Treating noise as signal (and signal as noise)
The over-fitter’s curse
Contact arrival data is noisy. A 30-minute interval that’s 12% above forecast might be a real pattern shift or might be Poisson volatility doing what Poisson volatility does. The planning team often can’t tell in the moment, and most of them over-react. They tune the forecast model to fit the recent variance, the model now overfits the noise, and the next month it predicts a shift that wasn’t real.
The mirror failure is just as common. A 1.5% week-on-week drift that lasts six weeks is below most operations’ awareness threshold; meanwhile it’s adding 9% to the volume base by week six, and nobody’s caught it because each individual week was “within tolerance.” The model underfits the real signal because the team was filtering it as noise.
Why this happens
The honest answer: Poisson arrivals look surprisingly “patterned” even when they’re completely random. A run of three high intervals followed by two low intervals isn’t evidence of anything; it’s what random data looks like. Humans are pattern-finders by nature, and planners trained to spot trends find them whether they’re there or not.
Meanwhile, small persistent drifts don’t look like patterns — they look like the noise floor moving by a tiny amount, week after week, in the same direction. Each week individually is unremarkable. The cumulative effect isn’t visible unless someone explicitly tracks it.
The four signs you’re reacting to noise
You changed the forecast model in response to one or two periods. A single bad interval almost never justifies a model change. If the model performed well historically, the recent miss is more likely to be variance than structural shift. Wait.
The “trend” reverses within a few periods. Pattern-finding minds latch onto direction, and most apparent trends in short windows are random walks that mean-revert. If the supposed trend reverses, it wasn’t a trend.
You can’t name a real-world cause. A real volume shift almost always has a cause: a marketing campaign, a system change, a regulatory event, a seasonal pattern. If you can’t name the cause, you’re probably looking at noise.
The variance is within the operation’s natural Poisson noise floor. For an interval with a mean of N contacts, the standard deviation is roughly √N. Intervals routinely vary by 10–30% from the mean without anything having happened. Below this threshold, the “variance” is just arithmetic.
The four signs there’s a real pattern you’re missing
The drift is small but persistent. Week-on-week movements of 1–3% in the same direction, sustained over five or more weeks, are almost always real. The cumulative effect is what matters, not the per-week movement.
The drift correlates with a known driver. Volume drifting up while marketing spend is also drifting up. AHT drifting up while a new product launched a month ago. Cumulative bias in the same direction while a regulatory deadline approaches. The correlation is the signal.
The forecast bias is one-sided over months. The model is “wrong” in the same direction in nine of the last ten months. That’s structural, not noise. The tracking-signal diagnostic catches this if anyone’s looking.
The pattern persists at multiple granularities. The drift is visible at daily, weekly, and monthly aggregations. Random noise mostly disappears under aggregation; real signal doesn’t.
The discipline that catches both
Three habits separate the operations that handle noise and signal well from the ones that don’t.
The tolerance grid. Pre-agreed thresholds for “observe” vs “act” on intraday and weekly variances. Anything inside tolerance is logged, not actioned. Anything outside triggers a defined response. This stops the over-reaction to noise without requiring real-time judgement.
The rolling-trend review. Once a week, a 12-week rolling chart of the key metrics (volume, AHT, contact mix, channel split). The eye picks up real drift on a rolling chart that’s invisible on a week-by-week dashboard. Operations that do this catch most small signals before they become big problems.
The cumulative-bias tracker. Forecast vs actual, cumulative, plotted over time. A real bias produces a steady walk away from zero; noise produces a wandering walk that drifts back. The chart is harder to misread than individual months’ data.
The honest case for restraint
The single most common forecasting improvement an operation can make is to react to fewer things. Most short-term variance is noise; most planning interventions in response to noise make the forecast worse, not better. The discipline of not intervening — of letting a 12% interval variance pass without re-tuning the model — is uncomfortable but right. Pair this with sometimes the best thing to do is nothing for the equivalent discipline at the real-time layer.
Conclusion
Treating noise as signal is the most common forecasting failure that comes from inside the planning team. Treating signal as noise is the most common one that comes from operations leadership ignoring small persistent drifts. Both are correctable with discipline rather than methodology: a tolerance grid, a rolling-trend review, a cumulative-bias tracker, and the cultural permission to do nothing when nothing is the right answer. Planners who develop these instincts spend much less time chasing phantom patterns and much more time catching the real ones.
Next in the series: The business-intelligence layer that closes the gap.
Pair this with Poisson and natural noise, sometimes the best thing to do is nothing, and your forecast is probably more accurate than you think.