No shortfalls in a cash simulation does not establish zero risk

A cash model that produces no shortfalls in 100 simulated paths has reported an observation about those paths. It has not demonstrated that a shortfall is impossible. Rare events can be absent from a small simulation even when the model assigns them a nonzero probability. Keep the number of qualifying paths and total independent runs beside the percentage.

Define the event first. It might mean usable cash falling below zero on any business day, cash falling below a chosen reserve, or an inability to meet a payment when due. A model checking only month-end balances can miss an intramonth funding need regardless of how many paths it runs.

NIST describes exact binomial confidence limits for event proportions and notes the limitations of normal approximations with small event counts. Applied to independent Monte Carlo paths from one fixed probability model, that mathematics helps quantify sampling uncertainty in an estimated event probability. It does not validate the underlying business assumptions.

Assume, solely for illustration, that the chosen model assigns a 1% probability to a funding shortfall on each independent simulated path. The probability of seeing no shortfall in 100 paths is 0.99 raised to the power of 100, approximately 36.6%. A zero count would therefore be quite ordinary under that model.

This example explains why a displayed 0.0% needs a denominator and rounding rule. Zero events out of 100 and one event out of 10,000 can both appear as 0.0% when a dashboard uses one decimal place, although the observations and precision differ.

Store the integer event count, unrounded estimated probability and run count. Also retain the definition of a path, such as a complete 13-week cash history. Counting each week as an independent trial would be inappropriate when weeks belong to the same dependent cash path.

For zero events in n independent trials, a one-sided 95% exact upper confidence bound is 1 minus 0.05 raised to the power of 1/n. It solves the probability of observing zero events at the boundary. With 100 runs the bound is approximately 2.95%; with 1,000 runs it is approximately 0.299%.

These are confidence bounds for the event probability inside the fixed simulation model under the stated sampling assumptions. They are not a 95% guarantee that the real company is safe. The calculation also differs from a two-sided interval, so record which convention the report uses.

If the observed count is greater than zero, use a suitable interval calculation for that count. Do not continue applying the zero-event formula or substitute a symmetric percentage range that can become negative. The interval method should match the sampling design and the precision needed for the decision.

A simulation can deliberately connect receipts, costs and dates within each path. Those dependencies may be essential to representing the business. The binomial calculation above concerns independence between complete paths, not independence between every input inside a path.

Repeating the same saved paths does not create a larger effective sample. Nor does rerunning an identical random seed and counting its duplicate output as new evidence. Preserve reproducibility, but distinguish a repeated calculation check from additional independent sampling.

Some simulation methods use structured or dependent sampling to improve numerical efficiency. Their uncertainty assessment may require a different treatment from the simple binomial calculation. Ask the analyst which sampling design was used before attaching the bound above to its output.

If every customer payment is forced to arrive within ten days, the model cannot produce a payment delayed for forty days. Increasing the run count only explores the permitted ten-day range more thoroughly. Review whether input bounds, event definitions and dependencies reflect the uncertainties relevant to the funding decision.

Separate numerical sampling uncertainty from uncertainty about the inputs. The first concerns how precisely the chosen model's probability has been estimated. The second concerns whether that model is a credible description of future conditions. A very narrow numerical interval can sit around a probability produced by weak collection assumptions.

Challenge material exclusions with specific operating evidence. A shared customer parent, a processor hold or a conditional funding draw may connect several receipts. Record how such a case enters the model or show it as a separate scenario if reliable probabilities are unavailable.

If management needs to distinguish a modeled 2% event probability from a 0.2% probability, 100 paths will rarely provide a useful answer. Set a precision objective and review the event count as the simulation grows. Avoid selecting a run count merely because it sounds large on a presentation slide.

Also examine the size and timing of a shortfall. Two models can estimate the same probability while one implies a small temporary draw and the other a large funding need that cannot be arranged quickly. A frequency estimate is only one part of the treasury response.

Preserve a few representative adverse paths with their input events and cash dates. They let the finance owner trace why the balance crossed the threshold and check whether the assumed response was feasible. The path evidence should accompany the probability result rather than disappear behind an aggregate chart.

A useful report states the model version, event definition, horizon, run count, observed count, sampling method and interval convention. For example, it can say that no shortfall appeared in 1,000 independent paths and give the corresponding approximate 0.299% one-sided upper bound within that model.

Attach the major input uncertainties and any excluded scenarios that affect the proposed commitment. Management can then decide whether to gather better evidence, increase the available buffer or change the spending schedule. A zero-event result becomes useful when its numerical precision and operational meaning are both clear.