Time Series in Practice: The Stationarity Assumption and When It Fails

Article 4 of 6 | Infrastructure of Risk Series

Photo by Umberto on Unsplash

The VaR estimate didn’t spike. That was the problem. For thirty days before the loss event, the model had been reporting orderly numbers, numbers that reflected a volatility surface that no longer existed in any meaningful sense. When the correlation shift hit, it didn’t signal a regime change. It seemed like just a bad week. The post-mortem later used the phrase “unprecedented market conditions.” However, it failed to address the model’s core assumption: that the statistical properties of the training data would hold in unseen conditions.

This assumption has a name: stationarity. It’s a costly oversight in quantitative risk management.

The Assumption Nobody Tests

A stationary time series boasts three key traits: a steady mean, unchanging variance, and autocorrelation that hinges solely on the distance between observations. However, financial time series frequently stray from these ideals. This isn’t merely an occasional hiccup; research reveals it’s a prevalent pattern.

Stationarity is a working assumption, but practitioners don’t fully believe in it. Indeed, many quants recognize it is not true. It persists because testing for stationarity is inconvenient. Models also seem to work well when times are calm, and their failures are often blamed on market conditions.

The difference between weak and strict stationarity is important. Weak stationarity requires only stable mean and variance over time. Strict stationarity demands the entire distribution to be time-invariant. Most risk models assume at least weak stationarity, but most failures involve shifts in the full distribution. This gap is where capital is lost.

Non-stationarity in financial series comes in four forms, and they are not interchangeable. Trending means indicate the unit root problem. Abrupt shifts show structural breaks. Volatility that changes with market conditions is heteroskedasticity. And autocorrelation that decays slowly is long memory. Each type affects models differently and requires different responses. None of these are rare; all appear regularly in equity returns, credit spreads, interest rates, and FX series. Ignoring them isn’t simplifying; it’s neglecting the data.

Unit Roots and the Cointegration Trap

Prices aren’t fixed; they change over time. If a price series has a unit root, it doesn’t have a stable average to fall back on. This means any disruptions have a lasting effect, making it tough to forecast. To handle this, analysts usually transform the series. This is accomplished by differencing or using log returns instead of actual prices. Another approach is to look at changes in yield rather than the yield itself.

Differencing is often seen as the end of the analysis, but differenced series can still have problems. They can show non-stationarity in higher moments, have structural breaks, or have unstable relationships, even if each series looks stationary on its own.

This is where cointegration comes in. Two series are cointegrated if a combination of them is stationary, meaning they have a stable long-run relationship that persists even as each series changes. Models like relative-value trading, spread models, and pairs strategies rely on these relationships. The Engle-Granger and Johansen procedures are used to test for cointegration.

But these tests can’t predict if the relationship will hold during a liquidity crisis.

The European sovereign debt market between 2005 and 2010 illustrates the risks of structural breaks. Many models used to calculate relative value relied on the connection between interest rates in peripheral and core countries. This gap got smaller and smaller in the mid-2000s as it seemed like the European Monetary Union was going to continue to integrate. These models did their job when tested with the data they had. But during the subsequent crisis, the relationship between the rates didn’t just change — it vanished. Models that had flagged small differences as trading opportunities were suddenly hit with huge gaps of hundreds of basis points, and the models failed to account for the regime change. They didn’t send out any signals to trade, resulting in significant risk exposure.

A cointegration test that passes in calm markets only shows that a relationship existed at that time. It does not guarantee that the relationship will persist or maintain its predictive power.

Structural Breaks: The Outlier That Isn’t

The most operationally dangerous form of non-stationarity is not the one that looks dangerous. A structural break — a permanent shift in the data-generating process — does not announce itself as a regime change in real time. It presents as a bad observation: an extreme return, a vol spike, a correlation reading that sits three standard deviations from the historical mean. The model flags it, the risk committee discusses it, and the observation is absorbed into the historical window. If the lookback period is long enough, it is eventually diluted. If it is short enough, it ages out entirely.

This is precisely the mechanism behind the 2008 underestimation problem. Historical simulation VaR models using pre-crisis calibration windows were not obviously misspecified. They were trained on a regime that had produced orderly, if occasionally volatile, return distributions. The crisis didn’t reveal that the models were wrong about the old regime. It revealed that the old regime was gone.

The Chow test checks for structural breaks at a specific date by seeing if regression parameters stay the same across two periods. CUSUM-based methods track cumulative forecast errors and flag instability when they go past a certain limit. Both methods have the same basic limitation: they look back at past data. By the time they show clear evidence of a break, the risky event has already occurred.

The Markov-switching model addresses this by assigning observations to different regimes and updating these assignments as new data arrives. However, this approach introduces specific challenges. The estimated regime probabilities can vary greatly based on the model setup. A poorly configured model might falsely signal regime changes due to noise or miss real ones.

In truth, there isn’t one best test to find. The trade-off exists between being sensitive to breaks and maintaining stability. Breaks need to be found early while avoiding false alarms. This issue can’t be solved with just better techniques; it requires governance. Model documents should clearly outline what triggers a human review of regime assumptions. They should specify who makes that decision and what actions they can take. Most model documents fail to mention this, which is a risk management failure.

GARCH: The Non-Stationarity Risk Managers Acknowledge

Time-varying volatility is widely accepted in the industry. The Generalized Autoregressive Conditional Heteroskedasticity model, or GARCH, backs this view. The GARCH(1,1) model is a popular tool in risk management. The mechanics are simple. Today’s conditional variance combines yesterday’s variance and yesterday’s squared return. The alpha parameter shows how quickly the model reacts to shocks. A high alpha means estimates change rapidly after big moves. Beta indicates persistence — how slowly the shock fades. If alpha plus beta approaches one, volatility shocks linger. This means a spike today can affect risk estimates for weeks.

GARCH: A Primer

Volatility is not constant. Anyone who has watched equity markets through a stress period knows this intuitively. The Generalized Autoregressive Conditional Heteroskedasticity model — GARCH — gives that intuition a formal structure.

The core idea: variance tomorrow is a function of variance today and what actually happened today. A large return, positive or negative, raises the estimate of tomorrow’s variance. A quiet day lowers it. The model is conditional because each variance estimate is conditioned on the information available at the time.

The GARCH(1,1) specification is the simplest practical form. The two parameters in parentheses indicate one lag of past squared returns and one lag of past variance. Today’s conditional variance equals a weighted sum of three terms: a long-run average variance, yesterday’s squared return scaled by alpha, and yesterday’s conditional variance scaled by beta.

Alpha measures reactivity. High alpha means the model responds quickly to new shocks. Beta measures persistence. High beta means shocks decay slowly. When alpha plus beta approaches one, a volatility spike today will still be visible in the model’s estimates weeks from now.

Most financial volatility series produce alpha plus beta estimates close to one. That is not a mathematical curiosity. It is a description of how markets actually behave.

GARCH addresses a real issue. It lets conditional variance adjust to market conditions. This means VaR estimates from GARCH-filtered returns are not tied to a single volatility estimate that overlooks past events. This is a big step up from basic historical simulations.

However, GARCH does not solve everything. It ignores mean non-stationarity and does not spot structural breaks. Most critically, as seen in 2008 and March 2020, it fails to model correlations between series. A GARCH model for individual assets may provide solid volatility estimates during a crisis, but portfolio risk can be misestimated. While individual volatilities rise, correlations tend to converge toward one, reducing diversification. GARCH captures the first issue but misses the second.

The EGARCH and GJR-GARCH extensions fill one more gap: the leverage effect. Negative equity returns increase implied volatility more than positive ones. This asymmetry has been noted across equity markets since the early 1990s. Using a symmetric GARCH model for equity risk is a mistake, leading to underestimated downside risk.

Long Memory and the Problem With Forgetting Slowly

Autocorrelations of returns decay rapidly. When you look at daily equity returns at short lags, the signal vanishes. This is why the random walk theory of prices still holds for returns.

Autocorrelations in squared returns act differently. They decline slowly, so shocks to volatility can affect variance estimates for weeks or months. This long-memory property of financial volatility goes against the GARCH(1,1) model’s assumption of geometric decay. The FIGARCH model — Fractionally Integrated GARCH — addresses this by allowing a fractional decay rate. This reflects the hyperbolic decay seen in real volatility series.

For those not using FIGARCH models, the message is clear: rolling-window estimators that give equal weight to all observations are biased. A 30-day window that includes a volatility spike from the first week treats that spike as equally important as an observation from the previous day. Over time, the spike’s influence fades to zero. This results in risk estimates that are noisy and consistently too low after stressful events, even when the actual risk is still high.

Exponentially Weighted Moving Average methods, or EWMA, provide some improvement. The RiskMetrics decay factor of 0.94 gives more weight to recent observations, reducing the sudden change at the window boundary. Although this is better than equal weighting, it doesn’t fully address the long-memory issue, as the exponential decay still underestimates the lasting effects of a shock.

As a result, firms using EWMA or short rolling windows after stressful events will see their risk metrics decrease faster than the actual risk. This difference is key when trading desks adjust positions. As risk appetite grows, accurate volatility assessments become vital for capital allocation.

What Honest Time-Series Risk Modeling Requires

This is not about making things complicated. Having more parameters doesn’t mean a model is better; in fact, it can be less reliable when things are uncertain. The point is to be honest: clearly state your assumptions, why they’re convenient, and what happens if they don’t work out.

There are four key practices that set good time-series risk models apart from average ones.

1. Clearly labeling regimes: A model based on data from 2013 to 2019 is specific to that time period, which had its own unique characteristics. This limitation should be clear to model validators and the risk committee. If the regime changes, the model owner must trigger a formal review.
2. Regular stationarity testing: Before recalibrating a model, run tests like the Augmented Dickey-Fuller (ADF) and Kwiatkowski-Phillips-Schmidt-Shin (KPSS) on the inputs. The ADF test checks for a unit root to identify non-stationarity, while the KPSS test evaluates the null hypothesis of stationarity. Using both gives a more complete picture. Ignoring a failed test is a serious governance issue, regardless of its impact on the model’s results.
3. Reporting conditional and unconditional risks: Risk reports that show just one loss estimate (VaR) miss key assumptions about the current situation. Showing both types of estimates highlights significant differences. When these estimates are far apart, it means the current period is different from the historical average.
4. Protocols for switching models: If stationarity tests fail or the model’s structure changes, what’s the next step? Most frameworks lack a clear plan for such situations. A key part of a robust framework is having a governance trigger that alerts senior managers, enabling them to adjust capital buffers or positions as needed. Without a plan, the default is to do nothing, which is rarely the best option.

The Forecast That Isn’t

Time-series risk models do not predict the future. They look at the past and use it to make projections, assuming things will stay the same. When this assumption holds, the projections are useful. But when it breaks, the model fails. It underestimates current risks by relying on a calmer past.

The assumption that things will stay the same is called stationarity. This assumption fails when the underlying conditions change. Markets change during times of stress, liquidity crises, rate shifts, and economic disruptions. These events are common features that risk models need to handle.

The key question for model validators and risk officers is not how complex the models are. It’s whether the people in charge of the models treat the possibility of change as a real risk or just a theoretical concern that someone else will deal with. For many institutions, the answer is not what they want to hear. This can be a wake-up call that leads to action, unlike a failed test of stationarity, which often gets ignored.

This is the fourth of six articles that help define the infrastructure of risk. Next: Model risk validation frameworks examined; what a rigorous validation process actually tests for, and why the most consequential model failures are rarely failures of mathematics.