14.2 Monte Carlo Simulation & Probability Distributions in Cost/Schedule

14.2 Monte Carlo Simulation & Probability Distributions in Cost/Schedule

This section covers 6.D review/interpret risk analysis and impact on a specific project. The blueprint verb is interpret, not run. CCP items on Monte Carlo typically hand you an output — an S-curve, a P-value, a tornado diagram — and ask what it means for a specific project decision, rather than asking you to configure the simulation.

Deterministic cost and schedule estimating relies on single-point values (e.g., 'Labor will cost $$4,500,000$' or 'Turbine installation will take 45 days'). However, complex industrial engineering projects rarely behave deterministically. As observed by mathematician Sam L. Savage, relying on single-point averages leads to the 'Flaw of Averages': plans based on average assumptions will, on average, fail.

Under AACE International Total Cost Management (TCM) standards, Monte Carlo Simulation (MCS) is the recognized standard for quantitative risk analysis (QRA). By replacing single-point inputs with probability distribution functions (PDFs) and recalculating the project model across thousands of randomized iterations, MCS generates a complete Cumulative Distribution Function (CDF)—commonly known as the S-Curve—allowing project stakeholders to establish risk-adjusted budgets and schedules at defined confidence levels (e.g., $P_{50}, P_{80}, P_{90}$).

---

1. Monte Carlo Simulation Mechanics & Sampling Computational Algorithms

+-----------------------------------------------------------------------------+
|                   MONTE CARLO SIMULATION COMPUTATIONAL FLOW                 |
|                                                                             |
|   1. DEFINE DETERMINISTIC BASE MODEL:                                       |
|      Cost Model: Total Cost = \sum (Quantity_i \times Unit_Rate_i)          |
|      Schedule Network: CPM Critical Path Logic (Precedence Diagramming)     |
|                                                                             |
|   2. ASSIGN PROBABILITY DISTRIBUTIONS TO UNCERTAIN VARIABLES:               |
|      Assign appropriate PDFs (Triangular, Beta/PERT, Lognormal) to items    |
|                                                                             |
|   3. DEFINE CORRELATION COEFFICIENTS (r_ij):                                |
|      Model dependencies between interrelated items (e.g., labor rates)      |
|                                                                             |
|   4. ITERATIVE STOCHASTIC SAMPLING (5,000 - 10,000 RUNS):                   |
|      Draw random values from each input PDF; compute total project outcome  |
|                                                                             |
|   5. GENERATE STATISTICAL OUTPUT DISTRIBUTIONS:                             |
|      Construct Histogram (PDF) and Cumulative Probability S-Curve (CDF)     |
+-----------------------------------------------------------------------------+

Simple Random Sampling vs. Latin Hypercube Sampling (LHS)

A critical distinction in simulation mechanics is the method used to sample input distributions:

+-----------------------------------------------------------------------------+
|                        SAMPLING METHOD COMPARISON                           |
|                                                                             |
|   A. SIMPLE RANDOM SAMPLING (Standard Monte Carlo):                         |
|      - Draws pseudo-random numbers purely by chance.                        |
|      - Risk of "clustering" in central regions and under-sampling tails.    |
|      - Requires very large sample sizes (e.g., 20,000+ iterations) to reach |
|        stable convergence in extreme percentiles (P10, P90).                |
|                                                                             |
|   B. LATIN HYPERCUBE SAMPLING (LHS - Stratified Sampling):                  |
|      - Divides the cumulative probability range [0, 1] into N equal, non-   |
|        overlapping intervals (where N = number of iterations).              |
|      - Samples exactly once from each interval without replacement.         |
|      - Guarantees uniform coverage across the entire distribution, including|
|        extreme low-probability / high-impact tails.                         |
|      - Reaches statistical convergence with 70% to 80% fewer iterations     |
|        (typically 1,000 to 5,000 runs).                                     |
+-----------------------------------------------------------------------------+

---

2. Common Probability Distribution Functions in Cost Engineering

Selecting the correct probability distribution is vital. Fitting the wrong distribution to a cost or schedule element introduces model error that invalidates simulation outputs.

+-----------------------------------------------------------------------------+
|                   COMMON COST ENGINEERING DISTRIBUTIONS                     |
|                                                                             |
|   1. TRIANGULAR: Tri(min, mode, max) -> Tri(a, m, b)                        |
|      - Mean: \mu = (a + m + b) / 3                                          |
|      - Variance: \sigma^2 = (a^2 + m^2 + b^2 - am - ab - bm) / 18           |
|                                                                             |
|   2. BETA / PERT: PERT(min, mode, max) -> PERT(a, m, b)                    |
|      - Mean: \mu = (a + 4m + b) / 6                                         |
|      - Standard Deviation: \sigma = (b - a) / 6                             |
|      - Variance: \sigma^2 = (b - a)^2 / 36                                  |
|                                                                             |
|   3. LOGNORMAL: Lognormal(\mu_L, \sigma_L)                                  |
|      - Strictly positive (X > 0); highly positively skewed (long right tail)|
|      - Models unbounded upside cost growth / severe schedule delays         |
|                                                                             |
|   4. NORMAL (GAUSSIAN): N(\mu, \sigma^2)                                    |
|      - Symmetrical bell curve (Mean = Median = Mode)                        |
|      - Governed by Central Limit Theorem for aggregated independent items   |
|                                                                             |
|   5. UNIFORM: U(min, max) -> U(a, b)                                        |
|      - Constant probability density: f(x) = 1 / (b - a)                     |
|      - Mean: \mu = (a + b) / 2; Models total uncertainty / maximum ignorance|
+-----------------------------------------------------------------------------+

Comparative Distribution Selection Matrix

---

3. Interpreting Simulation Outputs: S-Curves and Percentiles

The primary output of a Monte Carlo cost simulation is the Cumulative Distribution Function (CDF), commonly plotted as an S-Curve.

+-----------------------------------------------------------------------------+
|                        CONFIDENCE PERCENTILE METRICS                        |
|                                                                             |
|   - P10 (Optimistic):        10% probability that final cost will be <= P10 |
|                              (90% chance of cost overrun).                  |
|                                                                             |
|   - P50 (Median):            50% probability that final cost will be <= P50 |
|                              (Equal chance of overrun or underrun).         |
|                                                                             |
|   - Mean (Expected Value):   Statistical arithmetic average of all runs.    |
|                              In positively skewed models: Mean > P50.       |
|                                                                             |
|   - P80 (Sanction Budget):   80% probability that final cost will be <= P80 |
|                              (Standard corporate investment benchmark).     |
|                                                                             |
|   - P90 (Conservative):      90% probability of non-exceedance; used for    |
|                              firm fixed-price bidding and high-risk scopes. |
+-----------------------------------------------------------------------------+

Mathematical Derivation of Contingency from CDF

Under AACE standards, the quantitative contingency requirement is derived directly from the CDF percentile corresponding to the organization's risk tolerance:

$\text{Contingency Amount } (P_{80}) = \text{Simulated Cost at } P_{80} - \text{Deterministic Base Estimate}$ $\text{Contingency Percentage } (P_{80}) = \left( \frac{\text{Simulated Cost at } P_{80} - \text{Base Estimate}}{\text{Base Estimate}} \right) \times 100$

Sensitivity Analysis: Tornado Diagrams

During simulation, software tracks the correlation between each individual input variable and the total project outcome. This is displayed as a Tornado Diagram:

• Ranks risk drivers in descending order of their impact on total cost/schedule variance.
• Uses Spearman Rank Correlation Coefficients ($r_s$) or Pearson Correlation ($r$).
• Pinpoints the 'Vital Few' risk drivers (e.g., piping craft labor productivity, steel market escalation) that contribute $80%$ of total bottom-line uncertainty.

---

4. Correlation and Dependencies Between Cost Elements

In complex capital projects, cost elements rarely move independently. For example, if severe local wage inflation impacts structural steel labor, it will almost certainly impact piping, electrical, and mechanical labor simultaneously.

+-----------------------------------------------------------------------------+
|                   THE CRITICAL IMPACT OF CORRELATION (r_ij)                 |
|                                                                             |
|   1. ZERO CORRELATION (r = 0) ASSUMPTION (The Independence Fallacy):        |
|      - When items are modeled as independent, positive variances in one item|
|        cancel out negative variances in another (Diversification Effect).   |
|      - Result: ARTIFICIALLY NARROWS THE TOTAL S-CURVE; severely understates |
|        P80/P90 contingency requirements.                                    |
|                                                                             |
|   2. POSITIVE CORRELATION (0 < r <= 1.0):                                   |
|      - Items vary in tandem. When labor rates rise, all labor accounts rise.|
|      - Result: WIDENS THE TOTAL S-CURVE SPREAD; increases standard deviation|
|        and establishes realistic, robust contingency amounts.               |
|                                                                             |
|   3. CORRELATION MATRIX (Pearson / Spearman):                               |
|      - Cost engineers construct an n x n matrix defining correlation        |
|        coefficients between related WBS accounts.                           |
+-----------------------------------------------------------------------------+

---

5. Quantitative Schedule Risk Analysis (QSRA) & Merge Bias

Applying Monte Carlo simulation to project schedules is known as Quantitative Schedule Risk Analysis (QSRA). QSRA replaces deterministic activity durations with probability distributions and runs Critical Path Method (CPM) forward and backward passes on every iteration.

The Merge Bias Phenomenon (Parallel Path Convergence)

In deterministic CPM scheduling, the total project duration is governed strictly by the single longest path. However, when two or more parallel activity paths converge at a common milestone (a merge node), deterministic schedules systematically underestimate project completion dates.

+-----------------------------------------------------------------------------+
|                        THE MERGE BIAS MATHEMATICAL PROOF                    |
|                                                                             |
|   Suppose three independent parallel paths converge at Milestone M:         |
|   - Path 1 has an 80% probability of finishing by Day 100: P(T_1 <= 100) = 0.80
|   - Path 2 has an 80% probability of finishing by Day 100: P(T_2 <= 100) = 0.80
|   - Path 3 has an 80% probability of finishing by Day 100: P(T_3 <= 100) = 0.80
|                                                                             |
|   DETERMINISTIC FALLACY:                                                    |
|   Project manager assumes Milestone M has an ~80% chance of on-time finish. |
|                                                                             |
|   PROBABILISTIC TRUTH (Joint Probability):                                  |
|   Because Milestone M cannot start until ALL THREE paths are complete:       |
|   P(Milestone M <= 100) = P(T_1 <= 100) * P(T_2 <= 100) * P(T_3 <= 100)    |
|                         = 0.80 * 0.80 * 0.80 = 0.512 (51.2%)                |
|                                                                             |
|   CONCLUSION:                                                               |
|   Merge bias drops the actual milestone confidence from 80% down to 51.2%!  |
+-----------------------------------------------------------------------------+

Schedule Criticality Index ($CI$)

Unlike deterministic CPM where an activity is either $100%$ on the critical path or not on it ($0%$), QSRA calculates the Criticality Index ($CI$):

$CI = \left( \frac{\text{Number of Iterations Activity is on the Critical Path}}{\text{Total Simulation Iterations}} \right) \times 100$

• An activity with high float in the deterministic schedule may have a $CI = 45%$, meaning that in nearly half of all simulated scenarios, duration growth consumes its float and pushes it onto the driving critical path.

---

6. Comprehensive Worked Step-by-Step Engineering Case Study

Subsea Pipeline Installation Cost Risk Model: A cost engineer builds a risk model for three major work packages with the following three-point estimates:

1. Package 1: Subsea Trenching & Dredging (Beta/PERT Distribution):
   • Optimistic ($a_1$): $$12,000,000$
   • Most Likely ($m_1$): $$18,000,000$
   • Pessimistic ($b_1$): $$36,000,000$
2. Package 2: Pipe Lay Vessel Charter (Triangular Distribution):
   • Minimum ($a_2$): $$20,000,000$
   • Mode ($m_2$): $$25,000,000$
   • Maximum ($b_2$): $$42,000,000$
3. Package 3: Hyperbaric Tie-in & Hydrotesting (Normal Distribution):
   • Mean ($\mu_3$): $$10,000,000$
   • Standard Deviation ($\sigma_3$): $$2,000,000$

Deterministic Base Estimate (Sum of Most Likely / Mode values): $\text{Base Cost} = \$18,000,000 + \$25,000,000 + \$10,000,000 = \mathbf{\$53,000,000}$

---

Step 1: Calculate Mean and Variance for Each Package

1. Package 1 (Beta/PERT): $\mu_1 = \frac{a_1 + 4m_1 + b_1}{6} = \frac{\$12M + 4(\$18M) + \$36M}{6} = \frac{\$12M + \$72M + \$36M}{6} = \frac{\$120M}{6} = \mathbf{\$20,000,000}$ $\sigma_1 = \frac{b_1 - a_1}{6} = \frac{\$36M - \$12M}{6} = \frac{\$24M}{6} = \mathbf{\$4,000,000}$ $\sigma_1^2 = (\$4,000,000)^2 = 16.0 \times 10^{12} \; \text{($\$^{2}$)}$

2. Package 2 (Triangular): $\mu_2 = \frac{a_2 + m_2 + b_2}{3} = \frac{\$20M + \$25M + \$42M}{3} = \frac{\$87M}{3} = \mathbf{\$29,000,000}$ $\sigma_2^2 = \frac{a_2^2 + m_2^2 + b_2^2 - a_2 m_2 - a_2 b_2 - m_2 b_2}{18}$ $\sigma_2^2 = \frac{20^2 + 25^2 + 42^2 - (20)(25) - (20)(42) - (25)(42)}{18} = \frac{400 + 625 + 1764 - 500 - 840 - 1050}{18} = \frac{399}{18} = \mathbf{22.167} \times 10^{12}$ $\sigma_2 = \sqrt{22.167} \approx \mathbf{\$4,708,150}$

3. Package 3 (Normal): $\mu_3 = \mathbf{\$10,000,000}$ $\sigma_3 = \mathbf{\$2,000,000}$ $\sigma_3^2 = (\$2,000,000)^2 = 4.0 \times 10^{12} \; \text{($\$^{2}$)}$

---

Step 2: Combine Statistical Moments (Assuming Independence)

1. Total Expected Cost (Mean $\mu_{\text{total}}$): $\mu_{\text{total}} = \mu_1 + \mu_2 + \mu_3 = \$20,000,000 + \$29,000,000 + \$10,000,000 = \mathbf{\$59,000,000}$ (Notice that the Expected Mean of $$59.0M$ is already $$6.0M$ above the deterministic base estimate of $$53.0M$ due to positive skewness!).

2. Total Variance ($\sigma_{\text{total}}^2$) and Standard Deviation ($\sigma_{\text{total}}$): $\sigma_{\text{total}}^2 = \sigma_1^2 + \sigma_2^2 + \sigma_3^2 = 16.0 + 22.167 + 4.0 = \mathbf{42.167} \times 10^{12}$ $\sigma_{\text{total}} = \sqrt{42.167 \times 10^{12}} = \mathbf{\$6,493,612}$

---

Step 3: Compute Risk-Adjusted Percentile Budgets and Required Contingency

By the Central Limit Theorem, the sum of multiple distributions approximates a Normal distribution with $\mu = $59,000,000$ and $\sigma = $6,493,612$:

• $P_{50}$ (Median = Mean for Normal): $\mathbf{$59,000,000}$
• $P_{80}$ Confidence Level ($Z_{0.80} = +0.8416$): $P_{80} = \mu_{\text{total}} + (0.8416 \times \sigma_{\text{total}}) = \$59,000,000 + (0.8416 \times \$6,493,612) = \$59,000,000 + \$5,465,024 = \mathbf{\$64,465,024}$
• $P_{90}$ Confidence Level ($Z_{0.90} = +1.2816$): $P_{90} = \$59,000,000 + (1.2816 \times \$6,493,612) = \$59,000,000 + \$8,322,213 = \mathbf{\$67,322,213}$

---

Step 4: Contingency Allocation Table

+-----------------------------------------------------------------------------+
|                        CONTINGENCY DETERMINATION TABLE                      |
|                                                                             |
|   Metric                 Cost ($)          Contingency ($)   Contingency (%)|
|   --------------------   ---------------   ---------------   ---------------|
|   Deterministic Base     $53,000,000       $0                0.0%           |
|   P50 (Median)           $59,000,000       +$6,000,000       +11.3%         |
|   P80 (Sanction Budget)  $64,465,024       +$11,465,024      +21.6%         |
|   P90 (High Confidence)  $67,322,213       +$14,322,213      +27.0%         |
+-----------------------------------------------------------------------------+

[!IMPORTANT] Exam Takeaway: Setting a budget at the deterministic base estimate ($53M) provides less than a $17%$ probability of success ($Z = (53-59)/6.494 = -0.924 \rightarrow P = 17.8%$). To achieve corporate governance compliance at $P_{80}$, management must authorize $$11,465,024$ ($21.6%$) in contingency.