Paper: arXiv 2311.07597
Authors: Alexej Brauer
Abstract
Currently, there is a lot of research in the field of neural networks for non-life insurance pricing. The usual goal is to improve the predictive power via neural networks while building upon the generalized linear model, which is the current industry standard. Our paper contributes to this current journey via novel methods to enhance actuarial non-life models with transformer models for tabular data. We build here upon the foundation laid out by the combined actuarial neural network as well as the localGLMnet and enhance those models via the feature tokenizer transformer. The manuscript demonstrates the performance of the proposed methods on a real-world claim frequency dataset and compares them with several benchmark models such as generalized linear models, feed-forward neural networks, combined actuarial neural networks, LocalGLMnet, and pure feature tokenizer transformer. The paper shows that the new methods can achieve better results than the benchmark models while preserving certain generalized linear model advantages. The paper also discusses the practical implications and challenges of applying transformer models in actuarial settings.
Complexity vs Empirical Score
- Math Complexity: 7.0/10
- Empirical Rigor: 6.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper presents mathematically sophisticated transformer architectures for tabular data and combines them with established actuarial models, yet it validates these methods on a single real-world dataset with standard benchmarks, demonstrating practical backtest-readiness.
Research Flowchart
flowchart TD
A["Research Goal<br>Enhance Actuarial Pricing with Transformers"] --> B["Data<br>Real-world Claim Frequency Dataset"]
B --> C["Methodology<br>Feature Tokenizer Transformer Models"]
C --> D{"Benchmarks"}
D --> E["GLM, FNN, CAN, LocalGLMnet, FTT"]
D --> C
C --> F["Computational Process<br>Train & Evaluate Hybrid Models"]
F --> G["Key Findings<br>Superior Performance & GLM Advantages"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes novel transformer‑based methods (CAFTT, LocalGLMftt) to enhance actuarial non‑life pricing models for tabular data.
Source sentence
Our paper contributes to this current journey via novel methods to enhance actuarial non-life models with transformer models for tabular data.
Assumes 4
- The response variable Y follows an exponential dispersion family (EDF). statistical
Source sentence
The main assumption for a glm is that Y follows an exponential dispersion family (EDF).
- All covariates xi are numerical after dummy or one‑hot encoding. data
Source sentence
We assume here that all covariables xi are numerical (i.e. ∈Rk), so for example, a dummy or one hot encoding is applied beforehand.
- Poisson regression with exposure vi is appropriate for claim frequency. model
Source sentence
In the case of a Poisson regression with exposure vi, for i ∈{1, . . . , n} the i’th prediction would be given by: ˆyi = exp(β0 + ⟨β, xi⟩)vi
- Default hyper‑parameter settings are used for transformer models (no tuning). execution
Source sentence
the performance is still moderate. This can be traced back to the fact that we didn’t perform any hyperparameter tuning and used the default hyperparameter settings.
Data
- French motor third‑party liability claims frequency dataset (FreMTPL2freq) benchmark public
Source sentence
The dataset of choice is the well-known French motor third‑party liability (MTPL) claims frequency dataset FreMTPL2freq Dutang and Charpentier (2018).
Costs
Not applicable
Finds 3
- Transformer‑based models (FTTdef, CAFTTdef, LocalGLMfttdef) clearly outperform their FNN‑based counterparts. performance
Source sentence
At this stage, we already see that the transformer‑based models (FTTdef, CAFTTdef, LocalGLMfttdef) clearly outperform their FNN‑based counterparts (FNNEMB,CANN, LocalGLMnet).
- CAFTT sets a new benchmark on the FreMTPL2freq dataset. performance
Source sentence
Especially the results for the CAFTT are setting, to the best of our knowledge, a new benchmark on this very popular dataset.
- Average predicted claims frequency of transformer models is substantially far from the true mean of 7.35%. predictive-accuracy
Source sentence
the average predicted claims frequency avg(ˆy) on the test dataset is in comparison to the Poisson GLMs substantially far away from the true mean of 7.35%.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 17 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (46)
- LightGBM : A Highly Efficient Gradient Boosting Decision Tree (2017) — link
- Predictive learning via rule ensembles (2008) — link
- Nesting Classical Actuarial Models into Neural Networks (2019) — link
- EDITORIAL : YES , WE CANN ! (2019) — link
- From Generalized Linear Models to Neural Networks , and Back (2019) — link
- AI : Coming of age? (2022) — link
- SHAP for Actuaries : Explain any Model (2023) — link
- Enhancing Logistic Regression Using Neural Networks for Classification in Actuarial Learning (2023) — link
and 38 more
- LocalGLMnet : interpretable deep learning for tabular data (2023) — link
- Entity Embeddings of Categorical Variables (2016) — link
- Embeddings for Tabular Data : A Survey (2023) — link
- Revisiting Deep Learning Models for Tabular Data (2021) — link
- On Embeddings for Numerical Features in Tabular Deep Learning (2022) — link
- A unified approach to interpreting model predictions (2017) — link
- Statistical foundations of actuarial learning and its applications (2023) — link
- Combined Actuarial Neural Networks in Actuarial Rate Making (2021) — link
- Neural network theory (2022) — link
- Attention is All you Need (2017) — link
- Why do tree-based models still outperform deep learning on typical tabular data? (2022) — link
- Believing the Bot - Model Risk in the Era of Deep Learning (2019) — link
- Generalized Linear Models (1972) — link
- Greedy Function Approximation : A Gradient Boosting Machine (2001) — link
- Generalized Linear Models (1983)
- Deep learning (2016)
- Generalized Additive Models (1986) — link
- Generalized Additive Models (1990)
- Deep Residual Learning for Image Recognition (2016) — link
- Bayesian Deep Net GLM and GLMM (2020) — link
- TabNet : Attentive Interpretable Tabular Learning (2021) — link
- TabPFN : A Transformer That Solves Small Tabular Classification Problems in a Second (2023) — link
- TabR : Unlocking the Power of Retrieval - Augmented Tabular Deep Learning (2023) — link
- Varying- Coefficient Models (1993) — link
- TabTransformer : Tabular Data Modeling Using Contextual Embeddings (2020) — link
- The Transformer Model in Equations — link
- CASdatasets : Insurance datasets (2018) — link
- Peeking into the Black Box : An Actuarial Case Study for Interpretable Machine Learning (2020) — link
- Case Study : French Motor Third - Party Liability Claims (2020) — link
- GLU Variants Improve Transformer (2020) — link
- Delving Deep into Rectifiers : Surpassing Human - Level Performance on ImageNet Classification (2015) — link
- The balance property in neural network modelling (2022) — link
- CASdatasets : Insurance datasets (2018) — link
- API design for machine learning software: experiences from the scikit-learn project (2013)
- Python 3 Reference Manual (2009)
- TensorFlow : Large-Scale Machine Learning on Heterogeneous Systems (2015) — link
- Optuna: A Next-generation Hyperparameter Optimization Framework (2019)
- A multi-task network approach for calculating discrimination-free insurance prices (2023) — link
0 of 46 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.