How Analytical, Numerical, and Machine Learning Methods Differ in Option Pricing
Summary
The document clarifies that an option pricing model and a method for solving or evaluating it are separate concepts. Black–Scholes is a model with a closed-form solution under its assumptions, but the same model can also be evaluated using numerical techniques or Monte Carlo simulation. Monte Carlo and numerical solvers are therefore computational approaches, not distinct models by themselves.
The answer describes finite-difference solutions to pricing equations and simulation of underlying paths, noting that both can produce values for the same modeled dynamics. Simulation can be especially useful for path-dependent or exotic derivatives because the payoff depends on the evolution of the underlying through time, though it can require substantial computation. It also discusses limits on machine learning for pricing: market quotes and implied-volatility surfaces often already provide useful information, while complex models may require market data that is difficult to obtain. The appropriate choice depends on the product and the dynamics that need to be represented; the document offers conceptual guidance rather than empirical comparisons.
Key ideas
- A pricing model describes underlying dynamics and payoff valuation, while analytical, numerical, and simulation techniques are ways to obtain a price.
- Black–Scholes can have a closed-form solution and can also be solved numerically or through simulation.
- Monte Carlo methods are useful when path evolution matters, including for many exotic or path-dependent payoffs.
- Machine learning may face data access and market calibration limits in practical option pricing.
- The product being priced and its relevant dynamics should guide method selection.
Tags
Full text
# How to identify between Analytical, Numerical and ML Model based option pricing? # How to identify between Analytical, Numerical and ML Model based option pricing? I am new to Quantitiative Finance. Coming from Computer Science domain, I wanted to clear the key distinguishing factor between analytical, numerical and ML based models for option pricing. As far as I have learned, following are the differences between these three. 1- Given the required parameters, we can say that the price calculated for a call or put option using Black-Scholes Formula is an analytical approach as it gives a closed form answer while the underlying asset is operating with Black-Scholes asssumption. 2- If we use Monte-Carlo method to get the price for an option, it will be considered as a numerical approach because to get an answer for a given data point, the algorithm will have to extrapolate `n` paths for the under lying asset using geometric Brownian Motion formula hence making it numerical in nature. 3- If we use any Machine Learning or Deep Learning model given the input parameters from the data, it will be considered as a machine learning based approach as it will use machine learning or deep learning parameteric approach to learn the underlying pattern of the option price and hence predict the value. Please let me know if I am missing something or am mistaken in any of the above concepts. ## Answer by AKdemy (score 3) https://quant.stackexchange.com/a/66476 Leaving ML aside for now. Neither Monte Carlo simulation nor numerical solutions are a model generally speaking. They are methods that can be applied on a model to find solutions. For example, Black Scholes has - a closed form solution - which is the standard text book formula - can be solved numerically - can be used in Monte Carlo simulations In practice, I have not seen ML used (seriously) so far to price options. I think partially because - listed options have prices and you use the information these prices convey to get implied vol surfaces which can be used to price more complex structures - if OTC trading is the dominant way of pricing (e.g. FX options), you have direct IVOL quotes - you only need more complex models like Local Vol (LV), Stochastic Vol (SV), Stochastic Local Vol (SLV) or Shifted LMM (if you look at interest rate options) to price complex derivatives: However, to use ML, you would need a lot of market data - which you usually (unless you are the market maker) don't have access to Numerical methods or MC will usually only be used if there is no closed form solution. Generally, a finite-difference solver of the PDE or MC simulation of the SDE should result in the same value. However, for more exotic path-dependent structures MC is usually preferred despite being more computationally intensive. The reason is intuitively obvious if you require to know expected values of underlying(s) over time, simulating their path directly will be useful. In my opinion, the real question should be what you need to price. That will determine what to use to capture the underlying dynamics of the product you're pricing.
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.