Forecasting Insurance Company Commitments with Long Short-Term Memory Models

ObjectiveThis study aims to present a novel model for predicting the future commitments of insurance companies that can adequately address the potential challenges of traditional methods. Traditionally, insurance companies use the Chain Ladder approach as a statistical tool to forecast the trend of claims development. This statistical method is favored by regulatory authorities in various countries due to its simplicity in assumptions and clear interpretation. However, certain assumptions, such as the stability of data development and linear relationships between variables, can affect the efficiency of this model when faced with internal policies or external factors like the COVID-19 pandemic. Forecasting future commitments close to reality is closely related to the financial stability of insurance companies. The amount that insurance companies allocate to meet their future obligations is identified as reserves. Calculating reserves that are less than the required amounts can pose challenges for insurance companies in fulfilling their commitments while calculating more than necessary amounts can negatively impact the financial statements of insurance companies. MethodsIn this study, a dynamic model based on machine learning algorithms is proposed. The model's output, which combines the number and timing of bodily injury accidents, plays a crucial role in calculating reserves for non-life insurance products. This model is specifically trained to predict the frequency of accidents in Vehicle Third-Party Liability Insurance. It can identify hidden patterns and non-linear, complex relationships within claims data. A Long Short-Term Memory (LSTM) neural network algorithm is employed, recognized for its strong predictive capability in time series data. The model is trained using historical data from Karafarin Insurance Company covering the years 2017 to 2021. ResultsThe performance of the model is highly related to the hyperparameters chosen for the model. Two of the most common approaches for tuning the hyperparameters are tested in this study. These Two models are grid and random search. The Root Mean Square Error (RMSE) is used as a performance metric, and it indicates that the grid search has a lower RMSE than the random search for the training data with a slight difference (16.33 versus 17.4). However, the results for the test data in the grid search have a sign of overfitting. ConclusionThis study recommends using random search for tuning the hyperparameters of the model to predict the frequency of daily incidents. The evaluation of the two approaches for tuning hyperparameters indicates that random search is more suitable for working with unfamiliar data and managing overfitting situations. Overfitting occurs when the model becomes overly influenced by the training data, learning not only the actual patterns but also the noise and minor details of the data. This issue can negatively impact the model's generalization ability.