Balancing Time Steps and Simulations in CVA Monte Carlo

Summary

The document considers how to allocate limited computational capacity in a Monte Carlo system for credit valuation adjustment (CVA) on a portfolio of long-dated derivatives. The central choice is between finer time discretization with fewer simulated paths and coarser time discretization with more paths. It does not offer a universal preferred setting, since the best balance depends on the portfolio and how sensitive the result is to each source of numerical error.

The proposed practical method is to establish a reasonable baseline, then vary the number of time steps while holding the simulation count fixed and measure the result change. Next, vary the simulation count and compare its effect. The dimension that moves the valuation more may deserve additional computational resources. The suggested comparison should be repeated from another baseline to check robustness. This is a sensitivity exercise, not a formal error analysis: the document gives no convergence criteria, confidence intervals, portfolio-specific findings, or guidance on runtime beyond the general resource constraint.

Key ideas

• The useful balance between time discretization and path count depends on the portfolio.
• Start with a baseline configuration and change one computational dimension at a time.
• Compare how much the valuation moves when increasing time steps versus simulations.
• Repeat the sensitivity comparison from another baseline to check whether the result is robust.

Tags

# In CVA simulation, timesteps vs number of simulations?


# In CVA simulation, timesteps vs number of simulations?












On a CVA system with limited computational power.

For pricing, What is best, More timesteps and less number of simulations or less timesteps and more number of simulations?

for example with a whole portfolio of long term derivatives with several counterparties on a long term analysis: 200 timesteps (more granular at first) and 3,000 simulations vs 120 timesteps and 10,000 simulations?

What are your thoughts?

Thanks.

## Answer by Magic is in the chain (score 1)

https://quant.stackexchange.com/a/41610

Depends on the portfolio, know that is not very helpful! So here is a simple test. Try a reasonable base, say 100 times 10k, then increase the number of timesteps, see how much the results change by, and then do the same for the number of simulations. Whichever produces bigger delta, wins! Repeat with a different base to ensure results are robust!

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.