Monte Carlo Beats Single-Point Underwriting: A Deal Is a Distribution, Not One Number

A deal is a distribution, not one number, and single-point underwriting hides that fact behind a clean IRR. Monte carlo underwriting fixes the omission: instead of one set of assumptions producing one return, it replaces each uncertain input with a probability distribution, runs the model thousands of times, and reports the full range of outcomes. The base-case IRR you present to committee is one draw from that range. It tells you what the deal returns if every assumption holds. It never tells you the odds that they will.

The thesis is simple. A pro forma that ends in a single 10 percent IRR looks precise and behaves like a guess. The same model, simulated, tells you the probability of clearing your hurdle, the probability of losing equity, and how much of the outcome rides on two or three inputs. That is a different quality of information, and it is available from the same spreadsheet.

Key Takeaways

• Single-point underwriting reports one IRR from one set of assumptions. It cannot express uncertainty, so it cannot express risk. A deal is a distribution of outcomes, and one number is a single sample from it.

• Monte carlo underwriting replaces each uncertain input with a probability distribution, runs the discounted cash flow thousands of times, and returns the full spread of IRRs rather than one point estimate.

• Simulation converts a return into a probability of loss. In the worked example below, a 10 percent base-case IRR carries a 34 percent chance of missing an 8 percent hurdle and a roughly 3 percent chance of losing equity, facts the single number conceals.

• Academic work by Hoesli, Jani, and Bender (University of Geneva, 2005) found real estate valuation confidence intervals are most sensitive to the long-term interest rate and terminal value growth. Simulation shows you that concentration directly.

• Monte Carlo is only as honest as its inputs. Garbage distributions produce a precise-looking but false range, which is why the method disciplines assumptions rather than replacing judgment.

Why does single-point underwriting misrepresent a deal's risk?

Single-point underwriting misrepresents risk because it collapses an uncertain future into one number and presents that number as fact. You choose a rent growth, an exit cap rate, and a hold period, the model returns a 10 percent IRR, and the uncertainty vanishes from the page. The output looks decisive. The decisiveness is manufactured by the format, not earned by the analysis.

The problem is not that the base case is wrong. The problem is that it is one outcome among many, and the format gives it no error bars. Two deals can both underwrite to a 10 percent IRR while carrying entirely different risk. One clears 8 percent in nine cases out of ten. The other clears it barely half the time and loses equity in a meaningful tail. The single number treats them as identical because on paper they are.

This is where the CFA Institute curriculum places simulation. Its "Simulation Methods" reading teaches analysts to construct and interpret a Monte Carlo analysis precisely because a point estimate cannot represent the distribution of possible results. A Monte Carlo simulation is the standard answer to a question single-point models cannot even ask: not what will this deal return, but what is the range, and where inside it are the outcomes concentrated.

What is Monte Carlo underwriting, and how does it model a deal?

Monte carlo underwriting replaces each uncertain assumption in a discounted cash flow model with a probability distribution, then runs the model thousands of times, drawing a new value for each input on every pass. The result is not a single IRR but a distribution of thousands of IRRs, which you read at percentiles: the tenth, the median, the ninetieth.

The mechanics are ordinary. Your existing pro forma already computes an IRR from inputs like rent growth, vacancy, expense inflation, and exit cap rate. In a single-point model each of those is one number. In a simulation each becomes a distribution: rent growth might be normal around 3 percent with a standard deviation of 1.5 percent, the exit cap rate normal around 5.75 percent with a spread of 50 basis points. Each trial samples one value from every distribution, computes the resulting IRR, and stores it. Ten thousand trials produce ten thousand IRRs.

The academic literature is direct about why this matters for property. Hoesli, Jani, and Bender, in their 2005 University of Geneva and Swiss Finance Institute paper on Monte Carlo simulation for real estate valuation, use empirical data to build the input distributions and report that the resulting confidence intervals are most sensitive to the long-term equilibrium interest rate and to the expected growth rate of the terminal value. That is a finding you cannot get from a point estimate. It tells you which two assumptions govern the outcome, which is where diligence should concentrate. This is close kin to sensitivity analysis, which isolates one input at a time; simulation moves all of them together and reads the joint result.

How does simulation turn a deal into a probability of loss?

Simulation turns a return into a probability of loss by counting outcomes. Once you have thousands of simulated IRRs, the probability of missing a hurdle is the share of trials below it, and the probability of losing equity is the share below zero. The single number becomes a set of odds, which is the language risk is spoken in.

Consider a worked example. A five-year hold on a 10 million dollar asset, Year 1 net operating income of 550,000 dollars for a 5.5 percent going-in cap rate, financed at 60 percent loan to value with interest-only debt at 6 percent. The single-point base case uses 3 percent annual NOI growth and a 5.75 percent exit cap rate, and it returns a 10.0 percent levered IRR. That is the number that goes on the committee slide. Now hold the fixed terms and let the two most consequential inputs vary.

Running that model 200,000 times produces the distribution below. These IRR percentiles are derived directly from the stated inputs above, not asserted.

The single-point IRR of 10.0 percent is honest as far as it goes: it lands on the median. What it omits is everything around it. A third of the outcomes fall short of an 8 percent hurdle, and the P10 case delivers 3.5 percent, less than the cost of the debt. A committee looking at 10.0 percent sees a clean go. A committee looking at the distribution sees a deal that needs a wider equity cushion or a sharper entry price. Same model, same inputs, different decision. The expert-voice line: a single-point IRR tells you what the deal returns if you are right about everything, and simulation tells you the price of being wrong.

A CFA Institute analysis on the Enterprising Investor blog (Nicolas Rabener, 2024) shows the same shape in portfolio terms, reporting an expected return of 13.45 percent with a bottom-quartile outcome of negative 0.63 percent and an upper-quartile outcome of 25.71 percent. The lesson carries to real estate: the spread is the information, not the midpoint.

How is Monte Carlo different from sensitivity and scenario analysis?

Monte Carlo differs from sensitivity and scenario analysis in how many futures it examines and how it weights them. Sensitivity analysis flexes one input at a time. Scenario analysis bundles a handful of inputs into a few named cases, usually base, upside, and downside. Monte Carlo varies every uncertain input at once, across its full distribution, and produces thousands of probability-weighted outcomes rather than three hand-picked ones.

The three methods answer different questions and are complementary, not rival. The table sorts them.

The gap Monte Carlo closes is probability weighting. A downside scenario tells you the deal returns 4 percent if three things go wrong together. It does not tell you how likely that combination is. Simulation does, because it draws from the input distributions and lets the joint frequencies fall out. The tradeoff is discipline: a simulation is only as good as the distributions you feed it, and inventing tight distributions to force a comfortable range is the modeling equivalent of the optimism this method is meant to catch. For the manual, one-input-at-a-time version and where it still earns its place, see sensitivity analysis: model versus guess. And because simulation only sharpens the assumptions you start with, the discipline of interrogating them stands on its own: the pro forma assumptions buyers should challenge are the same inputs a simulation makes you specify as distributions.

Frequently Asked Questions

Does Monte Carlo underwriting require special software?

No. Any model that computes an IRR from inputs can be simulated. The requirement is not a tool but a shift in how inputs are specified: each uncertain assumption becomes a probability distribution rather than a single value, and the model is run repeatedly. The output is a distribution of returns read at percentiles.

How many simulation trials are enough?

Enough that the reported percentiles stop moving between runs, typically thousands to tens of thousands of trials for a property model. More trials tighten the estimate of the tails, where probability of loss lives. The worked example used 200,000 trials, well past stability, so its P10 and loss figures are steady.

Can Monte Carlo make a bad deal look good?

Yes, if the input distributions are dishonest. The method inherits the quality of its assumptions: narrow, optimistic distributions produce a narrow, optimistic range that looks rigorous and is not. Simulation does not replace judgment about rent growth or exit caps. It forces that judgment into the open by making you state a full distribution instead of hiding it in one number.

Conclusion

Single-point underwriting is not wrong so much as incomplete. It reports one outcome from one set of assumptions and formats it as certainty, which is exactly the information an underwriter should distrust. Monte carlo underwriting keeps the same model and asks a better question: not what does this deal return, but what is the range, how likely is the downside, and which inputs decide it. A 10 percent IRR that clears an 8 percent hurdle two-thirds of the time is a different deal from a 10 percent IRR that clears it nine times in ten, and only the distribution tells them apart. The number was always a sample. Simulation makes the operator look at the distribution instead.