Does a constant first difference imply a straight-line level sequence?

Yes, for equally spaced observations; adding the same amount each period produces a linear trend.

Does a zero first difference mean the original value is zero?

No. It means the value did not change from the previous observation.

Can first differences be negative?

Yes. Negative values represent decreases, even when every original observation is positive.

AI Glossary / Time series and signal processing / Intermediate

First difference

Definition

A first difference is the change between a time-series observation and the immediately preceding observation: Δyt = yt − y(t−1). It replaces levels with consecutive changes. In time-series modeling, first differencing can help remove certain forms of nonstationarity, but it does not automatically make every series stationary or suitable for forecasting.

What it means

A series of account totals answers how much exists at each date. Its first differences answer how much the total changed between consecutive dates. The units remain those of the original measurement, although the interpretation becomes change per observation interval.
First difference ordinarily means a lag-one difference. It is distinct from subtracting the value one seasonal cycle earlier, and from taking a second difference of the changes. Clear notation prevents these operations from being confused, especially when software exposes both lag and order settings.

How it works

Arrange observations in their correct time order and subtract each preceding value from the next. A complete series of N observations yields N−1 first differences unless an earlier reference observation is supplied. Missing dates require an explicit policy: adjacent stored rows are not necessarily adjacent periods.
Model the differenced series only after examining its behavior and the reason for transforming it. To recover level forecasts, add predicted changes cumulatively to the last known level. Forecast uncertainty must also accumulate through the model, including correlations between future changes when present. Differencing a series changes both its interpretation and its dependence structure.

Three practical examples

An operations team examines weekly changes in an unresolved-ticket backlog to distinguish growth from the current backlog size.
A utility converts cumulative meter readings into interval consumption, after checking for meter resets and corrections.
A forecasting analyst tests whether modeling monthly changes is more appropriate than modeling a drifting level, while separately considering seasonal behavior and changes in variance.

Think of it this way

Think of comparing successive photographs of a water tank. The level in each photograph tells you how much water is present. The difference tells you how much the level changed since the previous photograph. A positive change does not reveal whether inflow increased or outflow decreased; it describes the net movement, not its cause.

Worked example

Consider an original fictional backlog series at five consecutive weekly endpoints: 240, 252, 249, 267, and 276 unresolved tickets. The four first differences are 12, −3, 18, and 9 tickets. Their sum is 36, which independently matches the last level minus the first: 276 − 240 = 36.
The average observed change is 36/4 = 9 tickets per week. If, purely for illustration, a model forecasts a change of 9 in each of the next two weeks, the level forecasts are 276 + 9 = 285 and 276 + 9 + 9 = 294. This is a stipulated drift forecast, not a claim that four observed changes establish a reliable model.
The first observed percentage change is 12/240 = 5%, whereas the second is −3/252, approximately −1.19%. Thus first differences and percentage changes are different quantities. A negative difference of −3 means the backlog fell by three, not that the backlog itself was negative. If the third observation had actually been taken two weeks after the second, its difference would span two weeks and should not be silently treated as a one-week change.

Illustrative case study

Illustrative fictional case: Pine Logistics monitors a growing queue of delayed shipments. A level chart emphasizes how large the queue has become, while a first-difference chart shows whether it is expanding or shrinking each day.
The team notices a sharp negative change after a database cleanup and labels it as a correction rather than an operational improvement. It retains both charts because a shrinking queue can still be unacceptably large. This fictional case demonstrates why differenced values need context and do not replace the original series.

Common misconceptions and mistakes

First differencing is not percentage growth, and it is not a derivative unless an appropriate limiting or numerical interpretation is supplied. For unequal time intervals, divide by elapsed time only when a rate is the intended quantity and its interpretation is justified.
Do not difference automatically until a plot looks flat. Unnecessary differencing can amplify noise and introduce avoidable dependence. A structural break or seasonal pattern may require separate treatment. Conversely, subtracting a fitted trend and first differencing are different operations, with different implications for future behavior and uncertainty.

Business use

Businesses can use first differences to monitor net changes in inventory, backlog, balances, and other quantities. The operation is transparent, making it useful for explaining why a level moved. However, a net change alone does not reveal all underlying inflows, outflows, or causal drivers.
For forecasting pipelines, document the transformation and retain the final observed level needed for reconstruction. Evaluate forecasts on the decision's original scale as well as any transformed scale. Ensure that revisions, resets, and missing periods are handled consistently before comparing performance over time.

Check your understanding

What are the first differences of 20, 25, and 23?

Answer: They are 5 and −2.

How many first differences arise from five complete observations?

Answer: Four, unless an earlier observation is also available.

Why add both forecast changes to obtain the example's second future level?

Answer: That level is two periods after the last observed level, so both intervening changes contribute.

Would a decline in backlog prove improved service speed?

Answer: No. Arrivals, cancellations, and data corrections can also affect the net change.

Related terms

Differencing
Seasonal difference
Stationarity
Random walk
Drift (tracking)
second difference
percentage change
logarithmic difference
drift
ARIMA
cumulative sum

Learn next

reconstruct a level series from its first differences and starting value. Compare ordinary and seasonal differences on a recurring pattern, then examine why a random walk and a deterministic trend can require different modeling choices.

Random walk

Sources

Sources: Glossary of Forecasting Terms, First difference and Differencing entries: https://robjhyndman.com/mwh3/FG4.pdf . Hyndman and Athanasopoulos, Forecasting: Principles and Practice, third edition, Stationarity and differencing: https://otexts.com/fpp3/arima.html . Primary glossary and textbook consulted. The backlog calculation and fictional case are original educational material.

Previous: First-choice hill climbing Next: First-hitting-time model