Variance
Not all variation requires action. Knowing when to react and when to wait is half of operational judgement.
Two kinds of cause
Last month’s revenue was £1.2m. This month it’s £1.1m. Do you have a problem?
Depends on whether the drop is signal or noise.
Common cause variation is the natural fluctuation in any process. Sales vary month to month. Conversion rates wobble. Customer support tickets spike and dip. This variation is built into the system — it happens even when nothing changes.
Special cause variation is a departure from the normal pattern. A new competitor launched. A key salesperson left. The product broke. Something specific caused the change.
The distinction matters because the responses are completely different.
Why it matters
Reacting to noise is expensive. If monthly revenue naturally varies by ±10% and you investigate every dip, you’ll spend your life chasing phantoms. Worse, you might “fix” things that weren’t broken — adding process, changing incentives, reorganising teams — all in response to randomness.
Ignoring signal is dangerous. If a genuine problem emerges and you dismiss it as normal variation, you lose response time. By the time the pattern is undeniable, you’re months behind.
The skill is distinguishing between them before it’s obvious.
How to tell the difference
Understand the baseline. What’s the normal range? If revenue has fluctuated between £1.0m and £1.3m for the past twelve months, £1.1m is unremarkable. If it’s been steady at £1.25m ± £50k, a drop to £1.1m is a signal.
Look for patterns, not points. A single data point tells you almost nothing. Three months of decline is a pattern. A sudden step-change that persists is a signal. A spike that reverts is probably noise.
Check for known causes. Did something specific change? New pricing, lost customer, product issue, seasonal effect? If you can point to a cause, you’re probably seeing special cause variation. If nothing changed and the number moved, it’s more likely noise.
Compare to similar processes. Is this metric moving while others stay stable? If conversion rate dropped but traffic, pricing, and product are unchanged, the drop might be real. If everything is wobbling together, you might be seeing correlated noise — or a shared underlying cause.
The control chart intuition
Statistical process control uses control charts: plot the data over time, draw lines at ±2-3 standard deviations from the mean. Points inside the lines are common cause. Points outside, or patterns like seven consecutive points trending one direction, suggest special cause.
You don’t need formal control charts to apply the logic:
- Know the typical range (informally, “what’s normal”)
- Notice when something falls outside that range
- Notice when you see runs in one direction
- Investigate outliers and runs; don’t investigate every fluctuation
The goal is calibrated attention — responding to signals, ignoring noise.
Where people go wrong
Over-reacting to single points. One bad month triggers a strategy review. One good month triggers celebration. Both are premature. Wait for the pattern.
Under-reacting to trends. “It’s just noise” becomes an excuse to ignore gradual decline. If the trend persists across multiple periods, it’s not noise — even if each individual point is within historical range.
Assuming all variation is fixable. Some variation is inherent. You can’t reduce monthly revenue variance to zero unless you have perfect visibility into buying decisions. Trying to eliminate all fluctuation adds cost and complexity for no gain.
Treating correlation as causation. Revenue dropped the month you changed the website, so the website change caused the drop. Maybe. Or maybe revenue was going to drop anyway, or the sample is too small to draw conclusions. See: Small Samples.
The practice
Establish baselines before you need them. Track key metrics long enough to know what normal variation looks like. When something happens, you’ll have context for whether it’s unusual.
Set thresholds for investigation. Not every dip warrants a deep dive. Decide in advance: movements less than X% or lasting fewer than Y periods get monitored, not investigated.
Separate monitoring from reacting. Track everything; react selectively. The dashboard shows all the noise. Your attention should focus on the signals.
When in doubt, wait one more period. If you’re unsure whether a change is signal or noise, another data point usually clarifies. The cost of waiting is small; the cost of reacting to noise is large.
Variation is information. Some of it tells you something changed. Most of it tells you the system is running normally. Learning to tell the difference is how you stay calm when others panic — and alert when others shrug.
Connects to Library: Systems Thinking — Variation as system behaviour, not individual events.