DeepTriangle: A Deep Learning Approach to Loss Reserving

Kevin Kuo

Source-linked AI summary

DeepTriangle: A Deep Learning Approach to Loss Reserving

Kevin Kuo

arXiv:1804.09253v4Apr 24, 2018stat.APcs.LGq-fin.RM

TL;DR

Loss reserving must forecast delayed, uncertain claim payments that are important for insurer decisions and financial reporting. DeepTriangle uses deep neural networks to jointly model paid losses and claims outstanding with heterogeneous inputs, and it improves performance over several established models while supporting automated forecasting. The study does not provide uncertainty estimates as posterior predictive distributions and uses an aggregated dataset with limited features.

Problem
Delayed reporting, prolonged claim development, and uncertain claim amounts make accurate loss-reserve forecasting difficult, despite its importance to insurers, investors, and regulators.
Source: §1 · p. 1
Method
DeepTriangle is a deep-learning framework that jointly models paid losses and claims outstanding, incorporates heterogeneous inputs, and learns company representations through neural-network embeddings.
Source: §1 · p. 2, §3.4.3 · p. 6
Results
DeepTriangle improves performance over chain-ladder, ODP, common machine-learning, and Bayesian stochastic models, while capturing some individual-company development patterns but missing deteriorating loss ratios in one example.
Source: §4.3 · p. 9
Takeaways & Limitations
The framework maintains predictive accuracy while enabling automated, extensible model updating and reporting without expert input during routine forecasting.
Source: §5 · p. 12
The study does not estimate uncertainty or interpret forecasts as posterior predictive distributions, and it analyzes an aggregated dataset with limited features.
Source: §4.3 · p. 9, §5 · p. 12

Abstract

from arXiv · show

We propose a novel approach for loss reserving based on deep neural networks. The approach allows for joint modeling of paid losses and claims outstanding, and incorporation of heterogeneous inputs. We validate the models on loss reserving data across lines of business, and show that they improve on the predictive accuracy of existing stochastic methods. The models require minimal feature engineering and expert input, and can be automated to produce forecasts more frequently than manual workflows.

1. Introduction

Loss reserving is difficult because claims are reported and paid with delays and uncertain ultimate amounts, yet accurate reserves matter to insurers, investors, and regulators. DeepTriangle addresses this problem with an automated neural-network framework that jointly models multiple loss-related outcomes and heterogeneous inputs.

Motivation

Accurate unpaid-claim estimates inform insurer decisions, investor financial statements, and regulators’ assessments of financial soundness.Source: §1 · p. 1
Delayed reporting, prolonged claim development, and uncertain claim amounts make forecasting future payments difficult.These challenges motivate extensive research on loss reserving methods.Source: §1 · p. 1

Related work

Prior neural-network reserving research incorporates claims features, synthesizes claims data, or embeds classical reserving models into trainable networks.Examples include Wüthrich, Gabrielli and Wüthrich, Gabrielli et al., and Gabrielli.Source: §1 · p. 1

Contribution

DeepTriangle jointly models paid losses and claims outstanding for multiple companies while accommodating additional inputs such as claim counts and economic indicators.Source: §1 · p. 2
The framework requires no manual input during updates or forecasting, enabling more frequent predictions and faster management responses to portfolio changes.Source: §1 · p. 2

2. Neural Network Preliminaries

Neural networks transform inputs through successive parameterized layers, using nonlinear activations and learned weights to produce predictions. Training repeatedly updates these parameters on mini-batches to minimize a loss function.

Terminology

DeepTriangle is named as a portmanteau of deep learning and loss development triangle.Source: §2 · p. 2

Network structure

Hidden-layer activations transform input data into representations that become progressively more useful for predicting the output.Source: §2 · p. 3

Network parameters

Activation functions generate layer outputs, while weights and biases form the learned network parameters.Source: §2 · p. 3

Network structure

A neural network represents prediction as a composition of parameterized functions across layers, with nonlinearities enabling more than a trivial composition of linear models.Source: §2 · p. 3

Training

Training iteratively updates parameters to minimize prediction loss, typically using mini-batches over repeated passes through the data.Source: §2 · p. 3

3. Data and Model Architecture

This section introduces the dataset used for the experiments and the proposed neural-network architecture.

Data and Model Architecture

The paper’s data and model sections cover the experimental dataset and the architecture proposed for loss-reserving forecasts.Source: §3 · p. 3

3.1. Data Source

The experiments use publicly available NAIC Schedule P claims triangles, restricted to four business lines and 50 companies per line for comparison with existing results.

Data Source

The dataset contains NAIC Schedule P accident-year and development-year records for accident years 1988–1997, each with 10 years of development experience.Source: §3.1 · p. 3
The study covers commercial auto, private personal auto, workers’ compensation, and other liability, with 50 companies in each line of business.This restriction facilitates comparison with existing results.Source: §3.1 · p. 3
Inputs include line of business, company code, accident year, development lag, incurred loss, cumulative paid loss, and net earned premium.Claims outstanding is derived as incurred loss minus cumulative paid loss.Source: §3.1 · p. 4

3.2. Training/Testing Setup

The setup defines observed paid-loss and claims-outstanding triangles and forecasts their future development through the oldest maturity available. Training uses data through year-end 1997, with predictive performance evaluated on development-year-10 cumulative paid losses.

Observed data consist of paid losses and claims outstanding for accident years 1 through I and development years 1 through I − i + 1.Source: §3.2 · p. 4
Future paid losses and claims outstanding are forecast for accident years i = 2, ..., I and development years j = i + 1, ..., I.Source: §3.2 · p. 4
Training uses data as of year-end 1997, and predictive performance is evaluated on development-year-10 cumulative paid losses.Source: §3.2 · p. 4

3.3. Response and Predictor Variables

DeepTriangle represents each accident-year-development-year cell by normalized paid-loss and claims-outstanding responses together with observed history and company identity. Claims outstanding is included as an auxiliary prediction target alongside paid losses.

Each training sample corresponds to an accident-year-development-year cell whose response is a sequence of paid-loss and claims-outstanding loss ratios.Loss ratios divide each quantity by net earned premium, placing values on a similar scale for training.Source: §3.3 · p. 4
The predictor combines the observed history available at the cell’s calendar-year evaluation date with the associated company identifier.Source: §3.3 · pp. 4–5
Claims outstanding is modeled as an auxiliary output while the primary interest remains predicting paid losses.Source: §3.3 · p. 4
The architecture includes a company-code pathway for distinguishing experience from multiple insurers during training.Source: §3.3 · pp. 4–5

3.4. Model Architecture

DeepTriangle is a multi-task sequence-to-sequence network that jointly predicts paid losses and claims outstanding using sequential history, company embeddings, and separate output subnetworks. Its recurrent encoder-decoder processes development sequences, while shared temporal outputs support both targets.

3.4. Model Architecture

DeepTriangle uses one sequence-to-sequence model per line of business, trained across multiple companies with two prediction goals: paid loss and claims outstanding.Source: §3.4 · p. 5

3.4.1. Multi-Task Learning

Joint training is intended to improve prediction by supplying additional claims-outstanding signals, exposing shared features, and discouraging overfitting to noise.Source: §3.4.1 · p. 5

3.4.2. Sequential Input Processing

A GRU encoder summarizes historical values, and a decoder GRU produces hidden states across I − 1 future timesteps using 128 hidden units and dropout rate 0.2.The repeat count matches the I − 1 future steps required for the oldest accident year.Source: §3.4.2 · pp. 5–6

3.4.3. Company Code Embeddings

Company codes are learned as 49-dimensional embeddings whose geometry is optimized for future paid-loss prediction.The embeddings provide learned representations that can act as proxies for company characteristics.Source: §3.4.3 · p. 6

3.4.4. Fully Connected Layers and Outputs

Each decoded timestep is concatenated with the company embedding and sent to separate paid-loss and claims-outstanding subnetworks with 64-unit hidden layers and dropout rate 0.2.Source: §3.4.4 · p. 6
ReLU activations produce nonnegative cash-flow predictions, thereby excluding recoveries under the dataset’s assumption.The authors note that this output assumption may be modified for other use cases.Source: §3.4.4 · pp. 6–7

3.5. Deployment Considerations

Deployment considerations emphasize adapting company embeddings as portfolios change, integrating model forecasts with actuarial workflows, and incorporating richer insurer-held information when available.

For group insurers, the company-code predictor can infer relationships among companies from historical development behavior instead of relying on manual aggregation or clustering.Source: §3.5 · p. 7
New or reorganized companies require enlarging the embedding input and refitting the model, while the remaining architecture stays unchanged.The refitted network assigns embedding vectors to the new companies.Source: §3.5 · p. 7
Because the model forecasts every triangle cell, its outputs can be converted into traditional age-to-age loss development factors for workflow integration.Source: §3.5 · p. 7
The model can incorporate claim-count triangles in addition to paid and incurred losses to learn from and predict frequency information.Source: §3.5 · p. 7

4. Experiments

The experiments evaluate DeepTriangle with unweighted percentage errors, train ensembles using early stopping, and compare out-of-time forecasts against stochastic and machine-learning baselines. DeepTriangle improves performance across the compared models, while qualitative results show both successful and unsuccessful company-level forecasts.

4.1. Evaluation Metrics

Performance is measured with RMSPE and MAPE on ultimate losses through development year 10, aggregated across companies within each line of business.Percentage errors provide unit-free comparisons across companies with substantially different portfolio sizes.Source: §4.1 · pp. 7–8
The evaluation uses unweighted percentage measures because premium- or dollar-weighted results would concentrate on a handful of large companies.The authors identify model evaluation as an ongoing research area and cite Martinek.Source: §4.1 · p. 8

4.2. Implementation and Training

Training uses AMSGRAD ADAM, a 0.0005 learning rate, early stopping after 200 validation epochs without improvement, and up to 1000 epochs.The validation subset contains training data becoming available after calendar year 1995.Source: §4.2 · p. 8
For each line of business, the authors ensemble 100 identically structured networks with different random initializations to reduce neural-network variance.The implementation uses the keras R package and TensorFlow, with experiment code available online.Source: §4.2 · p. 8

4.3. Results and Discussion

The paper benchmarks DeepTriangle against Mack chain ladder, bootstrap ODP, AutoML, and Bayesian MCMC models including CIT and LIT.AutoML uses iterative forecasting because its component methods produce scalar outputs.Source: §4.3 · p. 9
DeepTriangle improves on chain ladder, ODP, common machine-learning, and Bayesian stochastic models in out-of-time performance.The comparison uses averaged ensemble predictions for DeepTriangle and predictive-distribution means for stochastic models.Source: §4.3 · p. 9
The model captures Company 1767’s development patterns reasonably well but fails to forecast deteriorating loss ratios for Company 337’s workers’ compensation book.Figures 4 and 5 examine paid-loss and claims-outstanding development for these two companies.Source: §4.3 · p. 9, Figure 4 · p. 10, Figure 5 · p. 11
The study does not estimate uncertainty or interpret ensemble forecasts as posterior predictive distributions.The authors suggest that applying randomness-based predictive-distribution techniques to reserve variability requires further research.Source: §4.3 · p. 9

5. Conclusion

DeepTriangle maintains predictive accuracy while adding automation and extensibility to loss reserving. The framework can incorporate heterogeneous inputs, but this study uses limited-feature aggregated data and focuses on point estimates.

DeepTriangle attains performance comparable to modern stochastic reserving techniques while enabling automated model updates and report production.The authors recommend expert review within the automated machine-learning process.Source: §5 · p. 12
Deep neural networks allow DeepTriangle to incorporate heterogeneous inputs, train on multiple objectives simultaneously, and customize models to available data.The paper identifies claims adjusters’ free-text notes and images as possible additional inputs.Source: §5 · p. 12
The study analyzes publicly available aggregated data with limited features, although DeepTriangle can be extended to incorporate additional data such as claim counts.Source: §5 · p. 12
The study focuses on point estimates, while future extensions may output distributions to address reserve variability.Source: §5 · p. 12