Why planners must understand Poisson and natural noise in data
The blind spot most planners share
Most workforce planners use Erlang formulas every working day. Most of them couldn’t tell you what statistical distribution Erlang is built on, why call arrivals fit that distribution, or what natural noise around a forecast looks like. The gap is forgivable — it’s rarely on a training agenda — but it quietly damages three things. Forecast accuracy reporting becomes a defensive exercise rather than a meaningful one. Staffing decisions chase variance the maths says you cannot eliminate. And the conversations with finance and operations get harder than they need to be because the planner can’t explain why the headline numbers refuse to behave.
This article walks through Poisson in plain language, why arrivals follow it, what natural noise means, the staffing implications, the accuracy-reporting implications, and the three sentences every planner should be able to say when somebody asks why the model was “wrong” for an interval that came in 30% above forecast.
What Poisson actually is — in one paragraph
A Poisson distribution describes how many independent, rare events arrive in a fixed window of time, when the average rate of arrival is known but the precise timing is not. Telephone calls arriving at a contact centre fit this almost perfectly. Each customer’s decision to call is independent of the others. The average rate is forecastable. The exact moment any individual call arrives is random. Given those three conditions, the count of arrivals in any 15- or 30-minute interval follows a Poisson distribution, with one mathematically inescapable property: the variance equals the mean. If your forecast for an interval is 100 calls, the standard deviation of the actual arrivals around that 100 is the square root of 100 — ten calls. Two-thirds of the time, you’ll see between 90 and 110. Around 5% of the time you’ll see fewer than 80 or more than 120. That spread isn’t forecast error. It’s the maths.
Why this matters — the variance equals the mean
The property “variance equals the mean” has consequences most planners haven’t consciously worked through. Three follow directly.
1. Small intervals look noisier than large ones, even when the model is right. A 15-minute interval forecast at 25 calls has a standard deviation of 5 (square root of 25). That’s a 20% coefficient of variation. A daily forecast of 4,000 calls has a standard deviation of about 63 — only 1.6% of the mean. Same forecasting model; same accuracy in absolute terms; very different appearance in percentage terms. The 15-minute interval will look horrendously inaccurate by MAPE; the daily forecast will look excellent. Neither is wrong; the maths is just unforgiving at small intervals.
2. There’s a hard floor on accuracy. No statistical model can do better than the natural noise of the underlying process. A perfect forecast of 100 calls per interval will still produce actuals from 80 to 120 quite often. The planner who promises “5% MAPE on intervals” for a queue averaging 25 calls is promising something the statistics make impossible. Knowing this saves the planner from being held to a target the maths can’t deliver.
3. Staffing has to absorb the variance. If you staff exactly to the mean forecast, half the intervals will be under-staffed. The Erlang inputs are not the only variability the staffing has to handle. The natural noise of arrivals around the mean has to be absorbed too — through buffer, through flex, through real-time response. Operations that staff to the forecast and treat noise as forecast error end up with a structurally unstable model.
What “natural noise” means in operational terms
Natural noise is the variance around the mean that no forecasting model, however sophisticated, can predict away. It comes from the random timing of independent customer decisions to call. You can model the rate — the seasonality, the time of day, the day of week, the weather effect — but you cannot model which individual customer chooses to ring at 14:23 versus 14:24. That randomness is irreducible. The natural noise produced by it is mathematically calculable but operationally unforecastable.
For practical planning, this means: the interval-level “variance” the planning team agonises over is partly genuine forecast error (signal you could have caught) and partly natural noise (signal nobody could have caught). Distinguishing the two is the difference between a planning improvement effort that pays back and one that chases unicorns.
The accuracy-reporting implications
Three reporting habits that change once the planner understands Poisson.
Report accuracy at multiple aggregation levels. Daily, weekly, monthly — not just interval. The 15-minute MAPE will look bad because the maths makes it look bad. The weekly and monthly figures will show the genuine forecasting performance. Pretending the interval number is the headline number gives operations management an unfair view of the team.
Use WAPE alongside MAPE. Weighted Absolute Percentage Error weights intervals by their actual volume. A 50% miss on a 5-call interval matters less than a 5% miss on a 500-call interval. WAPE captures that; MAPE doesn’t. See forecast accuracy metrics that matter.
Compute the “Poisson floor.” For each interval, calculate the natural-noise band — mean ± one standard deviation. Any actual that falls inside that band is statistically consistent with the forecast. Reporting “how often the actual fell outside the natural-noise band” gives a far better signal of model quality than raw MAPE. Operations that adopt this find their accuracy conversations land differently.
The staffing implications
The planner who has understood Poisson schedules differently. Three habits.
Staff to the mean plus a small buffer. Erlang already accommodates some randomness, but the variance in actual arrivals around the mean is a separate effect. Buffer of 5–10% on small queues, less on large ones (because the relative variance shrinks).
Use real-time flex, not over-staffing, to absorb noise. Standing capacity to cover the worst-case interval is expensive. Maintaining a real-time response capability — aux-code discipline, holding back optional training, having a small flex pool — is cheaper and produces the same outcome.
Don’t over-react to one bad interval. One 15-minute interval that came in 30% above forecast is, on small queues, well inside the natural-noise band. Two bad days in a row are signal; one bad interval is noise. Operations that escalate every interval miss to a planning review burn the planning team out chasing things that the maths predicts will happen routinely.
The conversation with finance and operations
The maths is also a leadership tool. The three sentences every planner should be able to say:
1. “Telephone arrivals follow a Poisson distribution, which means the natural variability around the forecast equals the square root of the volume. Below about 50 contacts an interval, the percentage spread looks dramatic even when the model is right.”
2. “The number we should hold the forecasting team to is daily WAPE, not interval MAPE. Interval-level accuracy is bounded by the statistics, not by the model.”
3. “The staffing model already includes a buffer for natural noise. When we miss an interval, the right question is whether the day or week as a whole was off — not whether one 15-minute window was.”
Planners who can say those three sentences calmly and clearly get a different reception from finance and operations than planners who get defensive about interval-level numbers.
Where Poisson doesn’t fit (and what to do instead)
Poisson’s assumption of independence breaks down in three contact-centre situations.
Marketing-send-driven volume. A burst of contacts triggered by the same external event aren’t independent — they’re correlated. The arrival pattern looks more like a clump than a Poisson process. Standard Erlang under-staffs these by a meaningful margin; the planner should model marketing sends as a known event with its own response curve, not as Poisson noise.
Cascading failures. When a system outage causes customers to ring back repeatedly, the arrivals are positively correlated. The variance is higher than the mean — sometimes much higher. Models that assume Poisson during outages catastrophically under-staff the recovery.
Email and back-office work. Arrival is closer to Poisson, but handle time has nothing to do with the call-centre case. The variance/mean relationship in queue terms still applies, but the staffing model is fundamentally different. See backlog management for non-real-time channels.
How to build the intuition without the textbook
Three exercises a planner can run in a quiet afternoon to build the working intuition.
1. Generate Poisson data in Excel. =POISSON.INV(RAND(),MEAN) produces a Poisson-distributed random number. Generate 100 intervals at mean 25, plot them. See the spread. Run it again. Note how often actuals fall above 30 or below 20. That’s the natural noise you’ll never forecast away.
2. Compute the standard deviation of your own arrivals. Pull the last quarter of 15-minute interval data for a stable queue. Compute the mean and standard deviation. Compare the standard deviation to the square root of the mean. In most queues they’ll be similar — that’s Poisson signing its name. Any persistent excess is real forecasting error worth chasing.
3. Plot the “Poisson floor” over a week of forecasts. For each interval, draw the forecast and the ±1 SD band. Overlay the actuals. Watch how many actuals fall inside the band even when individual intervals look catastrophically off. The picture is the most persuasive teaching tool you have.
Conclusion
Poisson isn’t arcane statistics. It’s the underlying maths of how customer contacts arrive, and the planner who understands it makes better staffing decisions, reports accuracy more honestly, and explains the operation more credibly to finance and operations leadership. The natural noise in arrivals is real, calculable, and unforecastable — not a failure of the model but a property of the world. Planners who internalise this stop chasing the noise and start managing the signal. Three sentences and an afternoon’s exercises are enough to build the intuition; the dividend is paid every working week.
Pair this with forecast accuracy metrics that matter, the Erlang C calculator, the Excel paradox, why weather belongs in your forecast, and composite metrics that hide the truth.