Seasonality in Mixed Causal-Noncausal Processes

This paper demonstrates that seasonal roots in mixed causal-noncausal autoregressive models can be isolated in their moving average representation without generating new joint seasonal effects, a finding that significantly impacts model selection procedures as validated through simulations and empirical applications.

The Gist

The Big Picture: Predicting the Future vs. Remembering the Past

Imagine you are trying to predict the weather.

• Standard Models (Causal): These are like looking at the clouds right now to guess if it will rain in an hour. They only look at the past to predict the future.
• The New Models (Noncausal): These are a bit magical. They assume that sometimes, the future "pulls" the present. It's like seeing a sudden drop in temperature and realizing, "Oh, a storm is coming tomorrow, so the air is cooling down today."

The paper focuses on Mixed Models, which use both the past (causal) and the future (noncausal) to understand time series data (like stock prices, virus cases, or crop prices). These models are great at spotting "bubbles"—sudden, explosive spikes in data that then crash.

The Problem: The "Seasonal" Confusion

The authors noticed something tricky. Real-world data often has seasonality (repeating patterns like winter holidays or summer sales).

• In standard models, these patterns are easy to spot.
• In these fancy "Mixed" models, there was a fear that the math might get messy. People worried that mixing the "past" and "future" parts together might create fake new seasons or distort the real ones. It was like worrying that mixing red and blue paint might accidentally create a new, invisible color that messes up your whole painting.

The Discovery: The "Lego" Breakdown

The authors used a mathematical tool called Partial Fraction Decomposition. Think of this like taking apart a complex Lego castle to see exactly which bricks make up the walls and which make up the roof.

Their Big Finding:
They proved that even in these complex Mixed models, the "seasonal bricks" (the roots that cause the repeating patterns) never mix to create new patterns.

• If you have a seasonal pattern in the "past" part, it stays in the past part.
• If you have one in the "future" part, it stays there.
• They don't combine to make a new kind of season.

The Analogy: Imagine you have a band with a drummer (the past) and a guitarist (the future). The paper proves that no matter how they play together, the drummer can't suddenly start playing a melody that sounds like a saxophone, and the guitarist can't suddenly start playing a drum beat. They keep their own distinct sounds. You can always separate them out.

Why This Matters: The "Recipe" for Better Models

This discovery is huge for statisticians and economists because it simplifies the recipe for building these models.

1. The "Pseudo-Causal" Shortcut
Instead of guessing how to split the data between the "past" and "future" parts, you can first build a simple, standard model (the "Pseudo-Causal" model). This acts like a rough draft.

• Once you have the rough draft, you can see all the "seasonal bricks" (roots) clearly.
• Because we know they don't mix, we know exactly how to sort them. If you see a pair of complex seasonal roots, you know they must stay together in either the "past" bucket or the "future" bucket. You don't have to try every possible combination.

2. Avoiding the "Local Maxima" Trap
Fitting these models is like trying to find the highest peak in a foggy mountain range. If you start in the wrong spot, you might get stuck on a small hill thinking it's the highest point (a "local maximum").

• The authors show that by identifying the seasonal roots first, you can narrow down your search. You know exactly which "hills" to climb, saving time and preventing errors.

Real-World Examples: Viruses and Soybeans

The paper tested this theory on two real-world datasets:

1. COVID-19 Deaths (Belgium & Italy)

• The Data: Daily death counts showed a "zig-zag" pattern (up one day, down the next) that got more intense over time.
• The Insight: The authors found that this zig-zag was a "seasonal bubble." It wasn't just random noise; it was a specific pattern driven by the "future" part of the model (noncausal).
• The Result: By correctly separating the seasonal roots, they built a model that perfectly captured the zig-zag behavior, whereas standard models missed it.

2. Soybean Prices

• The Data: Soybean prices have a 6-month cycle (harvest times) and occasional price explosions.
• The Insight: Standard models thought the cycle was 8 months long. But by using the new method, the authors realized the "future" part of the model was actually driving a 6-month cycle.
• The Result: They corrected the model to show a 6-month rhythm, which makes much more sense for farming.

The Takeaway

This paper is like a instruction manual for untangling a knot.
It tells us that even in the most complex, "magical" time-series models (where the future influences the present), the repeating seasonal patterns are stable and predictable. They don't get confused or mixed up.

For the average person: If you are trying to understand why something keeps happening in a cycle (like holiday sales or flu seasons), you don't need to fear that the math is creating fake patterns. You can use a simple "rough draft" to find the real patterns, sort them into "past" and "future" buckets, and build a much more accurate prediction tool.

Technical Summary

1. Problem Statement

The paper addresses a gap in the literature regarding Mixed Causal-Noncausal Autoregressive (MAR) models. While MAR models are well-established for capturing nonlinear dynamics like speculative bubbles (driven by roots at the zero frequency), their behavior regarding seasonality (driven by negative real roots or complex conjugate roots) is less understood.

Key questions include:

• How do seasonal roots propagate through the multiplicative structure of causal and noncausal polynomials in MAR models?
• Can the interaction between causal and noncausal components generate new seasonal effects not present in the individual polynomials?
• How does the presence of seasonal roots (specifically complex conjugate pairs) affect the identification, estimation, and model selection of MAR processes?

2. Methodology

The authors employ a combination of theoretical derivation, simulation, and empirical application:

• Theoretical Framework:

  • Partial Fraction Decomposition (PFD): The core theoretical tool used to decompose the autoregressive polynomials into monomials associated with specific inverse roots.
  • Root Characterization: Roots are analyzed using their exponential form ($\alpha_k = \rho_k e^{i\omega_k}$), distinguishing between real roots (zero and Nyquist frequencies) and complex conjugate pairs (harmonic frequencies).
  • Model Representations: The study contrasts the "strong form" (MAR with i.i.d. non-Gaussian errors) with the "pseudo-causal" or "second-order equivalent" (SOE) representation (a purely causal AR model that matches the second-order properties of the MAR process).

• Estimation and Selection Strategy:

  • Pseudo-Causal Step: Estimate a purely causal AR model to identify the total autoregressive order ($p$) and the set of all roots.
  • Root Allocation: Assign the identified roots to the causal ($r$) and noncausal ($q$) polynomials. A critical constraint is that complex conjugate pairs must be assigned jointly to either the causal or noncausal polynomial to ensure real-valued coefficients.
  • Approximate Maximum Likelihood Estimation (AMLE): Estimate the final MAR($r, q$) models using non-Gaussian error distributions (e.g., Student's $t$).
  • Model Selection: Select the model ($r, q$) that maximizes the log-likelihood function.

• Empirical & Simulation Tools:

  • Monte Carlo Simulations: Used to test the consistency of root recovery in the pseudo-causal model and the accuracy of model selection under various sample sizes and root configurations.
  • Periodogram Analysis: Used to detect seasonal frequencies and verify stationarity.
  • Datasets: Daily COVID-19 death variations (Belgium and Italy) and monthly soybean prices.

3. Key Contributions and Theoretical Findings

A. Isolation of Seasonal Effects

The paper proves that no new seasonal effects are generated by the multiplicative interaction of causal and noncausal polynomials.

• Using partial fraction decomposition, the authors demonstrate that the moving average (MA) representation of an MAR process can always be decomposed into terms corresponding to the individual roots of the causal and noncausal polynomials.
• Seasonal frequencies (associated with complex roots or negative real roots) remain isolated. The "seasonality" of the system is simply the sum of the seasonalities of its components; the multiplicative structure does not create cross-frequency interactions that generate new seasonal peaks.

B. Implications for Model Selection

The presence of seasonal roots, particularly complex conjugate pairs, significantly simplifies the model selection problem:

• Constraint on Combinations: In a standard MAR($r, q$) selection with total order $p$, one must test all combinations of assigning $p$ roots to $r$ and $q$. However, if a complex conjugate pair exists, they must be assigned together to either the causal or noncausal side.
• Reduction of Search Space: This constraint drastically reduces the number of feasible model specifications. For example, if $p=4$ and there is one complex pair, an MAR(1,1) specification becomes impossible because it would require splitting the pair, resulting in complex coefficients. This forces the model into purely causal/noncausal or higher-order mixed structures.

C. Identification via Pseudo-Causal Models

The study validates that the pseudo-causal (SOE) representation is a robust tool for:

1. Determining the total order $p$.
2. Recovering the set of all roots (including complex ones) consistently.
3. Providing starting values for AMLE, provided that complex pairs are kept intact during the allocation process.

4. Results

Monte Carlo Simulations

• Root Recovery: The pseudo-causal model consistently recovers the true roots (both modulus and frequency) as the sample size increases, even when the underlying process is mixed causal-noncausal.
• Model Selection Accuracy: The AMLE selection procedure successfully identifies the true MAR specification.
  • When complex conjugate roots are present, the procedure correctly rejects "naive" mixed specifications (like MAR(1,1)) that split the pair.
  • As sample size ($T$) and the magnitude of noncausal coefficients increase, the probability of selecting the correct model approaches 100%.

Empirical Applications

1. COVID-19 Data (Belgium & Italy):

   • Italy: The pseudo-causal model revealed a complex conjugate pair. This ruled out MAR(1,1). The final selected model was a purely causal MAR(2,0), which fit the data significantly better than mixed alternatives.
   • Belgium: The data showed a "zig-zag" pattern (Nyquist frequency) and complex roots. The selected model was MAR(2,2). Crucially, the authors showed that assigning the complex pair to the causal part and the real roots to the noncausal part yielded a superior fit (higher log-likelihood) compared to the reverse assignment. This highlights the importance of correct root allocation to capture specific dynamics (e.g., the noncausal part capturing the explosive "bubble" zig-zag).

2. Soybean Prices:

   • The pseudo-causal AR(5) model identified roots at zero, Nyquist, and a complex pair.
   • The selected model was MAR(2,3).
   • Refinement: The allocation of roots in the MAR model differed slightly from the pseudo-causal estimates (a root shifted from Nyquist to zero frequency). This suggests that while the pseudo-causal model provides a good starting point, the non-Gaussian likelihood maximization refines the root allocation to better fit the explosive nature of the data. The MAR model captured peaks and troughs better than the purely causal AR(5).

5. Significance and Conclusion

This paper provides a rigorous theoretical foundation for handling seasonality in MAR models, which are increasingly used for modeling financial bubbles and economic cycles.

• Theoretical Clarity: It dispels the notion that MAR models generate complex, unobservable seasonal interactions, proving instead that seasonal effects are additive and isolatable.
• Practical Guidance: It offers a concrete, step-by-step procedure for practitioners:
  1. Estimate the pseudo-causal model.
  2. Identify roots and check for complex conjugate pairs.
  3. Enforce the "joint assignment" rule for complex pairs to reduce the model selection space.
  4. Use AMLE to finalize the model, ensuring the correct allocation of roots to causal/noncausal polynomials to avoid local maxima.
• Econometric Impact: The findings suggest that explicitly formulating multiplicative seasonal MAR models (SMAR) is unnecessary, as the general MAR framework with seasonal roots is a more flexible and encompassing representation.

In summary, the paper demonstrates that understanding the spectral properties of roots (specifically complex conjugates) is the key to unlocking the correct specification of mixed causal-noncausal models, leading to more accurate modeling of nonlinear seasonal dynamics in economic and financial time series.