생명보험산업 보험료 성장률 예측계량모형 비교

In this article we evaluate the performance of forecasting models to predict the Korean life insurance premium growth rates. Comparisons are made for the Vector Auroregressive (VAR) predictive model, the multivariate leading indicator model, and the diffusion index model proposed by Stock and Watson (2002) against the Univariate Autoregressive (AR) predictive model as a benchmark. We compare each model’s predictability for the total premium incomes and the initial premium incomes of three types in the individual insurance; protection, endowment and annuity. The complete quarterly data spans from the second quarter of 1986 to the first quarter of 2014. The lag selection for AR and VAR forecasting models depends on the Bayesian Information Criteria (BIC) with the maximum number of lags set to 4. For the multivariate leading indicator model, we use 4 leading indicators of GDP growth rates, inflation, education, and age. In order to avoid data mining concerns, we select the variables that have been found to be the determinants of life insurance demands by previous studies. The diffusion index model is an approximate dynamic factor model that relates the future life insurance premium growth rates to a number of factors estimated by principal components using a large number of macroeconomic variables. The set of macroeconomic variables consists of 57 variables representing 6 main categories of macroeconomic time series: demand for final output; balance of payments and international trade; price indexes; money, interest rates and financial markets; labor, production and population; and world variables. We also include in the data set the variables related to the life insurance industry such as the industry’s total asset returns, claims paid and etc. In predicting total premium income growth rates, the AR predictive model produces the smallest mean squared predictive errors (MSPEs). For the premium incomes of protective insurance all the forecasting models have the MSPEs less than 3%. But they become unreliable in predicting the premium incomes of annuity with producing the MSPEs more than 10%. When we test the null hypothesis of no difference in MSPEs, it is rejected at 5% significance level for the VAR and the diffusion index predictive models when we forecast the total premium income growth rates of protective insurance. In predicting initial premium income growth rates, we find that there are no statistically significant differences in the MSPEs of each model. In addition, all the models have MSPEs more than 10% so that we may not be able to depend on any model to forecast in practice. We conclude that it may not be beneficial to take advantage of the information contained in macroeconomic variables for predicting life insurance premium growth rates. It may be due to the fact that most of the insurance contracts in Korea charge monthly premiums, which induces heavy autocorrelations among quarterly insurance premium data and makes an AR forecast very effective. Rho and Shin (1998) also found that macroeconomic variables are not likely to influence on insurance demands as is largely determined by insurer’s push-marketing. Policy holders cannot surrender their insurance contracts without heavy penalties, which makes them not so much responsive to macroeconomic environments. Lastly, life insurance demands are very sensitive to changes in insurance regulation and taxation, which could not be controlled for by our estimation procedure due to the lack of time series observations.