Comparative analysis of forecasting models in the nonlife insurance: Insights from the SARIMA and ETS approaches

Abstract

To ensure competitiveness and sustainability, insurance companies need accurate predictions. The paper analyzes nonlife insurance gross written premiums, technical premiums, claims, and the number of contracts through a multi‐model univariate approach comparing Seasonal Autoregressive Integrated Moving Average (SARIMA) with Exponential Smoothing models. The hypothesis suggests that SARIMA's superior handling of complex seasonal patterns enhances predictive accuracy. Data from 2012M01 to 2023M06 provides a comprehensive sector‐wide view, essential for decision‐making. The results show that SARIMA models outperform exponential smoothing in almost all major accuracy metrics over the training period. Over the test period, the exponential smoothing models show more accurate performances. However, the optimal exponential smoothing models fail in mitigating autocorrelation and non‐normality in residuals, which is successfully tackled by the weighted ensemble models. In all cases except for the number of contracts concluded, the rolling one‐step ahead forecast approach generates superior accuracy over the classic training‐test split. This research confirms that using SARIMA, exponential smoothing, weighted ensemble, and rolling window models can significantly improve forecasting accuracy and are better solutions than the seasonal naive benchmark. It provides important insights into insurance forecasting and risk management, which are often overlooked in existing literature, and can greatly benefit corporate governance and policymaking.