Monte Carlo Risk Analysis
Monte Carlo risk analysis is a quantitative approach to project evaluation under uncertainty in which input parameters are represented as probability distributions rather than fixed values, and the model is solved iteratively thousands of times by randomly sampling from those distributions. Each iteration produces one possible outcome; the full collection of outcomes reveals the probability distribution of the result, from the pessimistic tail through the median to the optimistic tail. In oil and gas, the technique is applied across the entire project lifecycle: prospect volumetrics, EUR forecasting, well economics, portfolio optimization, and reserve reporting under SEC and SPE-PRMS frameworks.
Probability Distributions Used in Oil and Gas
The choice of distribution shape for each input parameter is as important as the bounds themselves. Triangular distributions are the most common starting point because they require only three values: minimum, most likely, and maximum. They are used for net pay thickness, net-to-gross ratio, and recovery factor when analogue data is limited. Log-normal distributions are appropriate for parameters that cannot go negative and that have a long right tail, including gross rock volume, permeability, and EUR per well in unconventional plays. Normal distributions apply to well-constrained parameters such as reservoir temperature, fluid gravity, and formation volume factor derived from PVT lab data. Uniform distributions represent complete ignorance within a bounded range and are sometimes used for areal extent in early-stage exploration. Discrete distributions capture binary or categorical risk, including the presence or absence of a trap, seal integrity, and hydrocarbon charge, and are often applied as separate geologic risk multipliers before the volumetric sampling begins.
Correlation Between Input Variables
A common error in Monte Carlo models is treating geologically correlated inputs as statistically independent. In many formations, higher gross rock volume correlates positively with greater net pay and higher average porosity because the same depositional environment that created the structural closure also influenced reservoir quality. Ignoring this correlation artificially narrows the output distribution and underestimates both the upside and the downside. Modern reservoir risk tools allow the analyst to specify a correlation matrix that links input pairs with a rank-correlation coefficient between negative one and positive one. A rank correlation of 0.6 between gross rock volume and porosity, for example, means that when a high GRV value is drawn, there is a 60 percent tendency to also draw a high porosity value. Getting the correlation structure right requires analogue field data, geologic judgment, and often discussion between geologists, petrophysicists, and reservoir engineers before the model is run.
Number of Iterations and Convergence
The number of iterations required to produce stable percentile estimates depends on the number of uncertain inputs, the shape of the output distribution, and the precision required at the tails. For a typical four-variable STOIIP model, 5,000 iterations usually produces a P90 estimate stable to within two or three percent. For highly skewed distributions, for complex economic models with 15 or more inputs, or when the P5 and P95 tails are needed for value-at-risk reporting, 25,000 to 50,000 iterations are commonly used. Running 1,000 iterations is almost always insufficient for stable tail estimates. Most commercial tools (Crystal Ball, @RISK, Petrel Uncertainty, and Python-based frameworks using numpy and scipy) monitor convergence by tracking how much the percentile estimates change as iterations accumulate and can automatically stop when a defined stability criterion is met.
Sensitivity Analysis and Tornado Charts
Once a Monte Carlo run is complete, sensitivity analysis identifies which input variables are driving the most variance in the output. The standard visual tool is the tornado chart, which ranks inputs by their contribution to output variance, with the largest driver at the top. Contribution to variance is typically measured using the Spearman rank correlation between each input and the output across all iterations. An input that accounts for 40 percent of NPV variance warrants the most attention: additional seismic acquisition, a pilot well, or laboratory tests targeted at that variable will reduce overall project uncertainty more efficiently than data collection on lower-ranked inputs. Tornado charts also guide capital allocation decisions within a portfolio by highlighting which parameters, if better constrained, would shift the most projects above the investment hurdle rate.
Portfolio Risk Analysis and Project Ranking
At the portfolio level, Monte Carlo risk analysis aggregates the probabilistic economics of multiple prospects or development projects to estimate the range of total portfolio NPV, capital requirements, and production outcomes. Correlation between projects must again be specified: two prospects in the same basin sharing the same source rock have correlated geologic risk, and two wells in the same play share correlated commodity price exposure. When these correlations are ignored, the portfolio-level distribution is too narrow, underestimating both the chance of a very bad year and the chance of an exceptional one. Operators and investors use portfolio Monte Carlo to answer questions such as: what is the probability that the capital program generates a positive return if oil averages below 60 dollars per barrel for the next three years, and which combination of prospects maximizes expected NPV subject to a capital constraint? These outputs directly inform annual budget submissions, board presentations, and SEC reserve filings.
Key Takeaways
- Monte Carlo risk analysis samples thousands of input combinations from defined probability distributions to produce a full probability distribution of project outcomes, replacing single-point estimates with P10, P50, and P90 results.
- Distribution shape selection matters: log-normal is standard for volumetric and EUR inputs, triangular for expert-bounded estimates, and discrete distributions for binary geologic risk factors.
- Correlation between input parameters must be explicitly modeled; ignoring it artificially narrows the output distribution and can cause significant underestimation of project risk and upside.
- 5,000 to 50,000 iterations are needed for stable results; fewer than 1,000 iterations is insufficient for reliable tail percentile estimates.
- Tornado charts from sensitivity analysis direct data acquisition budgets to the parameters that drive the most output variance, maximizing the value of additional information.