Threshold Cointegration: Modeling Non-Linear Relationships in Financial Time Series

Beyond Linearity: The Need for Threshold Models

Standard cointegration models, such as those developed by Engle-Granger and Johansen, are fundamentally linear. They assume that the adjustment back to the long-run equilibrium occurs at a constant speed, regardless of the size of the deviation. The Vector Error Correction Model (VECM), for instance, uses a constant adjustment coefficient, $\alpha$, to represent this process. However, this assumption may not hold in real-world financial markets. Transaction costs, market frictions, and strategic behaviors of large market participants can create a scenario where the mean-reverting force only kicks in when the deviation from equilibrium becomes sufficiently large.

Consider a pairs trade. If the spread between two cointegrated assets is very small, it might not be profitable to execute an arbitrage trade due to transaction costs (commissions, bid-ask spread). As a result, arbitrageurs will not act, and the spread may wander randomly within a certain band. Only when the spread widens beyond a certain threshold, making the potential profit large enough to overcome the costs, will arbitrageurs step in and push the spread back towards its mean. This implies that the adjustment process is non-linear and state-dependent. The system behaves differently inside the "no-arbitrage" band than it does outside of it. This is the core intuition behind threshold cointegration.

The Threshold Autoregressive (TAR) Model

The pioneering work on threshold models was done by Howell Tong. The basic model is the Threshold Autoregressive (TAR) model. For a mean-reverting spread, $Z_t$, a simple TAR model can be specified with two regimes, defined by a threshold value, c. Typically, the threshold is set symmetrically around the long-run mean (assumed to be zero).

$Z_t = \begin{cases} \rho_1 Z_{t-1} + \epsilon_{1,t} & \text{if } |Z_{t-1}| \leq c \ \rho_2 Z_{t-1} + \epsilon_{2,t} & \text{if } |Z_{t-1}| > c \end{cases}$

• Inner Regime ($|Z_{t-1}| \leq c$): This is the "no-arbitrage" band. Inside this band, the autoregressive coefficient $\rho_1$ is expected to be close to 1, implying that the process behaves like a random walk. There is no significant mean reversion.
• Outer Regime ($|Z_{t-1}| > c$): When the deviation exceeds the threshold c, the system enters this regime. Here, we expect strong mean reversion. The coefficient $\rho_2$ should be significantly less than 1 (and greater than -1 for stability). The farther away from 1 (i.e., the smaller $\rho_2$ is), the stronger the pull back towards the mean.

This simple model captures the essential non-linearity. The adjustment process is dormant for small shocks but activates for large ones.

The Momentum-Threshold Autoregressive (M-TAR) Model

Enders and Granger (1998) proposed a refinement to the TAR model, known as the Momentum-Threshold Autoregressive (M-TAR) model. It allows for asymmetric adjustments. The M-TAR model considers not just the magnitude of the deviation, but also its direction of change, $\Delta Z_{t-1}$. This can be particularly relevant in finance, where markets might react differently to positive and negative news, or where it might be easier for a spread to widen than to narrow.

The M-TAR model is specified as:

$\Delta Z_t = I_t \rho_1 Z_{t-1} + (1-I_t) \rho_2 Z_{t-1} + \epsilon_t$

where $I_t$ is an indicator function. The specification of $I_t$ depends on the model:

• TAR: $I_t = 1$ if $Z_{t-1} > c$ and $I_t = 0$ if $Z_{t-1} \leq c$. This allows for different reversion speeds when the spread is above or below a certain level.
• M-TAR: $I_t = 1$ if $\Delta Z_{t-1} > c$ and $I_t = 0$ if $\Delta Z_{t-1} \leq c$. This tests whether the speed of adjustment depends on the momentum of the spread in the previous period.

By testing the hypothesis $\rho_1 = \rho_2$, we can test for the presence of threshold effects. If the null is rejected, it implies an asymmetric, non-linear adjustment process.

Practical Implementation and Testing

Let's consider a spread $Z_t$ from a pair of assets.

1. Test for Standard Cointegration: First, run a standard ADF test on $Z_t$. If you can't reject the null of a unit root, it doesn't mean there's no relationship. It might be that the linear model is misspecified and a non-linear model is more appropriate.

2. Estimate the Threshold Model: The next step is to estimate the TAR or M-TAR model. This involves an additional challenge: the threshold value 'c' is usually unknown and needs to be estimated. A common method, proposed by Chan (1993), is to perform a grid search. We test a range of possible values for 'c' (e.g., the middle 70% of the values of the threshold variable) and choose the value of 'c' that minimizes the sum of squared residuals from the model.

3. Test for Threshold Cointegration: Once we have our estimated threshold 'c' and the coefficients $\hat{\rho}_1$ and $\hat{\rho}_2$, we need to test for cointegration. We test the null hypothesis $H_0: \rho_1 = \rho_2 = 0$ (which corresponds to no cointegration) against the alternative that at least one of them is less than zero. The F-statistic for this joint hypothesis is compared against special important values tabulated by Enders and Siklos (2001), as the distribution is non-standard.

4. Trading Strategy Implications:

If threshold cointegration is confirmed, it has direct implications for trading strategy design.

• Zone-based Trading: The model explicitly defines trading zones. No trades should be initiated when the spread is in the inner regime ($|Z_{t-1}| \leq c$). This avoids churning and trading costs when there is no statistical evidence of mean reversion.
• Dynamic Entry/Exit: The entry signal is now clear: enter a trade only when the spread crosses the threshold 'c'. The exit signal could be when the spread crosses back into the inner regime, or when it reaches the long-run mean.
• Asymmetric Strategies: If an M-TAR model shows that reversion is much stronger for positive deviations than for negative ones (or vice versa), a trader might choose to only trade in one direction. For example, if the spread reverts quickly when it's wide, but drifts slowly when it's narrow, one might only implement the short side of the strategy (shorting the spread when it's wide) and ignore the long side.

Example: The Law of One Price with Transaction Costs

Consider the price of a stock, say Royal Dutch Shell, traded on the London Stock Exchange (RDS.A) and the Amsterdam Stock Exchange (RDS.B). The law of one price suggests their prices (adjusted for currency) should be identical. However, transaction costs and other frictions create a band. Let $Z_t$ be the price differential.

A TAR model might find a threshold of, say, 5 basis points.

• If $|Z_{t-1}| \leq 0.05%$, the model finds $\hat{\rho}_1 \approx 1$. The price difference wanders randomly.
• If $|Z_{t-1}| > 0.05%$, the model finds $\hat{\rho}_2 = 0.7$. This indicates strong mean reversion. Arbitrageurs are now active.

A trading strategy would ignore any noise within the +/- 5 bps band and only place trades (e.g., buy RDS.B, sell RDS.A) when the differential exceeds this threshold, expecting a relatively quick reversion with a half-life determined by $\rho_2$.

Conclusion

Threshold cointegration provides a more realistic and nuanced model of equilibrium relationships in financial markets by incorporating the role of transaction costs and non-linear adjustments. By moving beyond the assumption of linear, constant-speed mean reversion, traders can build more robust strategies that filter out market noise and act only when the statistical evidence for reversion is strong. The TAR and M-TAR models provide a formal framework for identifying these non-linear dynamics, defining clear trading zones, and potentially uncovering profitable asymmetries in the adjustment process.