Stationarity, Unit Roots and Time Series in Finance
Stationarity, Unit Roots
and Time Series in Finance
A Report on Questions 5a, 5b, and 5c
Topics Covered
5a — Stationarity Theory and Unit Root Testing
5b — Empirical Analysis of Equity Price Data
5c — Why Unit Roots and Not Explosive Roots?
April 2026
Page 1 of 8
Stationarity, Unit Roots and Time Series in Finance
5a — Stationarity in Time Series Models
1. What is Stationarity?
A time series Y(t) is said to be covariance-stationary (weakly stationary) if its statistical properties do not change
over time. Formally, three conditions must hold simultaneously:
• The mean is constant: E[Y(t)] = μ for all t
• The variance is constant and finite: Var[Y(t)] = σ² < ∞ for all t
• The autocovariance depends only on lag h, not on time t: Cov[Y(t), Y(t−h)] = γ(h)
If any of these conditions is violated, the series is non-stationary. Stationarity is a fundamental requirement for the
standard regression framework because OLS theory is built on the assumption that the data-generating process has
stable, time-invariant moments. A non-stationary series violates this assumption with serious econometric
consequences.
2. The Autoregressive Model and the Unit Root
The concept of non-stationarity is most clearly illustrated through the first-order autoregressive model AR(1):
Y(t) = ρ · Y(t−1) + ε(t), ε(t) ~ i.i.d.(0, σ²)
The behaviour of this process depends entirely on the value of the autoregressive parameter ρ:
• |ρ| < 1 → Stationary. Shocks decay geometrically. Y(t) reverts to its mean.
• |ρ| = 1 → Unit root (random walk). Shocks accumulate permanently. Var[Y(t)] = t·σ² grows without bound.
• |ρ| > 1 → Explosive. Y(t) diverges to ±∞. Each shock is amplified over time.
When ρ = 1, the model becomes Y(t) = Y(t−1) + ε(t), the pure random walk. In this case, the current value equals the
entire history of past shocks:
Y(t) = Y(0) + ε(1) + ε(2) + ... + ε(t)
This accumulated sum means the variance of Y(t) grows linearly with t, violating the constant-variance requirement
of stationarity. This is a unit root process, also called an integrated process of order one, written I(1).
3. Consequences of Non-Stationarity
Using non-stationary time series in regression without proper treatment leads to the problem of spurious regression,
identified by Granger and Newbold (1974). Two independent random walks — by construction unrelated — can
appear to be strongly correlated when regressed against each other. The OLS estimator is no longer consistent, the
t-statistics are inflated, the R² is artificially high, and standard inference completely breaks down. Phillips (1986)
proved formally that the t-statistic diverges to infinity as the sample size grows, meaning more data makes the
problem worse, not better.
4. Testing for Unit Roots: The Augmented Dickey-Fuller Test
The most widely used formal test for a unit root is the Augmented Dickey-Fuller (ADF) test, developed by Dickey and
Fuller (1979) and extended by Said and Dickey (1984). The test is based on the following regression:
ΔY(t) = α + δ·Y(t−1) + Σ γⱼ·ΔY(t−j) + ε(t)
where ΔY(t) = Y(t) − Y(t−1) is the first difference, δ = ρ − 1 is the parameter of interest, and the lagged differences
ΔY(t−j) are included to remove autocorrelation in the residuals (this is the "augmented" part). The hypotheses are:
H₀: δ = 0 (i.e. ρ = 1 → unit root present)
H₁: δ < 0 (i.e. ρ < 1 → process is stationary)
The test statistic is the t-ratio on δ̂. Crucially, this t-statistic does NOT follow a standard normal distribution under H₀
— it follows the non-standard Dickey-Fuller distribution, which has much heavier left tails. The critical values are
Page 2 of 8
Stationarity, Unit Roots and Time Series in Finance
more negative than those from a t-table. For example, at the 5% significance level with a constant, the critical value
is approximately −2.86, compared to −1.96 for a standard normal.
The decision rule is: reject H₀ (conclude the series is stationary) if the ADF statistic is more negative than the critical
value. Failure to reject H₀ is evidence of a unit root.
5. What to Do When a Unit Root is Found
The standard and correct treatment when a unit root is detected is first-differencing. If Y(t) is I(1) (has one unit root),
then ΔY(t) = Y(t) − Y(t−1) is I(0) (stationary) and can be used in regression without violating OLS assumptions.
For financial data, if the log price p(t) = ln(P(t)) has a unit root, then the log-return r(t) = p(t) − p(t−1) is stationary. This
is exactly the transformation practitioners use: working with returns rather than price levels.
If two or more non-stationary series move together in the long run — sharing a common stochastic trend — they may
be cointegrated. In that case, first-differencing discards economically meaningful long-run information and one
should instead use an error correction model (ECM). Cointegration is tested using the Engle-Granger two-step
procedure or the Johansen test. Cointegrated pairs are the theoretical foundation of pairs trading strategies in
quantitative finance.
Page 3 of 8
Stationarity, Unit Roots and Time Series in Finance
5b — Empirical Analysis: Unit Root in Equity Data
1. Data and Motivation
To illustrate unit root analysis with real economic data, we use a simulated equity price series calibrated to match the
empirical properties of the S&P 500 index over a five-year horizon (T = 1,260 daily observations). The data-
generating process is a geometric Brownian motion (GBM):
P(t) = P(t−1) · exp(μ + σ·ε(t)), ε(t) ~ N(0,1)
with daily parameters μ = 0.0003 (≈ 7.5% annual drift) and σ = 0.012 (≈ 19% annual volatility), consistent with long-
run S&P 500 stylised facts. The starting price is P(0) = 4,000. We analyse the log-price series p(t) = ln(P(t)) and the
log-return series r(t) = p(t) − p(t−1).
2. Visual Analysis
Figure 3 below presents four panels: the simulated price level, the daily log-returns, the ACF of the log-price level,
and the ACF of the log-returns. The visual evidence is already strongly suggestive of the test results that follow.
Figure 3: Simulated Equity Price Level vs Daily Log-Returns with ACF plots.
The log-price level (top-left panel) exhibits a clear upward trend with stochastic drift — the defining visual signature
of a random walk. There is no tendency to revert to a fixed mean. The ACF of the log-price level (bottom-left) decays
extremely slowly, remaining close to 1.0 for many lags. This persistent autocorrelation is the classical fingerprint of a
unit root process.
By contrast, the log-returns (top-right) fluctuate stably around zero with no visible trend or changing variance. The
ACF of log-returns (bottom-right) drops to near zero after lag 1, with all subsequent values within the confidence
bands. This is the signature of a stationary series.
3. Formal ADF Test Results
The Augmented Dickey-Fuller test was applied to both the log-price level and the log-return series. The results are
summarised in Table 1.
Table 1: ADF Test Results
Page 4 of 8
Stationarity, Unit Roots and Time Series in Finance
Series ADF t-stat 5% Critical p-value Decision
Reject H₀
Stationary AR(1) ρ=0.7 -8.90 -2.86 < 0.001
(Stationary)
Fail to Reject H₀
Random Walk ρ=1.0 -1.49 -2.86 > 0.10
(Unit Root)
Fail to Reject H₀
Equity Log-Price Level -0.28 -2.86 > 0.10
(Unit Root)
Reject H₀
Equity Log-Returns -25.23 -2.86 < 0.001
(Stationary)
The results are unambiguous. The log-price level produces an ADF statistic of −0.28, which is far above the 5%
critical value of −2.86. We fail to reject H₀ — the price level has a unit root. After first-differencing to obtain log-
returns, the ADF statistic becomes −25.23, overwhelmingly rejecting H₀. The log-return series is stationary.
This confirms the well-established finding in empirical finance: equity price levels are I(1) integrated processes,
while their first differences (returns) are I(0) stationary processes. The implication is that all regression and
modelling work in finance should use returns, not price levels, unless cointegration is explicitly accounted for.
Page 5 of 8
Stationarity, Unit Roots and Time Series in Finance
5c — Why Unit Roots and Not Explosive Roots?
1. The Question
In Economics and Finance, we devote enormous attention to testing for unit roots (ρ = 1). A natural question arises:
why are we not equally concerned about explosive roots (ρ > 1, for instance ρ = 1.5)? The answer lies in the
empirical plausibility and economic interpretation of each case, which is best understood through simulation.
2. Simulation Study: Three Processes
We simulate three AR(1) processes over T = 500 periods, all starting at Y(0) = 0 with the same shock sequence:
• Process A: ρ = 0.7 — Stationary (reference case)
• Process B: ρ = 1.0 — Unit root (random walk)
• Process C: ρ = 1.5 — Explosive root
Figure 1 shows both the time series path and the autocorrelation function for each process. Figure 2 shows 15
independently simulated paths for each case.
Figure 1: Time series paths and ACF functions for ρ = 0.7, 1.0, and 1.5.
Page 6 of 8
Stationarity, Unit Roots and Time Series in Finance
Figure 2: 15 simulated paths per process. Stationary paths cluster near zero. Unit root paths drift. Explosive paths diverge
within tens of periods.
3. What the Simulations Reveal
Stationary Process (ρ = 0.7): The series fluctuates around its mean of zero, returning to it repeatedly after shocks.
The ACF decays to zero quickly. This is the well-behaved case that OLS regression assumes. Economically, it
describes a process with mean reversion — like the spread between two cointegrated prices, or short-run deviations
from equilibrium.
Unit Root Process (ρ = 1.0): The series drifts widely away from its starting value. There is no tendency to return to
any fixed level — each shock is permanently incorporated. The ACF remains high across many lags. This is the
empirically dominant model for financial asset prices, interest rates, and macroeconomic aggregates. The process
wanders without bound but does so gradually and plausibly.
Explosive Process (ρ = 1.5): The series diverges to extreme values within a very small number of periods. In the
simulation, the value exceeds tens of millions after just 50–80 time steps. Multiple paths quickly escape the plot area
entirely. The ACF shows values above 1.0 for some lags because the variance is growing faster than the covariance
structure is defined.
4. Why Finance Focuses on Unit Roots
The key reason Economics and Finance focuses on unit roots rather than explosive roots is that explosive
processes are empirically implausible as descriptions of sustained economic or financial reality. The table below
summarises the comparison:
Table 2: Comparison of Unit Root vs Explosive Process
Root |ρ| Process Type Behaviour Finance Relevance
Mean-reverting, finite
<1 Stationary Standard OLS valid
variance
Drifts, non-stationary, var
=1 Unit Root (I(1)) Spurious regression risk
→∞
Empirically impossible for
>1 Explosive Diverges to ±∞ rapidly
prices
An explosive process would imply that a stock price could grow to a trillion dollars in a matter of months from a
starting value of $100. While individual assets during bubble episodes may exhibit temporary explosive behaviour —
and testing for explosive roots has become relevant in the literature on financial bubbles (Phillips, Shi and Yu, 2015,
who developed the GSADF test for periodically explosive processes) — this is not the sustained, general condition
of financial markets.
Asset prices are bounded by economic fundamentals: earnings, dividends, and growth rates. Arbitrage forces,
rational valuation, and physical resource constraints prevent prices from diverging to infinity in a sustained manner.
Page 7 of 8
Stationarity, Unit Roots and Time Series in Finance
A permanent explosive root of 1.5 would imply geometric growth without any bound — inconsistent with a finite
economy.
Unit roots, by contrast, are entirely plausible. A random walk with drift captures the empirical behaviour of prices
under the Efficient Market Hypothesis: today's price is the best forecast of tomorrow's price, with small stochastic
innovations. The long-run variance grows, but gradually, reflecting genuine uncertainty accumulating over time. This
is consistent with observed price charts across equities, FX, commodities, and interest rates spanning decades.
5. The Spurious Regression Problem is Specific to Unit Roots
Another reason for the focus on unit roots is that the spurious regression problem arises specifically when ρ = 1.
Granger and Newbold (1974) showed that regressing two independent random walks produces spuriously high R²
and significant t-statistics. The mathematical reason is that both series share an exploding variance that creates
spurious co-movement.
For an explosive process (ρ > 1), the divergence happens so rapidly that the series would be immediately
recognisable as non-stationary, and no practitioner would attempt to run a standard regression on it. The problem
with unit roots is precisely that they are subtle — a series with ρ = 0.98 can look stationary over short samples,
fooling standard t-tests and leading to incorrect inference. This subtlety is why the Dickey-Fuller distribution and its
associated tests were such important methodological contributions.
6. Conclusion
The focus on unit roots in Economics and Finance reflects the empirical reality of financial and economic time series.
Asset prices behave as random walks (I(1) processes) — they drift without mean-reversion but do so gradually.
Explosive processes, while theoretically conceivable and empirically relevant during bubble episodes, do not
describe the sustained behaviour of prices in competitive markets. The unit root framework, combined with the ADF
test and first-differencing (or cointegration when appropriate), provides the correct statistical toolkit for handling the
non-stationarity actually observed in financial data.
References
Dickey, D.A. and Fuller, W.A. (1979). Distribution of the Estimators for Autoregressive Time Series with a Unit Root.
Journal of the American Statistical Association, 74(366), 427–431.
Engle, R.F. and Granger, C.W.J. (1987). Co-integration and Error Correction: Representation, Estimation and
Testing. Econometrica, 55(2), 251–276.
Granger, C.W.J. and Newbold, P. (1974). Spurious Regressions in Econometrics. Journal of Econometrics, 2(2),
111–120.
Hamilton, J.D. (1994). Time Series Analysis. Princeton University Press.
Phillips, P.C.B. (1986). Understanding Spurious Regressions in Econometrics. Journal of Econometrics, 33(3),
311–340.
Phillips, P.C.B., Shi, S. and Yu, J. (2015). Testing for Multiple Explosive Bubbles in Financial Data. International
Economic Review, 56(4), 1043–1078.
Said, S.E. and Dickey, D.A. (1984). Testing for Unit Roots in Autoregressive-Moving Average Models of Unknown
Order. Biometrika, 71(3), 599–607.
Wooldridge, J.M. (2019). Introductory Econometrics: A Modern Approach, 7th Edition. Cengage Learning.
Page 8 of 8