. E., Hoeting, J. A., & Madigan, D. (1997) · Cited by 1,746
Predictive learning via rule ensembles
Friedman, J. H., & Popescu, B. E. (2008) · The Annals of Applied Statistics · doi:10.1214/07-aoas148 · Cited by 1,315
Bayesian Model Averaging for Linear Regression Models
Raftery, A. E., Madigan, D., & Hoeting, J. A. (1997) · Journal of the American Statistical Association · doi:10.1080/01621459.1997.10473615 · Cited by 1,287 · cited 1×
“…(Breiman 2001), artificial neural network ensembles (Hansen and Salamon 1990; Shu and Burn 2004), Bayesian model ensembles (Bravo and Ayuso 2021; Raftery et al. 1997), bootstrap aggregating (bagging), boosting and meta-learning strategies for expert combination such as stacking (Ashofteh and…”
Research works printed but not linked: 1
- Gao, 2023, Dependence modeling of frequency-severity of insurance claims using waiting time Author links open overlay panel, Insurance: Mathematics and Economics, 109, 29
Cited by
Panjee, P., & Amornsawadwatana, S. (2024) · Risks · doi:10.3390/risks12020025 · Cited by 17
“The study from Clemente et al. (2023) also shows that when studying both machine learning and GLMs, the results from assessing performance outside the sample indicate that the gradient boosting model (GBM) demonstrates better predictive accuracy than the standard generalized linear models (GLMs) in…”
ResultFrom point to probabilistic gradient boosting for claim frequency and severity prediction
Chevalier, D., & Côté, M. (2025) · European Actuarial Journal · doi:10.1007/s13385-025-00428-5 · Cited by 18
“Gradient boosting proliferates in actuarial applications, such as auto insurance [4, 5], hierarchical non-life reserving [6], and health insurance [7, 8].”
Machine Learning Applications for Predicting High-Cost Claims Using Insurance Data
Brati, E., Braimllari, A., & Gjeçi, A. (2025) · Data · doi:10.3390/data10060090 · Cited by 15
“[10] used a dataset from a European insurance firm from 1 January 2016 to 31 December 2019 to model auto insurance claims through gradient boosting (GB) and generalized linear models (GLMs).”
Ensemble Methods for Stock Market Prediction
Bravo, J. M. (2024) · Communications in computer and information science · doi:10.1007/978-3-031-74643-7_31 · Cited by 7
Self-citation
Metadata
- Authors
- Clemente, C., Guerreiro, G. R., & Bravo, J. M.
- Affiliations
- Universidade Nova de Lisboa (PT)
- Iscte – Instituto Universitário de Lisboa (PT)
- Corresponding author
- Guerreiro, Gracinda R.
- Venue
- Risks
- Publisher
- MDPI AG
- Published
- 2023-09-12
- Volume
- 11
- Issue
- 9
- Pages
- 163
- ISSN
- 2227-9091 (ISSN-L)
- OpenAlex
- W4386714714
- Semantic Scholar
- eac714dccf0b89ce13d1ea2e4760fcad9c6f961f
- Cited by
- 28
- References
- 64
- Citations by year
- 2026: 7 · 2025: 11 · 2024: 9 · 2023: 1
- FWCI
- 8.51
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Risks (publishedVersion, cc-by)
- Repositório do ISCTE-IUL (submittedVersion, other-oa)
- New University of Lisbon's Repository (New University of Lisbon) (submittedVersion)
- RePEc: Research Papers in Economics (submittedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- FCT—Fundação para a Ciência e a Tecnologia (UIDB/00297/2020, UIDP/00297/2020, UIDB/04152/2020, UIDB/00315/2020)
- Center for Mathematics and Applications (UIDB/00297/2020, UIDP/00297/2020, UIDB/04152/2020, UIDB/00315/2020)
- Centro de Investigação em Gestão de Informação (MagIC) (UIDB/00297/2020, UIDP/00297/2020, UIDB/04152/2020, UIDB/00315/2020)
- BRU-ISCTE-IUL (UIDB/00297/2020, UIDP/00297/2020, UIDB/04152/2020, UIDB/00315/2020)
- Topics
- Social Sciences › Economics, Econometrics and Finance › Economics and Econometrics › Insurance and Financial Risk Management
- Social Sciences › Decision Sciences › Management Science and Operations Research › Probability and Risk Models
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
- Keywords
- automobile insurance, claim frequency, claim severity, gradient boosting, decision tree, non-life insurance, Poisson regression, generalized linear models
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Modelling Motor Insurance Claim Frequency and Severity Using Gradient Boosting. https://app.cochat.ai/papers/10.3390/risks11090163
- The paper
- Clemente, C., Guerreiro, G. R., & Bravo, J. M. (2023). Modelling Motor Insurance Claim Frequency and Severity Using Gradient Boosting. Risks. https://doi.org/10.3390/risks11090163
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.