This paper reports on the acceleration of a standard, lattice-based numerical algorithm that is widely used in finance for pricing a class of fixed-income vanilla derivatives. We start with a high-level algorithmic specification, exhibiting irregular nested parallelism, which is challenging to map efficiently to GPU hardware. From it we systematically derive and optimize two CUDA implementations, which utilize only the outer or all levels of parallelism, respectively. A detailed evaluation demonstrates (i) the high impact of the proposed optimizations, (ii) the complementary strength and weaknesses of the two GPU versions, and that (iii) they are on average 2.4× faster than our well-tuned CPU-parallel implementation (OpenMP+AVX2) running on 104 hardware threads, and by 3-to-4 order of magnitude faster than an OpenMP-parallel implementation using the popular QuantLib library.