Is drift always random?

The process is random, but a drift coefficient can be deterministic.

Can drift depend on the state?

Yes; local expected movement can vary with the current value.

Is this concept drift in AI?

No. That phrase usually refers to changes in predictive relationships or distributions, depending on the definition used.

AI Glossary / Statistics and probability / Intermediate

Stochastic drift

Definition

Stochastic drift describes the systematic tendency in the change of a random process, usually expressed through an expected increment or an instantaneous conditional expected rate of change. A process with positive drift tends to move upward on average, even though individual paths can move downward. The term concerns the probabilistic evolution of a process. It should not be confused with every visible trend, with random fluctuation itself, or with the separate machine-learning concept of data or concept drift.

What it means

Random processes can combine a predictable tendency with unpredictable variation. In a simple random walk, drift is the mean change per step. In a continuous-time diffusion model, the drift coefficient describes the local expected movement, while a diffusion coefficient describes the random component's scale.

The tendency need not be constant: it can depend on time or the current state. In a mean-reverting process, for example, local drift may point upward below a reference level and downward above it. An overall mean can remain constant even though local conditional drift is nonzero. Stating the model and conditioning information therefore makes the word more precise than simply saying that the average changes.

How it works

Define the process, time unit, and available information. In discrete time, examine E[X next − X current | current information]. For a random walk with independent, identically distributed increments of mean m, this conditional expected increment is m.

Over n steps, that simple model has expected position X initial + nm. If increments have variance v and are independent, position variance grows by nv. Drift and variability describe different features: one governs the expected displacement, while the other governs spread around it.

For Brownian motion with constant drift, a common model is X(t) = x + mt + sB(t), where B(t) is standard Brownian motion. The mean is x + mt and variance is s squared times t. More general diffusion models use state- or time-dependent coefficients.

When estimating drift from observations, account for noise, sampling frequency, dependence, and changing conditions. A slope drawn through a short realization may be a poor estimate of the underlying tendency. Model adequacy and parameter uncertainty matter alongside the fitted drift value.

Three practical examples

1. An engineer models sensor calibration offset as a gradual average change combined with random fluctuations and checks whether the assumptions fit observed behavior.

2. An analyst simulates a workload index with positive expected increments while preserving the possibility of temporary declines.

3. An AI researcher distinguishes a drift coefficient in a stochastic process model from a change in the distribution of incoming model inputs.

Think of it this way

Think of a person walking on a moving walkway while making random steps forward and backward. The walkway supplies a directional tendency, but the person's next observed movement can oppose it. Looking at one short stretch does not necessarily reveal the walkway's speed.

Worked example

For an original discrete teaching example, let X start at zero. Each independent step is +1 with probability 0.6 and −1 with probability 0.4. The mean increment is 0.6 × 1 + 0.4 × (−1) = 0.2. Its second moment is 1, so its variance is 1 − 0.2 squared = 0.96.

After 25 steps, expected position is 25 × 0.2 = 5. Position variance is 25 × 0.96 = 24, giving standard deviation sqrt(24), about 4.899. The positive drift does not imply every trajectory ends at 5, stays positive, or rises at every step. Even the first step is negative with probability 0.4.

For comparison, consider X(t) = 10 + 0.5t + 2B(t). At t = 4 its mean is 12 and variance is 4 × 4 = 16, so its standard deviation is 4. The drift coefficient 0.5 and diffusion scale 2 have different units relative to time. These are stipulated mathematical examples, not fitted forecasts or evidence that a particular business process follows either model.

Illustrative case study

In a fictional engineering company, Pine Instruments sees an upward pattern in a sensor's offset. An analyst initially calls every positive weekly change drift. The team instead specifies a model separating average movement from random variation, then examines repeated measurements and environmental conditions. Some short downward movements remain compatible with positive estimated drift. The report includes uncertainty and compares alternative models, helping maintenance staff avoid interpreting a single temporary decrease as proof that the underlying tendency has disappeared.

Common misconceptions and mistakes

Positive drift does not guarantee a positive outcome at a chosen time. Random variability can overwhelm the expected movement over a short interval.

Zero drift does not mean a process stays fixed. A random walk can wander widely while having zero expected increment.

A visible upward run does not establish nonzero drift, and a drift coefficient need not be constant. In addition, machine-learning data drift usually concerns changing distributions, which is a different use of the word. A process can have a constant drift coefficient without that coefficient itself changing over time.

Business use

Drift models can support simulation, monitoring, and interpretation of noisy measurements. They make the distinction between expected movement and uncertainty explicit, which helps teams compare scenarios and set investigation priorities.

Before using a fitted drift operationally, define its units and horizon, examine whether the process has boundaries or mean reversion, and check whether increments behave as assumed. A model that permits negative values may be unsuitable for an inherently nonnegative quantity without modification. Do not interpret a convenient constant-drift model as a guaranteed long-term trajectory.

Check your understanding

Answer: A random walk moves +2 or −2 with equal probability each step.

What is its drift per step?

Answer: Zero, because the expected increment is zero. Its nonzero increment variance means the path can still move substantially.

Related terms

Learn next

Study random-walk moments and Brownian motion. Then compare constant drift with the state-dependent drift of a mean-reverting process.

Random walk

Sources

Pitman and Yor, A guide to Brownian motion and related stochastic processes: https://www.stat.berkeley.edu/~aldous/205B/pitman_yor_guide_bm.pdf . The author text defines Brownian motion with drift and variance parameters. The entry uses the probability-process meaning, explicitly separating AI distribution drift. Both numerical examples and the fictional engineering case are original.

Previous: Stochastic dominance Next: Stochastic edit