Paper: arXiv 2611.21326
Authors: Eva Flonner
Abstract
Health insurance claims exhibit temporal dependence, as previous claims may provide information about the timing, type and amount of subsequent claims. Such dependencies are difficult to capture with traditional actuarial models based on aggregated claim counts. We investigate a neural marked point process in which a causal Transformer represents individual claim histories and determines type-specific conditional intensities, allowing for self-excitation and interactions between claim types. The framework also incorporates conditional severity distributions, reporting delays and stochastic medical-cost inflation. We develop a likelihood-based calibration procedure that accounts for incomplete claim reporting and evaluate the model using both simulated and empirical data. A simulation study with 10,000 insurance contracts examines the recovery of known claim dynamics, while an application to MEPS-derived medical-event data assesses predictive performance. The fitted model reproduces the main patterns in claim frequencies and severities, although some discrepancies remain in the tails of the distributions. On the MEPS data, the neural model achieves better predictive performance than two generalized linear model benchmarks in terms of claim-count deviance and expenditure discrimination.
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 8.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper presents a sophisticated neural marked point process model, demonstrating high mathematical complexity in its formulation. It is rigorously evaluated using both simulated and empirical data, including a detailed simulation study and application to MEPS data, showcasing strong empirical rigor. The novelty lies in integrating neural networks (Transformers) with marked point processes for health insurance claims, addressing temporal dependencies and actuarial specificities like reporting delays.
Research Flowchart
flowchart TD
A["Research Goal: Capture Temporal Dependence in Health Insurance Claims"] --> B{"Methodology: Neural Marked Point Process"};
B --> C["Key Components: Causal Transformer for Conditional Intensities, Severity Distributions, Reporting Delays, Cost Inflation"];
C --> D["Data/Inputs: Simulated Insurance Contracts (10,000), MEPS-derived Medical-Event Data"];
D --> E["Computational Process: Likelihood-based Calibration accounting for Incomplete Reporting"];
E --> F["Findings (Simulation): Recovers Known Claim Dynamics"];
E --> G["Findings (MEPS Data): Better Predictive Performance than GLM Benchmarks (Claim-Count Deviance, Expenditure Discrimination)"];
F & G --> H["Outcomes: Model Reproduces Main Patterns in Frequencies/Severities (with some tail discrepancies)"];
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 a neural marked point‑process model that combines history‑dependent, type‑specific claim intensities with conditional severity distributions, reporting delays and stochastic medical‑cost inflation.
Source sentence
In this paper, we propose a neural marked point-process model for individual health insur-ance claims that combines history-dependent, type-specific claim intensities with conditional distributions of claim amounts.
Assumes 6
- Claim intensities are modeled as conditional intensities derived from a causal Transformer representation of past claims. model
Source sentence
We investigate a neural marked point process in which a causal Transformer represents individual claim histories and determines type-specific conditional intensities, allowing for self-excitation and interactions between claim types.
- Medical‑cost inflation follows a stochastic process with a common long‑run drift centered at 3% per year. statistical
Source sentence
Weakly informative priors constrain the drift and innovation scales; in particular the common long-run drift is centered at 3% per year in the fitted configuration.
- Reporting delays and exposure periods are correctly captured by the observation structure described (policy exposure, pre‑sample period, etc.). data
Source sentence
Particular attention is given to the observation structure of insurance data, including differences in policy exposure, reporting delays and changes in medical costs over calendar time.
- The logarithmic transformation of claim amounts and a mixture of three Gaussian components adequately model the conditional severity distribution. model
Source sentence
The logarithmic transformation allows the model to describe positive claim amounts on an unrestricted numerical scale, while Mk(t) accounts for the cumulative effect of medical inflation on claim amounts of type k up to time t. The remaining variation is modeled using a mixture of three Gaussian distributions.
Show 2 more
- The simulated portfolio (10,000 contracts) is representative for evaluating recovery of known claim dynamics. execution
Source sentence
A simulation study with 10,000 insurance contracts examines the recovery of known claim dynamics, while an application to MEPS-derived medical-event data assesses predictive performance.
- The MEPS‑derived medical‑event data are sufficiently detailed to assess predictive performance of the model. data
Source sentence
The processed sample contains 5,447 individuals, with 21,001.5 person-years and 436,542 recorded events.
Data
- Simulated insurance portfolio synthetic public
Source sentence
A simulation study with 10,000 insurance contracts examines the recovery of known claim dynamics, while an application to MEPS-derived medical-event data assesses predictive performance.
- MEPS‑based medical‑event data (Panel 24) other public
Source sentence
The processed sample contains 5,447 individuals, with 21,001.5 person-years and 436,542 recorded events.
Costs
Not applicable
Finds 5
- On the MEPS data, the neural model achieves better predictive performance than two generalized linear model benchmarks in terms of claim‑count deviance and expenditure discrimination. predictive-accuracy
Source sentence
On the MEPS data, the neural model achieves better predictive performance than two generalized linear model benchmarks in terms of claim-count deviance and expenditure discrimination.
- The fitted model reproduces the main patterns in claim frequencies and severities, though some discrepancies remain in the tails of the distributions. statistical
Source sentence
The fitted model reproduces the main patterns in claim frequencies and severities, although some discrepancies remain in the tails of the distributions.
- Simulation results indicate that the proposed model captures several important features of the underlying data‑generating process, including claim frequencies, amounts, reporting delays and dependence between successive claims. performance
Source sentence
Overall, the simulation results indicate that the proposed model captures several important 19 Neural marked point processes for health claims Working manuscript Figure 7: Medical cost indices relative to their initial level. The model reproduces the main patterns in claim frequencies and severities, while the agreement is less precise for some claim types and for the development of medical costs over time.
- The model captures the broad development of medical costs more reliably than precise type‑specific trajectories, with reasonable posterior uncertainty bands. statistical
Source sentence
Given the relatively short observation period and the potential overlap between calendar-time effects, attained age and policy duration, the results suggest that the model captures the broad development of medical costs more reliably than their precise type-specific trajectories.
- Results from the simulation describe performance for a single realized portfolio and seed, not variation across independent portfolios. other
Source sentence
Since the experiment uses one simulated portfolio and one random seed, the reported results describe performance for this particular realization rather than variation across independently generated portfolios.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 26 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-17 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (20)
- Evaluating the Role of Order and Time Information for Classifying Sequences of Healthcare Events Derived from Claims Data (2026) — link
- Healthcare cost prediction for heterogeneous patient profiles using deep learning models with administrative claims data (2025) — link
- Predicting Medical Event Occurrence Using Medical Insurance Claims Big Data (2024) — link
- Markov‐modulated marked Poisson processes for modeling disease dynamics based on medical claims data (2022) — link
- Machine learning versus regression modelling in predicting individual healthcare costs from a representative sample of the nationwide claims database in France (2021) — link
- Learning Hidden Patterns from Patient Multivariate Time Series Data Using Convolutional Neural Networks: A Case Study of Healthcare Cost Prediction (2020) — link
- Neural Temporal Point Processes For Modelling Electronic Health Records (2020) — link
- Deep learning for prediction of population health costs (2020) — link
and 12 more
- Transformer Hawkes Process (2020) — link
- Healthcare cost prediction: Leveraging fine-grain temporal patterns (2019) — link
- Lecture Notes: Temporal Point Processes and the Conditional Intensity Function (2018) — link
- Individual claims reserving: a survey (2017)
- Attention is All you Need (2017) — link
- Understanding Reporting Delay in General Insurance (2016) — link
- Prediction of Outstanding Liabilities in Non-Life Insurance (1993) — link
- Spectra of some self-exciting and mutually exciting point processes (1971) — link
- The Neural Hawkes Process: A Neurally Self-Modulating Multivariate Point Process (2017)
- Supervised Learning Methods for Predicting Healthcare Costs: Systematic Literature Review and Empirical Evaluation (2017)
- Medical Expenditure Panel Survey: Household Component Data File Types
- Neural marked point processes for health claims Working manuscript
0 of 20 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.