On the use of case estimate and transactional payment data in neural networks for individual loss reserving

arXiv:2601.05274 · q-fin.ST, cs.LG, stat.ML · Submitted 2025-12-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: We're looking at the paper: "On the use of case estimate and transactional payment data in neural networks for individual loss reserving".

Paper summary: Ruth: So, to recap this paper by Avanzia, Lambrianidis, Taylorb and Wongb aCentre for Actuarial Studies and UNSW Business School it compares two neural network types for individual loss reserving. Warren They look at a feed-forward neural network trained on summarized transactions versus a recurrent neural network that analyzes the entire payment history or case estimate development history. Iris The core thesis here is that case estimates will improve predictions significantly, but equipping the neural network with memory only leads to meagre improvements when using those histories. Ruth It matters because they tested this on multiple, complex datasets simulated from SPLICE which represents highly complex scenarios where data gets lost upon aggregation, as Taylor and McGuire and Sullivan showed >

Warren: So the paper is testing how well a recurrent neural network equipped with full claim history performs against a feed-forward network trained only on summaries. Iris And they found evidence that case estimates will improve predictions, specifically showing a significant improvement in absolute mean reserve error from twenty point five percent down to seven point seven percent >

Ruth: That drop is substantial when you compare the models without those case estimates to the ones with them. Warren It suggests that having access to these features may lead to more robust estimates of individual claims, which in turn could lead to more accurate estimates for the group >

Iris: The paper also points out that this framework isn't just about using machine learning; it's about providing a way for insurers to assess the value of their own case estimates, since they acknowledge that how case estimation happens varies significantly between different insurers >

Ruth: That makes sense because they aren't trying to give one single answer on whether case estimates are universally better or worse for every insurer, but rather a framework for them to explore their own data >

Warren: So the paper is setting up these research questions, like whether there’s a significant difference in performance between a feed-forward network trained solely on summary data compared to an RNN trained on full transactional data >

Iris: And another question they ask is whether a neural network, whether recurrent or feed-forward, performs better when it's trained with case estimate data than the exact same model without those case estimates >

Conclusion: Ruth: Thinking about the whole paper by "On the use of case estimate and transactional payment data in neural networks for individual loss reserving" and its authors, it seems they are focused on giving insurers a practical tool. Warren They are not just proving that a deep learning model is cool; they are trying to show how incorporating specific pieces of information like case estimates can actually help their reserving process work better >

Iris: The implication for me is that we need to think about the assumptions these models make about the people involved in the claims, especially since Ruth mentioned earlier how case estimation varies a lot between insurers >

Ruth: Right, because if an insurer has a system that produces very different case estimates than another one, then their neural network approach might give them a much better assessment of *their* specific estimate >

Warren: And from the quantitative side, it’s important that they are comparing the RNN with full history against the feed-forward model trained on summaries because that tests whether modeling sequence matters for this type of data >

Iris: The paper does have some limitations they mention, like they are not including claims where there is no incurred but no payment, and their models only produce point estimates without stochastic properties >

Ruth: That’s a fair caveat. They are giving us a snapshot of the model's performance at a specific time rather than showing how much uncertainty is built into the prediction >

Warren: So what this means for us is that while we see significant gains when we feed in case estimates, we still need to be careful about what the model can actually do once it's running in a real world environment >

Centre for Actuarial Studies, University of Melbourne · School of Risk and Actuarial Studies, UNSW Business School, UNSW Sydney

q-fin.ST, cs.LG, stat.ML

Submitted: 2025-12-28

Updated: 2026-11-17

Code: https://github.com/agi-lab/reservingRNN

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 78/100

The gist: The use of neural networks trained on individual claims data has become increasingly popular in the actuarial reserving literature In this paper…

Key concepts

Feed-forward Neural Network (FNN)
This type of neural network processes input data as a static vector. It maps an input directly to an output through a fixed sequence of transformations that do not depend on the order or history of the inputs presented at the prediction time.
Recurrent Neural Network (RNN)
An RNN is designed to model sequential processes, like claim development. It uses recurrent connections and a persistent latent state, meaning information from earlier transactions influences later representations, allowing it to capture ordered information over time.
Case Estimate Data
These are time series inputs representing the estimated payment amounts for a claim over time. The paper tests whether including this historical estimate data into the neural network model leads to more robust and accurate reserve predictions.
Individual Claims Reserving
This is the actuarial task of estimating the future cost or liability for a single, specific insurance claim. The models aim to predict these individual loss amounts based on historical payment and development data.

Terminology

Summary

The use of neural networks trained on individual claims data has become increasingly popular in the actuarial reserving literature In this paper, we compare a feed-forward neural network trained on summarised transactions to a recurrent neural network equipped to analyse a claim’s entire payment history and/or case estimate development history We find evidence that case estimates will improve predictions significantly, but that equipping the neural network with memory only leads to meagre improvements

Model Architectures Comparison

The paper compares two primary neural network architectures for individual loss reserving: feed-forward networks and recurrent neural networks Feed-forward networks map an input X to an output y through a fixed sequence of transformations that depend only on the inputs presented at the prediction time Recurrent neural networks explicitly model claim development as a sequential process, where information from earlier transactions influences later representations through recurrent connections and a persistent latent state The distinction is substantive because the recurrent formulation embeds the assumption that claim histories contain ordered information whose relevance evolves over time, whereas feed-forward networks process the input as a static vector once presented to the network

Data Inputs and Model Variants

The paper defines two main types of inputs for the models: time series inputs and static inputs Time series inputs include calendar time, transaction times, cumulative payments, and case estimates for each claim Static inputs consist of prediction quarters, accident quarters, and default covariates simulated in the SPLICE package The models are compared based on whether they include case estimate data or not, resulting in variants like LSTM+ and FNN+ versus LSTM and FNN For instance, the model that includes case estimate histories is denoted as the LSTM+, while the model without this data will be referred to as the LSTM

Performance Metrics and Findings

The study evaluates performance using several metrics, including M1vsM2, Mean Absolute Logarithmic Error (MALE), and Mean Squared Logarithmic Error (MSLE) at the valuation date Research Question 1 finds that transactional data provides a slight benefit to the reserve forecasts but does not improve results at an individual claim level Research Question 2 demonstrates that a neural network trained with case estimate data performs better than the same model without case estimates, showing a significant improvement in absolute mean reserve error from 20.5% to 7.7%

Conclusion and Limitations

The overall findings suggest that the addition of full transactional data appears to slightly improve forecasts in aggregate but deteriorates results at an individual claim level The inclusion of case estimates adds significant value across all claim durations, as seen through a markedly smaller mean MALE and MSLE for the FNN+ relative to the FNN However, limitations include the exclusion of IBNR claims and the models only producing point estimates without stochastic properties The authors stress that their framework is key, encouraging others to perform tests in their own environments

Data and Code Availability

All results in this paper are reproducible using Python and R, with code and data available on Zenodo (https://zenodo.org/records/18005906) and GitHub (https://github.com/agi-lab/reservingRNN) Datasets were simulated in R, while modelling was conducted using Python Hyperparameter tuning and model training was run in an Azure ML notebook on a Standard E8s v3 virtual machine with 8 cores, while the dataset simulation and evaluation of models on the test set were run on a local machine with an Nvidia RTX 3070 GPU and an AMD Ryzen 5 5600X CPU with 6 cores Creation and processing of datasets took roughly 24 hours It took approximately 300 hours to tune and train all models Evaluation of all models, including those in the appendix and supplementary material, on the test sets required around 24 hours

References

Al-Mudafer, M.T., Avanzi, B. p. 0 p. 0, Taylor, G. p. 0 p. 0 p. 0 p. 0C., Wong, B., 2022. Stochastic loss reserving with mixture density neural networks Avanzi, B. p. 0, Dong, E., Laub, P.J., Wong, B., 2025a. Distributional refinement network: Distributional forecasting via deep learning Avanzi, B. p. 0, Li, Y., Wong, B., Xian, A., 2024. Ensemble distributional forecasting for insurance loss reserving Avanzi, B. p. 0, Richman, R. p. 1 p. 1 p. 1, Wong, B., W¨uethrich, M. p. 1, Xie, Y., 2025b. Reinforcement learning for individual claims reserving Avanzi, B. p. 1, Taylor, G., Wang, M., Ho, W., 2021a. SPLICE: Synthetic paid loss and incurred cost experience (splice) simulator Avanzi, B. p. 8, Taylor, G., Wang, M., Wong, B., 2021b. SynthETIC: An individual insurance claim simulator with feature control Avanzi, B. p. 2, Taylor, G.C., Wang, M., Wong, B., 2023. SPLICE: a synthetic paid loss and incurred cost experience simulator Ba, J. p. 8L., Kiros, J.R., Hinton, G.E., 2016. Layer normalization Breiman, L. p. 1, 2001. Statistical modeling: The two cultures Cai, P. p. 1, Abdallah, A. p. 1, Jeganathan, P. p. 1, 2025. Recurrent neural networks for multivariate loss reserving and risk capital analysis Chaoubi, I. p. 0, Besse, C., Cossette, H., Cˆot´e, M.P., 2023. Micro-level reserving for general insurance claims using a long short-term memory network Cho, K. p. 5, Van Merri¨enboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y., 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation Cybenko, G. p. 2, 1989. Approximation by superpositions of a sigmoidal function Delong, L. p. 2, W¨uthrich, M. p. 1 p. 1 p. 1 p. 1 p. 1 p. 1V., 2020. Neural networks for the joint development of individual payments and claim incurred Denuit, M. p. 2, Charpentier, A. p. 8, Trufin, J. p. 8, 2021. Autocalibration and tweedie-dominance for insurance pricing with machine learning Duan, N. p. 1, 1983. Smearing estimate: a nonparametric retransformation method Duval, F. p. 1, Pigeon, M. p. 1, 2019. Individual loss reserving using a gradient boosting-based approach Elman, J. p. 1L., 1990. Finding structure in time Embrechts, P. p. 6, W¨uthrich, M.V., 2022. Recent challenges in actuarial science England, P. p. 1, Verrall, R. p. 8, 2002. Stochstic Claims Reserving in General Insurance Gabrielli, A. p. 9, 2021. An individual claims reserving model for reported claims Gabrielli, A. p. 9, Richman, R., W¨uthrich, M.V., 2019. Neural network embedding of the over-dispersed Poisson reserving model Gabrielli, A. p. 1, W¨uthrich, M.V., 2018. Individual claims history simulation machine Hochreiter, S. p. 1, Schmidhuber, J. p. 5, 1997. Long short-term memory Huang, J. p. 5, Wu, X., Zhou, X., 2016. Asymptotic behaviors of stochastic reserving: Aggregate versus individual models Ioffe, S. p. 1, Szegedy, C., 2015. Batch normalization: Accelerating deep network training by reducing internal covariate shift, in: International conference on machine learning, Proceedings of Machine Learning Research. p. 1 pp. 448–456 James, G. p. 9, Witten, D., Hastie, T., Tibshirani, R., Taylor, J. p. 0, 2023. An introduction to statistical learning with applications in Python Kohavi, R. p. 0, 1995. A study of cross-validation and bootstrap for accuracy estimation and model selection, in: International Joint Conference on Artificial Intelligence, pp. p. 2 1137–1145 Kuo, K. p. 1, 2019. Deeptriangle: A deep learning approach to loss reserving Kuo, K. p. 0, 2020. Individual claims forecasting with bayesian mixture density networks Lambrianidis, M. p. 0, 2024. Individual Loss Reserving with Recurrent Neural Networks Laub, P. p. 1J., Pho, T., Wong, B., 2025. An interpretable deep learning model for general insurance pricing Lopez, O. p. 1, Milhaud, X. p. 1, Th´erond, P. p. 1E., 2019. A tree-based algorithm adapted to microlevel reserving and long development claims Loschilov, I. p. 1, Hutter, F., 2017. Decoupled weight decay regularization Merz, M. p. 2, W¨uthrich, M.V., 2010. Paid–incurred chain claims reserving method Mulquiney, P. p. 2, 2006. Artificial neural networks in insurance loss reserving, in: 9th Joint International Conference on Information Sciences (JCIS-06), Atlantis Press Pigeon, M. p. 2, Antonio, K. p. 1, Denuit, M. p. 1, 2014. Individual loss reserving using paid-incurred data Poon, J. p. 2H., 2019. Penalising unexplainability in neural networks for predicting payments per claim incurred Rumelhart, D. p. 1E., Hinton, G.E., Williams, R.J., 1986. Learning representations by back-propagating errors Schelldorfer, J., W¨uthrich, M.V., 2019. Nesting classical actuarial models into neural networks Schneider, J. p. 0C., Schwab, B. p. 1, 2025. Advancing loss reserving: A hybrid neural network approach for individual claim development prediction Srivastava, N. p. 1, Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R., 2014. Dropout: a simple way to prevent neural networks from overfitting. p. 1 The journal of machine learning research 15, 1929–1958 Tashman, L. p. 1J., 2000. Out-of-sample tests of forecasting accuracy: an analysis and review. p. 0 International journal of forecasting 16, 437–450 Taylor, G. p. 0, 2019. Claim models: granular and machine learning forms Taylor, G. p. 2, McGuire, G. p. 1 p. 1, 2016. Stochastic loss reserving using generalized linear models Taylor, G. p. 1, McGuire, G., Sullivan, J. p. 1, 2008. Individual claim loss reserving conditioned by case estimates W¨uthrich, M. p. 0V., 2018. Machine learning in individual claims reserving W¨uthrich, M. p. 1V., 2020. Bias regularization in neural network models for general insurance pricing W¨uthrich, M. p. 0V., Richman, R., Avanzi, B., Lindholm, M., Maggi, M., Mayer, M., Schelldorfer, J. p. 1, Scognamiglio, S., 2025. Ai tools for actuaries W¨uthrich, M. p. 1V., 2022. The balance property in neural network modelling

--- Page 1 ---

On the use of case estimate and transactional payment data in neural networks for individual loss reserving Benjamin Avanzia, Matthew Lambrianidisa, Greg Taylorb, Bernard WongaCentre for Actuarial Studies, Department of Economics, University of Melbourne VIC 3010, Australia bSchool of Risk and Actuarial Studies, UNSW Business School, UNSW Sydney NSW 2052. Abstract The use of neural networks trained on individual claims data has become increasingly popular in the actuarial reserving literature We consider how to best input historical payment data in neural network models Additionally, case estimates are also available in the format of a time series, and we extend our analysis to assessing their predictive power In this paper, we compare a feed-forward neural network trained on summarised transactions to a recurrent neural network equipped to analyse a claim’s entire payment history and/or case estimate development history We draw conclusions from training and comparing the performance of the models on multiple, comparable highly complex datasets simulated from SPLICE (Avanzi, Taylor and Wang, 2023) We find evidence that case estimates will improve predictions significantly, but that equipping the neural network with memory only leads to meagre improvements Keywords: Loss Reserving, Neural Networks, Case Estimates, Individual Claims, Deep Learning, RBNS Reserves JEL codes: G22, C45, C53 MSC classes: 91G70, 91G60, 62P05, 91B30

--- Page 2 ---

data that is lost upon aggregation (Taylor, McGuire and Sullivan, 2008) At a minimum, this would include information about an individual claim’s payment history However, case estimates are also available in the format of a time series Having access to these features may lead to more robust estimates of individual claims, which in turn could lead to more accurate estimates for the group Huang et al.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Add a framework for insurers to assess their own case estimate values by standardizing their assessment methodology, as proposed in Section 1.2: we provide a standardised methodology for assessing their value.

  2. Enhance the model's ability to capture sequential dependencies by explicitly modeling claim development as an ordered process using Recurrent Neural Networks (RNN) like LSTMs, which explicitly models claim development as a sequential process, in which information from earlier transactions influences later representations through recurrent connections and a persistent latent state.

  3. Implement a mechanism to assess the relative value of case estimates by comparing model performance against case estimates directly using the M1vsM2 metric with appropriate weighting structures, such as w OCL (unless otherwise specified), to emphasize larger claims and earlier in their development.

  4. Develop a robust training methodology that mitigates data leakage risks by employing a proposed split where we randomly move 20% of the validation set into the training set to ensure some slightly longer development claims are included, balancing the need for extrapolation with comprehensive training.

  5. Improve performance in early development periods by leveraging insights from naive splits and bias correction, as suggested by Figure E.16: both models show a remarkably tight forecast across all stages of development, particularly in quarters 1-5 which accounts for almost 50% of the true outstanding claim amounts at the valuation date.

Sources

Related papers