Does I(0) mean no correlation over time?

No. A stationary process can have serial dependence.

Can an I(1) series have drift?

Yes. Its first differences can have a stable nonzero mean.

Can five observations establish the order?

Not reliably. The example demonstrates arithmetic, while identifying a process requires much more evidence and assumptions.

AI Glossary / Statistics and probability / Intermediate

Order of integration

Definition

The order of integration describes how many ordinary differences are needed to remove a time series' stochastic trend and obtain an integrated-of-order-zero process. In the usual integer-order setting, a series is written I(d) when its dth difference is I(0), while lower differences retain the relevant nonstationarity. Deterministic trends and seasonal integration require separate attention; not every changing series is a random walk.

What it means

A level series may wander persistently while its changes fluctuate around a stable mean. Such behavior is consistent with an I(1) process. If changes themselves contain a stochastic trend but changes in changes are I(0), the level process is I(2).
Integration here is the inverse of differencing: accumulating a stationary sequence can produce an integrated process. It is unrelated to the order of numerical integration in calculus software or to combining datasets in a business system.

How it works

Write the first difference as delta y_t = y_t - y_(t-1). Applying the operation twice gives delta-squared y_t = y_t - 2y_(t-1) + y_(t-2). Ordinary differencing always compares adjacent observations; a seasonal difference compares observations separated by the seasonal period.
In common short-memory time-series models, I(0) means stationary behavior with finite, stable second moments and appropriate weak dependence, after accounting for specified deterministic terms. Technical definitions can impose additional conditions, so stating only that a plot looks flat is insufficient.
A random walk y_t = y_(t-1) + e_t with nondegenerate independent, identically distributed shocks of zero mean and finite variance is a standard I(1) example. Its first difference is e_t, while the level variance grows with time from a fixed initial value. A random walk with drift also has stationary differences, now with a nonzero mean.
Distinguish a stochastic trend from a deterministic trend plus stationary noise. Removing a specified deterministic trend can produce a stationary residual without the same permanent-shock interpretation as a unit-root process. Structural breaks can further complicate diagnosis.
Use subject knowledge, plots, autocorrelation, and appropriately specified tests. Unit-root tests and stationarity tests have different null hypotheses. Their conclusions depend on trend terms, lag choices, sample size, and breaks; failure to reject is not proof of a particular order.
In ARIMA notation, d records ordinary differencing, while seasonal differencing has a separate order D. Choose differences judiciously: unnecessary differencing can introduce awkward dependence and discard useful level information. Fractional integration extends the concept beyond integer d.

Three practical examples

Forecasting: assess whether demand levels or changes should enter a time-series model.
Operational monitoring: distinguish a persistent evolving level from temporary fluctuations around a stable baseline.
Multivariate analysis: check integration properties before interpreting relationships among trending series or investigating cointegration.

Think of it this way

Think of a running total of daily additions. The additions may have stable behavior even though the total keeps wandering. Looking at changes reveals the additions again. But a total that follows a known scheduled trend is a different situation from one accumulating unpredictable shocks.

Worked example

Take the original illustrative sequence 10, 12, 11, 14, 14. Its first differences are 2, -1, 3, and 0. Its second differences are -3, 4, and -3. These calculations show the operation but cannot establish integration order from only five values.
Now specify a theoretical random walk starting at a fixed value, with independent shocks of mean zero and variance 4. After t steps, its level equals the starting value plus the sum of t shocks. Independence gives level variance 4t: it is 4 after one step, 20 after five, and 40 after ten.
The first difference is exactly one shock, whose variance remains 4 at every step. Thus the stipulated level process is nonstationary, while its first differences are I(0), making it I(1).
The second difference would be the current shock minus the previous shock. It remains stationary but adds unnecessary moving-average structure. The order is the minimum needed, not any number of differences that happens to yield stationary behavior.

Illustrative case study

Fictional case: Harbor Forecasting builds an AI-assisted workload forecast from a slowly changing service-volume series. Analysts compare a deterministic-trend explanation with a model of accumulated shocks.
They inspect changes, test specifications, and known changes in data collection, then evaluate forecasts on later periods. The report records the chosen differencing order and uncertainty about the underlying process. A short visual trend is not presented as conclusive evidence of a unit root.

Common misconceptions and mistakes

A rising series is not automatically I(1), and a flat-looking finite sample is not proof of I(0). Changes in variance, structural breaks, and deterministic seasonality can create nonstationarity that ordinary differencing does not adequately resolve.
Do not confuse autoregressive order p with integration order d. Nor should seasonal order D be silently added to ordinary d without describing the operators.
Differencing a series does not guarantee independent observations; stationary differences can remain autocorrelated. Also, applying more differences is not inherently safer. Overdifferencing can make modeling and interpretation harder.
Two I(1) series can have a stationary linear combination, so differencing everything without considering cointegration can lose information about their long-run relationship.

Business use

Businesses need this concept when forecasting persistent time series or analyzing relationships between evolving indicators. The choice affects whether a model predicts levels, changes, or changes in changes, and how forecasts return to original units. Document transformations and retain the necessary starting values. Evaluate forecast usefulness rather than treating an integration-order label as an end in itself.

Check your understanding

Why is the theoretical random walk nonstationary?

Answer: Its level variance grows with elapsed time.

Why is it not called I(2) after taking two differences?

Answer: One difference already produces I(0); integration order records the minimum required in this setting.

Related terms

Stationarity
Differencing
Random walk
Cointegration
unit root
ARIMA
deterministic trend
seasonal differencing
fractional integration

Learn next

compare a random walk with a deterministic linear trend plus stationary noise, focusing on how each responds to a one-time shock.

Random walk

Sources

Sources: Hyndman and Athanasopoulos, Forecasting: Principles and Practice, stationarity, differencing, and ARIMA: https://otexts.com/fpp3/arima.html . Relevant primary textbook sections were opened and checked. The finite sequence and theoretical variance calculation are original; the sequence is not used as evidence of an estimated integration order.

Previous: Order of accuracy Next: Order selection criteria