Seeing Signal in a Time Series

Most mistakes in data analysis come from one of two errors: reacting to noise, or ignoring signals. Donald J. Wheeler frames it this way in Understanding Variation, a book about applying statistical process control to management decisions. His central question is useful even outside that context: does a data point represent routine fluctuation, or a genuine change? This guide builds a chart design, layer by layer, that helps answer that question.

1. The Level

The simplest view is to plot the raw values. This reveals the broad trajectory but masks the dynamics. A line rising steadily at 5% per year looks nearly identical to one that grew 8% last year and 2% this year. As Wheeler puts it: "No data have meaning apart from their context."

2. Year-over-Year Change

The year-over-year (YoY) transformation compares each period to the same period twelve months prior. This removes seasonality and reveals acceleration and deceleration that the level chart hides. The chart splits at zero: green above (expansion), red below (contraction).

Toggle between the two views. The level chart shows a smooth recovery from the 2020 shock; the YoY chart reveals a massive positive spike in early 2021, not because anything extraordinary was happening, but because the comparison base from 2020 was so depressed. This is a base effect: a distortion caused by an unusual prior period, not by current conditions.

3. The Trend

A single YoY reading is hard to interpret in isolation. Is +5.2% good or bad? A linear regression fitted to the YoY values provides a rough baseline. The slope tells you whether the series is accelerating or decelerating, and readings above the line are outperforming relative to recent history, while readings below are underperforming.

But a single deviation from trend doesn't necessarily mean anything has changed. Some variation is routine. The question is how much variation is too much.

4. Deviation from Trend

Wheeler's answer is the process behavior chart. Subtract the trend from each observation to get the residual, then compute limits at ±2 standard deviations. Points within the limits are common cause variation: routine fluctuation you should not react to. Reacting to it is what Wheeler calls "tampering," and it usually makes things worse. Points outside are special cause: something has genuinely changed, and it warrants investigation.

In this sample, 7 readings breach the limits, all concentrated around the 2020 shock and its aftermath. The readings in between, even those that look volatile in the YoY chart, fall within the limits.

Wheeler uses ±3σ limits. We use ±2σ as an earlier-warning threshold, accepting more false positives in exchange for faster detection. Wheeler also describes run tests: eight or more consecutive points on the same side of the center line suggest the process has shifted, even without a single breach.

5. The Write-Up

Each chart is paired with a concise sentence that tells the reader what the charts say, so they don't have to reverse-engineer it themselves. Here is the template:

Each clause maps to a specific computation. The trend value and slope come from the linear regression. The gap in basis points and the σ reading come from the residual. The base effect compares the year-ago YoY reading to where the trend was at that date: if the base period was above trend, the current reading is depressed by a tough comp, and vice versa.

The point is to front-load interpretation. A reader scanning twenty series should be able to read the sentence and know immediately whether to look closer or move on.

6. The Complete Picture

Stacked together, the three panels tell different parts of the same story. Today the series is near an all-time high, growing at a moderate rate that tracks its trend, with deviations well within normal bounds. In 2020, the level showed a V-shaped dip, the YoY chart revealed the dip followed by a base-effect spike, and the deviation chart flagged both as special cause.

Every data point invites a reaction. The discipline is knowing which ones deserve one.