The paper deals with approximations of random sums. By random sum we mean a sum of random number of independent and identically distributed random variables. Distribution of this sum is called a compound distribution. The model is especially important in non-life insurance. There are many methods for approximating compound distributions, one of the most popular one is approximation with shifted gamma distribution. In this work we show an alternative way – using kernel density, Fast Fourier Transform and numerical optimization methods – for finding shifted gamma approximations and show results suggesting its superiority over classical method.