Skip to main content

​
Overview

A single simulation run uses one random draw from every distribution in your model — it’s one possible version of reality. Monte Carlo runs the same configuration many times with different random seeds and reports the distribution of outcomes instead of a point estimate. That’s the difference between “throughput is 847 units/day” and “throughput is 847 ± 23, with a 95th percentile of 891.” Once you can read the distribution, you can reason about confidence: a tight one means the answer is robust to randomness; a wide one means the model’s performance depends heavily on how stochastic events play out. Monte Carlo is the right tool for single-configuration uncertainty quantification. For comparing different configurations side by side, see Experiments. For ad-hoc dashboards across runs, see Insights.
Monte Carlo varies the random seed, not your parameters. It tells you how confident you should be in one Snapshot’s behavior. For parameter sweeps — what if I add a CNC machine? — capture multiple Snapshots and compare them in an Experiment.

​
Before You Run

A Monte Carlo batch runs against the current Run’s configuration. The configuration is fully captured at queue time:
  • The Snapshot the run executed against
  • The schedule that was active
  • The KPIs and charts as defined when you queued the batch
Once queued, the batch is fixed. Editing KPIs, charts, or the model afterwards does not affect the in-flight or completed batch.
KPIs are locked at queue time. If you queue a batch and then edit a KPI definition, the batch keeps the old definition through completion. To pick up KPI changes, queue a fresh batch after the edit lands.

​
Running a Batch

Monte Carlo can be triggered from two places:
  • Single-run Monte Carlo. On any Run’s Results page, click Monte Carlo in the top bar to switch the page into the Monte Carlo view, then click Run Monte Carlo. The modal asks for N Runs — the number of seeds to sweep.
  • Per-Snapshot Monte Carlo inside an experiment. Inside an Experiment, trigger Monte Carlo on each Snapshot’s individual Run page when you want both cross-Snapshot comparison and per-Snapshot confidence intervals.
The N Runs field defaults to 100 and accepts any value between 1 and 1,024.
Default 100 seeds is the right starting point for most models. Push to 200+ when you care about tail metrics (P95, P99) — tails need more samples to stabilize. The hard maximum per batch is 1,024; if you need more, run multiple batches and aggregate.

​
Reading Your Results

A completed batch produces a Monte Carlo Results view with three layers of detail:
  • Statistical summaries for each KPI — mean, standard deviation, median, and percentiles (25th, 75th, 90th, 95th, 99th).
  • Histograms of each KPI’s distribution across all seeds — how often each outcome occurred.
  • Box plots showing the spread, quartiles, and outliers at a glance.
Each individual seed’s run is also accessible — click into a single seed’s row to open its Results page if you want to inspect a specific replication’s event log or charts.

​
Building Charts

Monte Carlo–scoped charts are best for visualizing across replications rather than over simulation time. Useful types:
  • Histograms — distribution of a KPI across seeds
  • Box plots — quartiles and outliers per group (e.g., per resource)
  • Scatter — relationship between two KPIs across seeds (e.g., does throughput correlate with WIP?)
Author them through Dexter the same way you’d author any other chart — describe what you want and Dexter writes the underlying query against the multi-seed result set.

​
Interpreting Results

A few things to watch for as you read the distribution:
  • Tight distributions are reassuring. A small standard deviation relative to the mean means the model’s behavior is robust to randomness — the point estimate is trustworthy.
  • Wide distributions are a signal. A wide spread means stochastic events drive material outcome variance; the mean by itself can be misleading.
  • Tails matter for SLAs. “Average throughput is fine” doesn’t tell you anything about the worst case. P95/P99 is where you find the days a downstream commitment can’t be hit.
  • The model didn’t change. If two batches’ distributions look identical, that’s the system telling you the change you made doesn’t matter under noise — sometimes a useful answer.

​
Common Pitfalls

  • Editing the model mid-batch. The batch uses the configuration captured at queue time. Edits don’t propagate; queue a fresh batch.
  • Confusing Monte Carlo with Experiments. MC = many seeds, one configuration. Experiments = many configurations, one seed each (or many seeds each). Use both when both questions matter.
  • Reading P95 with too few seeds. Tail percentiles are noisy at 50–100 seeds. Bump to 200+ before drawing conclusions about worst-case behavior.
  • Snapshots — the immutable freeze a Monte Carlo batch runs against
  • Experiments — comparing multiple Snapshots side by side, with Monte Carlo per Snapshot
  • Results — the Run-scoped surface where individual seed runs render
  • Insights — durable dashboards that can summarize across batches