Research on China insurance demand forecasting: Based on mixed frequency data model

China's insurance industry, while experiencing significant growth, lags behind developed nations in insurance density and depth. The global economic downturn, rising unemployment, and uncertainties stemming from public health emergencies highlight the importance of accurately forecasting insurance demand in China. The insurance industry plays a crucial role as a social stabilizer and economic booster, making understanding its influencing factors vital for ensuring high-quality economic development. This paper proposes using a MIDAS model to analyze the macroeconomic factors affecting China's insurance demand and provide real-time forecasts. Macroeconomic fluctuations and risk, particularly economic policy uncertainty (EPU), significantly impact insurance purchase decisions. Consumer confidence (CCI) and consumer price index (CPI) also play important roles. The study uses CCI, EPU, and CPI (monthly high-frequency data) to predict insurance demand measured by the quarterly low-frequency indicator of insurance depth (TID). This research contributes by applying the MIDAS model to insurance demand forecasting, optimizing the model's design and prediction methods, and providing short-term forecasts for China's insurance demand.

Existing research on insurance demand often focuses on micro-level data (household life cycle, financial status, asset mix, gender), revealing heterogeneity in demand across different demographics. Studies show that older households hold less commercial life insurance than younger ones, financial status significantly impacts insurance consumption, and married female household heads are more likely to purchase commercial insurance. However, macroeconomic volatility risk is an important contextual factor influencing insurance purchase decisions. The EPU index, measuring uncertainty in economic policies, is a key indicator of macroeconomic fluctuations. Previous studies have established a positive relationship between insurance demand and fluctuations in EPU. Consumer expectations and disposable income levels are also significant drivers of insurance premium expenditures. Traditional forecasting models using co-frequency data, such as AR, ARMA, GARCH, neural networks, and GM(1,N), often face challenges due to data frequency mismatches, leading to information loss and potential distortion. The MIDAS model overcomes these limitations by incorporating variables of different frequencies, improving forecasting accuracy and timeliness.

This study employs a MIDAS model using monthly data (CCI, EPU, CPI, 2006M1–2022M3) and quarterly data (TID, 2006Q1–2022Q1). Data were seasonally adjusted using the X-12-ARIMA method. The univariate h-step ahead prediction model is defined as: yt = β0 + β1B(L1/m; θ)x(m)t + εt, where yt is the low-frequency variable, x(m)t is the high-frequency variable, B(L1/m; θ) is the weighted polynomial function, L1/m is the lag operator, and εt is the error term. m is the frequency ratio (3 in this case). The MIDAS model incorporating CCI, EPU, CPI, and TID(-p) is specified as: TIDt = β0 + β1B1(L1/3; θ)CCIt(3) + β2B2(L1/3; θ)EPUt(3) + β3B3(L1/3; θ)CPIt(3) + β4TIDt-p + εt. The study uses U-MIDAS, Beta, Almon, and Exp Almon weighting functions to examine the forecasting effects. Combination forecasting using EW, BIC, MSFE, and DMSFE weighting methods is also employed. Model selection is based on the Akaike Information Criterion (AIC). In-sample predictions are evaluated using Root Mean Square Error (RMSE). The study compares the MIDAS model with traditional AR and ARDL models. Different prediction windows (fixed, rolling, and recursive) are used to analyze their impact on prediction performance.

The MIDAS model consistently outperforms the AR and ARDL models in terms of in-sample prediction accuracy, as measured by RMSE. The Exp Almon and U-MIDAS models generally provide the best results. The optimal lag orders for CCI, EPU, and CPI vary depending on the sample period (full sample vs. in-sample excluding the COVID-19 period). Rolling window and recursive predictions show higher accuracy than fixed window predictions. Combined forecasting, particularly using BIC weights, further enhances the prediction accuracy. CCI demonstrates the strongest predictive power and explanatory capacity for TID, especially during the COVID-19 period. Nowcasting and short-term forecasting using the optimal MIDAS model suggest that China's TID is expected to recover to pre-COVID-19 levels by 2023Q2. The multivariate MIDAS model (excluding EPU) shows the best fitting performance.

The study's findings directly address the research question by demonstrating the superiority of the MIDAS model in forecasting China's insurance demand. The model's ability to handle mixed-frequency data effectively utilizes the information contained in both high- and low-frequency variables, leading to improved prediction accuracy compared to traditional co-frequency models. The identification of consumer confidence as the key driver of insurance demand highlights the importance of considering behavioral indicators in policymaking. The model's forecasting capabilities provide valuable insights for policymakers, insurers, and investors, enabling better resource allocation and risk management.

This study successfully demonstrates the application of the MIDAS model in forecasting China's insurance demand, highlighting its superior performance compared to traditional methods. The identification of consumer confidence as a key driver and the successful short-term forecasting provide valuable insights for policymakers. Future research could explore the impact of other macroeconomic variables, examine the long-term relationship between the variables, and further refine the MIDAS model's weighting functions.

The study primarily focuses on the predictive ability of the MIDAS model, neglecting a deeper exploration of the long-term relationships and potential bidirectional causality between the variables. The analysis does not comprehensively investigate the limitations of the different MIDAS weighting functions and their potential impact on predictive performance. While these aspects were not the central focus of the current study, they represent valuable avenues for future research.