Notes
🔍

Unit Root Process

A lot of time-series data “remembers” its past: today’s value is strongly tied to yesterday’s. A unit root process is an extreme version of that memory, where shocks don’t fade away—they stick around.

What it means technically
A common way to describe it is with an AR(1) model: yt = ρyt−1 + εt. When ρ = 1, the series has a unit root and becomes a random walk: yt = yt−1 + εt. Here, εt is a shock (often treated as white noise). The key consequence is non-stationarity: the mean and variance aren’t stable over time, and the effect of a shock is permanent. If you add a constant drift term, it becomes a random walk with drift, which can look like a trend even when it’s just accumulated noise.

Intuition
Think of a unit root series like a person taking steps in random directions. Each step changes where they are, and there’s no force pulling them back to a typical location. That “no pull-back” is what breaks stationarity.

Real-world examples

  • Stock prices are often modeled as close to a random walk; returns may be stationary, prices often are not.
  • House price indices can behave like they have a unit root, especially over long periods.
  • Macroeconomic levels (GDP, price level) are frequently treated as unit-root-like, while their growth rates are closer to stationary.

Why it matters in AI/ML
Many ML models quietly assume stable distributions. With a unit root, training/validation splits can mislead, and regressions can become spurious (high R² with no real relationship). Common fixes include differencing (modeling Δyt) as in ARIMA, or using unit root tests like ADF and KPSS to decide transformations.

Unit Root Process is a time series whose autoregressive behavior includes a root at 1 (e.g., AR(1): yt=yt−1+εt), making it nonstationary: shocks have permanent effects and variance grows over time. It matters in AI/ML forecasting because many models assume stationarity; ignoring a unit root can yield spurious relationships and poor generalization. Example: stock prices often resemble a random walk and are differenced to model returns.

Imagine a drunk person taking steps: each step is random, and there’s no strong pull back to the starting point. Over time, they can wander far away, and their position doesn’t settle around a typical level. That’s the feel of a unit root process in time-series data.

In statistics and AI/ML, a unit root process is a time series where today’s value is basically yesterday’s value plus some random shock. Because shocks “stick,” the series tends to drift and its average and variability can change over time. This matters because many forecasting methods assume stable behavior, so unit roots often require differencing (looking at changes) first.