A comparative machine learning approach for predicting incurred but not reported (IBNR) insurance reserve data in the presence of censored and truncated data

Subject Areas : Financial Economics

Akbar Pilehvar Soltanahmadi ¹ , Kiumars Shahbazi ^{2 *} , Hamzeh Didar ³

1 - Economic Sciences, Faculty of Economics, Urmia University, Urmia/Iran
2 - Department of Economic, Faculty of Economic and Management, Urmia University, Urmia, Iran
3 - Department of Accounting, Faculty of Economics and Management, Urmia University, Urmia, Iran

Received: 2024-11-13 Accepted : 2024-12-02 Published : 2025-09-06

Keywords: Incurred But Not Reported Reserves, Random Forest, Multi-Layer Perceptron, Long Short-Term Memory,

Abstract :

This study aims to predict incurred but not reported (IBNR) reserves in various insurance lines by employing advanced machine learning models and analyzing censored and trimmed data. The dataset includes information on incident and report dates for five major insurance lines: third-party financial, vehicle, third-party bodily injury and driver accidents, fire, and liability. The methods applied in this study are Multiple Linear Regression (MLR), Generalized Linear Model (GLM), Generalized Additive Model (GAM), Random Forest (RF), Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM) networks, using data from Iran Insurance Company for the period of 2021-2022. The data were censored and trimmed based on specific periods, such as holidays, Nowruz, peak travel seasons, and construction periods, to model impactful features according to the insurance line type. Results indicate that LSTM and RF models outperform linear models in predicting delays; specifically, RF achieved errors of 10.64 and 11.02 in vehicle and third-party financial lines, while LSTM attained errors of 9.83 and 10.72, respectively. These models effectively identified complex patterns in the data, revealing that considering factors such as holidays, weekends, and data structure can help capture intricate insurance data patterns. The findings underscore that LSTM and Random Forest models significantly enhance prediction accuracy, serving as valuable tools for risk assessment and optimal reserve allocation in the insurance industry.