MonteGrid | Articles | Sampling and precision control in MonteGrid
Sampling and precision control in MonteGrid
MonteGrid offers four sampling methods, two run modes and precision control. A run can have up to 1,000,000,000 trials.
Updated
Sampling methods
| Method | Sampling |
|---|---|
| Monte Carlo | Independent random draws. The default. |
| Latin Hypercube (random) | One draw from each of as many equal-probability strata as there are trials, at a random position in the stratum. Less sampling error in the tails for the same number of trials. |
| Latin Hypercube (median) | The same strata, each drawn at its midpoint. The sample mean of each assumption equals the distribution's mean, and histograms are smooth at low trial counts. |
| Quasi-Monte Carlo (Sobol) | A digitally shifted Sobol low-discrepancy sequence. Each assumption takes one dimension of the sequence. |
| MonteGrid | Crystal Ball | @RISK | |
|---|---|---|---|
| Methods | Monte Carlo, Latin Hypercube (random or median), Sobol quasi-Monte Carlo. | Monte Carlo and Latin Hypercube (Oracle). | Latin Hypercube, the recommended setting, and Monte Carlo (Lumivero). |
Note: Correlation reorders the samples of each assumption. In a correlated run, the Sobol sequence keeps its distributions but not its joint structure.
Run modes and trials
| Run mode | Calculation |
|---|---|
| Fast (MonteGrid's engine) | MonteGrid's engine calculates the workbook outside Excel. The default. |
| Excel | Excel recalculates the workbook for each trial. |
If MonteGrid's engine cannot calculate a formula, the run falls back to Excel, and About this run on the Results page names the cell.
A run has 5,000 trials by default and up to 1,000,000,000. Above 1,000,000 trials, a run needs the Fast run mode and Monte Carlo sampling.
A Fast Monte Carlo run with more than 1,000,000 trials, or more than 200 MB of trial data, streams: MonteGrid calculates the statistics as the trials run, and keeps an evenly spaced sample of 100,000 trials for the histogram, the scatter plot and Extract Data.
The same model, settings and random seed give the same results. The default seed is 20260802.
Precision control
Precision control stops the run when the chosen statistic of every forecast is known within the tolerance at 95% confidence, or at the maximum number of trials. The statistic is the mean or a percentile. Precision control requires Monte Carlo sampling.
| MonteGrid | Crystal Ball | @RISK | |
|---|---|---|---|
| Rule | Precision control: the mean or a percentile of every forecast, within a tolerance, at 95% confidence. | Precision control: the run stops "when the calculated confidence interval drops to less than the specified precision" (Oracle). | Convergence monitoring, with a tolerance and a confidence level for each statistic tested (Lumivero). |
The procedure is in Set up run settings.