The tracking signal

Definition

The tracking signal is an alarm that monitors the forecast continuously. It divides the cumulative signed error by the mean absolute deviation: while it stays near zero, the forecast is healthy; when it crosses a threshold, a bias has set in and calls for intervention. It is the planner's sentinel.

Why it matters

The sentinel that warns before the damage.

Bias, measured once, is an after-the-fact finding. The tracking signal is active surveillance: it runs every period and warns the moment a drift begins, before it costs dearly in tied-up stock or stockouts. It is the whole difference between an annual check-up and a sensor watching the vital signs continuously.

Inherited from the founding work on exponential smoothing, it is one of the few truly reliable forecast alarms: most alert systems cry wolf so often that they end up switched off. The tracking signal fires only when errors genuinely and durably pile up on the same side. Still too rarely deployed in practice, it deserves a place on any planning dashboard worthy of the name.

Business impact

The real value of the tracking signal is to turn an impossible human watch into an automatism. No planner can hand-inspect the drift of several thousand references every week. The tracking signal does it for them and surfaces only the cases that cross the threshold.

It is a principle of management by exception: instead of watching everything, you watch what steps out of line. Well calibrated, it focuses planners' scarce attention on the 2 or 3% of references that genuinely drift, and frees their time from the rest. That is where a demand team's productivity is won.

The mechanism

A cumulative that drifts, a threshold that alerts.

While the forecast oscillates around actuals, positive and negative errors offset and the cumulative stays near zero. When they fall always on the same side, the cumulative drifts inexorably and eventually crosses the alert threshold.

On the chart, each bar is the tracking signal value at one period. At first it stays quietly in the neutral zone around zero: the forecast is healthy, errors offset. Then the bars stretch gradually downward, period after period: an over-forecasting bias is silently setting in.

The moment the normalized cumulative crosses the -4 threshold, the bars turn red: the alarm fires. That threshold, conventionally set at plus or minus 4, is not magic but offers a compromise proven over decades of practice between detecting early and not firing on noise. The crossing is not a sanction, it is an invitation to go and see what is happening on that reference.

+4-4 periodes consecutives →
Healthy zoneThreshold crossed±4 threshold

Figure 1. The tracking signal period by period. It drifts downward as over-forecasting sets in, then crosses the -4 threshold: the alarm fires. Illustrative schematic.

TS = Σt ( Dt − Ft )MAD  with  MAD = Σt | Dt − Ft |n
TS tracking signal  ·  the numerator is the cumulative signed error (the same as bias)  ·  MAD mean absolute deviation, the average of absolute gaps over n periods. Dividing the cumulative by the dispersion normalizes the signal: whatever the volume scale, a TS between −4 and +4 indicates a healthy forecast; beyond that, errors pile up too much on one side to be chance.

The traps

Three setup errors.

The tracking signal is robust, but poorly tuned it becomes either screaming and useless, or blind and dangerous.

01

A threshold too wide or too tight

Too tight, it cries constantly and the team ends up ignoring it out of fatigue. Too wide, it lets costly biases slip under the radar. The threshold is calibrated on each portfolio's real dispersion, not on a universal value copied from a textbook.

02

Computing it at too high a grain

At the consolidated global level, opposite drifts of two families cancel in the cumulative and the alarm never sounds, even as each family drifts. The tracking signal must be set at the grain where decisions are made.

03

Watching it without an action protocol

An alarm that rings into the void is strictly useless and discredits the tool. Crossing the threshold must trigger a pre-defined action: who looks, at what precisely, and within what delay. Without that protocol, the tracking signal is just a gadget.

Computing it

Three steps, from cumulative to alarm.

The tracking signal uses exactly the same data as bias: forecasts and actuals, period by period, at the right grain.

Accumulate the signed errors

Each period, add the signed gap (actual minus forecast, with its sign) to the previous period's cumulative. This is the numerator: cumulative = Σ(D − F).

Compute the mean absolute deviation (MAD)

Average the absolute gaps over the same periods. This is the denominator, the dispersion measure that normalizes: MAD = mean( |D − F| ).

Set the tracking signal

Divide the cumulative by the MAD, each period (see formula above). Watch for the bounds being crossed: above +4 or below -4, trigger the defined investigation protocol.

Neighboring concepts

Read next.

The tracking signal is the natural extension of bias and builds on error measurement.

From knowledge to action

Are your forecasts under watch?

Installing the right alarms prevents silent drifts and frees planners' time. Our Planning, Forecasting & S&OP file sets them up with your teams.