| Vose Software

Industry: Construction and Infrastructure
Product: ModelRisk
Application: Labor Productivity Analysis


A Highway Schedule Built on P=4.6 m³/Worker-Day — and a 35% Chance of Missing the Deadline

The $310M highway-upgrade package was planned around a single deterministic productivity figure: 4.6 cubic metres of concrete placed per worker-day. Schedules, crew counts and contract durations were all derived from it. The probabilistic rebuild in ModelRisk — sampling weather, crew skill, equipment availability and fatigue as separate distributions, with weather treated as a single project-level shared factor — showed the realised productivity median was closer to 4.3 m³/wd, and the probability of meeting the 396-working-day contract was only 58%. A productivity number that looked precise to two significant figures was, in fact, a distribution whose right tail mattered to the contract.

The error is structural: industry-standard productivity rates are averages over conditions that vary day-to-day and crew-to-crew. Using the average as if it were a constant is the deterministic mistake. The probabilistic alternative — treating productivity as the product of four independent-but-shared factors — gives the schedule its real shape.

Four factors, one shared

Productivity P (m³/worker-day) is modelled as a multiplicative factor stack:

\[ P = P_0 \cdot W \cdot S \cdot E \cdot F \]

  • P₀ = 4.6 m³/wd (the deterministic baseline from the estimating manual).
  • W = weather index, LogNormal(0, 0.16) — sampled once per day and applied to every crew that day. This shared factor is the critical one: independence assumption would have given every crew its own weather and dramatically tightened the project-finish distribution.
  • S = crew skill, Triangular(0.78, 1.00, 1.15) per crew — sampled once per crew at project start (skill mix is persistent).
  • E = equipment availability, Bernoulli mixture — 95% of days at 1.0, 5% of days at 0.55 (an equipment shortage shaves productivity by 45%).
  • F = fatigue cycle, a 6-day rolling decay from 1.0 down to 0.86, capturing the well-documented productivity decline over consecutive working days without a rest.

The shared weather factor is the single biggest reason a probabilistic model differs from a deterministic one. On a wet day every crew is slower, every cubic metre takes longer, and every dependent task starts late. Independence-assumption Monte Carlo gives back something close to the deterministic mean and almost nothing in the tail.

Where the shared weather actually shows up

A single representative project realisation — 14 crews × 18 weeks — lays the productivity field on a heatmap. Whole columns (project weeks) move together because the weather index is shared across crews; whole rows (crews) drift because of the persistent crew-skill multiplier:

Productivity by crew x week — shared weather drives whole columns together

The yellow-green columns are the dry, well-staffed weeks (productivity 4.4–5.1 m³/wd); the red columns are the bad-weather weeks where every crew slows together (productivity dips to 3.4–3.8 m³/wd). The row variation is real but smaller — the difference between the strongest and weakest crew on a typical week is about 0.5 m³/wd, while a bad-weather week shaves 1.0+ off every crew. Aggregated across 20,000 such project realisations, the simulated project-average productivity has mean ≈ 4.28 m³/wd, P10 ≈ 3.90, P90 ≈ 4.65 — strictly below the deterministic baseline 4.6 by about 7%. The deterministic plan implicitly assumed a productivity higher than the simulated median, because the asymmetric tail to the low side is dominated by the shared-weather columns.

Cumulative concrete placed — the trajectory that matters

The 142,000 m³ scope, at the deterministic productivity, plots to a single straight line that crosses the scope target at day 376 — about a 20-day cushion against the 396-day contract. The simulated trajectories tell the real story:

Cumulative concrete placed — 142,000 m3 scope, 396-day contractual deadline

The deterministic line is the dashed straight trajectory. The P50 simulated path crosses the 142,000 m³ scope line at roughly day 389; the P10 (lucky) path crosses around day 354; the P90 (unlucky) path crosses around day 433. P50 finish ≈ 389 days, P90 ≈ 433 days; the probability of missing the 396-day contract is 42%. The 20-day deterministic buffer was a coin-flip, not a guarantee. The wedge between the P10 and P90 trajectories widens steadily through the project — bad weeks compound rather than self-correct, which is exactly the asymmetric tail that an independence-assumption Monte Carlo misses.

What actually drives the late tail

Tornado: drivers of P90 finish (concrete placement)

The shared weather index is the largest single mover of P90 finish — bigger than the per-crew skill mix or per-day equipment shortage. Crew skill is second; equipment availability is third. The fatigue cycle and quantity-scope-creep show up as smaller but real contributors. This ranking directs mitigation investment: weather sheds and outdoor-work scheduling first; targeted crew training second; backup equipment third.

How much mitigation buys what confidence?

A three-part mitigation package — crew training, backup equipment pool, shift-rotation policy — collectively reduces the weather-shock sigma from 0.16 to 0.06. The threshold-sweep below treats spend as a continuous lever and reads off the probability of meeting the contract:

P(finish on time) vs productivity-mitigation spend

At zero mitigation spend, the probability of meeting the 396-day contract is 58% (red curve) — and the 416-day cushion (a 5% schedule buffer) is met 79% of the time (green curve). The full $1.2M package — training + spares + shift rotation — lifts the on-time probability to ~99%, with the red curve crossing the 90% target around $800k of spend and the green cushion curve clearing 90% far earlier, around $330k. The curve is steepest in the first ~$500k of spend (where training and basic spares retire the largest chunk of weather-shock dispersion) and then flattens — additional spend buys diminishing returns because the residual variability is the fundamental weather randomness no training programme can shape away. The mitigation cost is roughly $1.2M; the expected liquidated-damages exposure avoided (at $80k/day × expected days late) is several times that. The tail protection — going from a 42% to a ~1% probability of breach — is what convinced the executive committee.

What the model changed

  • Productivity baseline rebuilt as P₀ = 4.6 with the shared-weather and skill-mix structure on top — not as a single point estimate.
  • Crew-training programme funded ($420k) ahead of schedule pressure, justified by the 7-point reduction in P(late).
  • Three days of float added to the critical-path schedule at the planning stage, eliminating an early-warning sign the deterministic schedule masked.
  • Weather contingency clauses written into the construction-management contract using the simulation's exceedance probability — replacing the boilerplate "abnormal weather" language with quantitative thresholds.
  • Two pending bids re-priced by the estimating department after the same productivity stack was applied, lifting margins on weather-exposed packages by 3%.

ModelRisk Functionality Used

  • Shared-factor LogNormal weather index sampled once per simulated day and applied to every crew — the correlation structure that independence-assumption tools cannot represent.
  • Triangular crew-skill draws persistent across the project life — capturing the fact that the skill mix you start with is mostly the skill mix you have.
  • Bernoulli-mixture equipment availability — separating the typical day (1.0) from the small-probability shortage day (0.55) where productivity is materially impaired.
  • Tornado of P90 finish drivers — pointing to weather mitigation first, training second, spares third, in a different order than the deterministic ranking would have given.
  • Mitigation-package CDF comparison that priced the value of training + spares + rotation against the tail of missed-deadline probability, not against the mean.

Productivity is not a number; it is the product of factors that share weather and crews. Monte Carlo simulation in ModelRisk is what makes that product visible — and once it is visible, the difference between "the schedule has a 20-day buffer" and "the schedule has a 65% chance of holding" stops being a rhetorical distinction and becomes a contracting one.