By: Unknown
Re-posted from: https://juliasnippets.blogspot.com/2017/10/pydata-warsaw-2017-example.html
Following several requests in this post I am presenting the Asian option pricing example that I have discussed during PyData Warsaw 2017 (after the first day of talks I highly recommend everyone to come tomorrow as it is an excellent event).
The problem is taken from the book Foundations and Methods of Stochastic Simulation by Barry Nelson where it is solved using VBA in section 4.5.
I will not repeat the whole discussion of the model, but focus on the numerical aspect. We are asked to calculate the value of the expression:
where X is generated by a continuous state geometric Brownian motion. We will use Monte Carlo simulation to approximate the above expected value.
A single pass of the Monte Carlo simulation is approximated by a discrete sum:
The parameters we will use in the simulation are: T=1, r=0.05, K=55, σ=0.3, m=1000 and starting value of X at time 0 is 50. We will run 100,000 replications of Monte Carlo simulation and calculate the average.
I want to compare five implementations of the above problem:
- Julia using loops
- Julia using vectorized code
- Python using loops
- Numba using loops
- NumPy using vectorized code
The key take-aways are the following:
- standard Python is a no-go solution for this problem;
- loops in Julia are fastest;
- somewhat surprisingly vectorized Julia code is faster than Numba although the former has to allocate more memory;
- NumPy implementation is around three times slower than vectorized Julia;
- Vectorized Julia code hugely benefits from in-place operations (that avoid memory allocation); however, even without these optimizations it was faster than Numba.