📈 economics

Fitting Dynamically Misspecified Models: An Optimal Transportation Approach

This paper proposes a sequential optimal transportation framework to address dynamic model misspecification in state-space models by mapping observations to a model-consistent sample, providing new methods for filtering, parameter estimation, and specification testing.

Original authors: Jean-Jacques Forneron, Zhongjun Qu

Published 2026-04-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jean-Jacques Forneron, Zhongjun Qu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to follow a recipe to bake the perfect sourdough bread. The recipe (your Model) tells you exactly how much flour, water, and salt to use and how long to let it rise.

However, when you actually start baking, you realize your kitchen is different from the chef's kitchen. Your flour might be a different brand, your oven might run five degrees too hot, and the humidity in your house is much higher than the chef's. If you follow the recipe exactly as written, your bread will come out wrong. You might try to "force" it to work, but you'll end up with a weird, dense loaf that doesn't make sense.

This paper is about a new, smarter way to "adjust the recipe" so that the results actually match the reality of your kitchen, without losing the soul of the original recipe.

The Problem: The "Rigid Recipe" Trap

In economics, scientists use complex mathematical models (the Recipes) to explain things like inflation or unemployment. These models are "structural," meaning they try to explain why things happen (e.g., "Inflation rose because people spent more money").

The problem is that these models are almost always "misspecified." They are simplifications of a messy, complicated world. When economists use standard tools (like the Kalman Filter) to fit these models to real-world data, two bad things happen:

The "Frankenstein" Effect: The model produces results that contradict itself. It might say "shocks are independent," but the data shows they are moving together. It’s like a recipe that says "use only cold water," but you end up using hot water just to make the dough stick.
The "Perfect Fit" Lie: Standard tools are so desperate to match the data that they "distort" the underlying causes. They might make it look like a sudden economic shock happened, when really, it was just a slow, natural trend. They "force" the fit, which makes the explanation untrustworthy.

The Solution: Optimal Transport (The "Smart Translator")

The authors propose a new method called Optimal Transport Filtering (OTF).

Think of Optimal Transport as a highly skilled Translator.

Imagine you have a book written in a very strict, formal language (the Model), but the real world is speaking in a messy, colorful slang (the Data). If you try to translate the slang word-for-word using the formal rules, the meaning is lost.

Instead, the "Optimal Transport" approach looks at the essence of what is being said. It finds the most efficient way to map the messy slang into the formal language so that the meaning (the statistical patterns, the trends, and the volatility) remains identical.

It doesn't just "force" the data to fit the model; it creates a "Model-Consistent Sample." It essentially says: "If the world actually followed this recipe, given how messy our kitchen is, this is what the bread would actually look like."

How it Works (The Three Steps)

Listen to the Data: First, the tool uses a flexible method (like a VAR) to listen to the real-world data without any preconceived notions. It learns the "slang" of the economy.
The Transport Map: It then calculates a mathematical "bridge" (the Transport Map). This bridge moves the data from the "messy reality" to the "structured model" in the most efficient way possible, minimizing the "distance" between them.
The Consistent Result: The result is a version of the data that looks like the real world but obeys all the rules of the model. Now, when an economist looks at the results, they can trust them. If the model says "shocks are independent," the filtered data will actually show them as independent.

Why This Matters

The paper proves this works using famous economic models (like DSGE models used by central banks).

It catches lies: It provides a "Specification Test." If the "bridge" between the data and the model is too long or too difficult to build, the test screams, "Hey! This model is fundamentally wrong!"
It handles "Singular" models: Some models are so complex they are mathematically "broken" (singular) for standard tools. This new method handles them easily.
It makes sense of the past: When looking at history (like US GDP), standard tools often produce "ghost" cycles that don't exist. This new method produces much cleaner, more interpretable economic cycles.

In short: This paper gives economists a way to use their theoretical "recipes" in the "messy kitchen" of the real world without the results turning into a mathematical disaster.

Technical Summary: Fitting Dynamically Misspecified Models: An Optimal Transportation Approach

Authors: Jean-Jacques Forneron and Zhongjun Qu
Date: April 27, 2026 (as per document)

1. The Problem: The Cost of Dynamic Misspecification

In structural econometrics (e.g., DSGE or asset pricing models), researchers use state-space models to recover unobserved latent variables (like productivity or risk) from observed data. However, these models are almost always misspecified—the mathematical functions $g$ and $h$ are simplifications of the true, more complex data-generating process (DGP).

The authors identify three critical failures that occur when standard filtering (like the Kalman Filter) is applied to misspecified models:

Violation of Model Restrictions: Filtered latent variables (shocks) often fail to satisfy the model's own structural constraints. For example, shocks specified as independent may appear cross-correlated in the filtered series, or a mean-zero process may yield a non-zero filtered mean.
Interpretability Crisis: Because the filter tries to force the model to fit the data, it "distorts" the shocks to achieve a perfect fit of the observables. This makes the filtered series (e.g., the output gap) difficult to interpret economically.
Likelihood and Dimensionality Issues: Standard Maximum Likelihood Estimation (MLE) is undefined when there are fewer structural shocks than observables (stochastic singularity), limiting the ability to perform model-based dimension reduction.

2. Methodology: Sequential Optimal Transportation (OT)

The paper proposes a novel framework that uses Optimal Transportation to construct a "model-consistent" sample that is as close as possible to the observed data.

A. The Optimal Transport Filter (OTF):
Instead of filtering the raw observations $\tilde{y}_t$ , the OTF iteratively maps them to a new series $y_t$ that satisfies the model's dynamics.

Auxiliary Model: The researcher first uses a flexible, semiparametric reduced-form model (a Vector Autoregression, VAR) to approximate the true dynamics of the data.
The Transport Step: At each time step $t$ , the algorithm solves a minimization problem: it finds a mapping that transforms the conditional distribution of the observed data into the conditional distribution predicted by the structural model, minimizing the mean-squared error (Wasserstein distance).
Linear Case (Closed-Form): For linear state-space models, the authors derive a closed-form "plugin" rule for the transport map $P$ , making the algorithm computationally efficient and easy to implement via standard matrix operations.

B. Optimal Transport Estimation (OTE):
To estimate the structural parameters $\theta$ , the authors minimize the discrepancy between the original observations $\tilde{y}_t$ and the model-consistent "coupled" sample $y_t$ generated by the OTF. This is a semiparametric approach because the auxiliary VAR captures the data's moments without requiring a specific distribution.

C. Specification Testing:
The authors propose a new test based on the residual discrepancy between the actual sample path and the model-consistent sample path. If the model is correctly specified, the discrepancy should be statistically indistinguishable from zero.

3. Key Contributions

Mathematical Innovation: Derivation of a closed-form optimal transport map for linear processes that preserves the linearity of the system while ensuring model consistency.
Robustness to Singularity: Unlike MLE, the OTE framework remains well-defined even in "stochastically singular" models (where there are fewer shocks than observables), allowing for more parsimonious structural models.
Theoretical Rigor: The paper provides proofs for the consistency and asymptotic normality ( $\sqrt{n}$ -rate) of the OTE under both correct specification and misspecification.
New Inference Tools: Development of a specification test that can pinpoint exactly which variables (e.g., inflation vs. consumption) are causing the model to fail.

4. Empirical Results

The authors validate the methodology across several complex domains:

DSGE Models (Lubik & Schorfheide; Smets & Wouters): The OTF produces latent shocks that actually obey the model's covariance structure, unlike the Kalman Filter. In the Smets-Wouters model, the OTF successfully identifies that the model fails to capture the volatility and persistence of consumption.
Affine Term Structure Models: The method demonstrates effective dimension reduction, showing how three latent factors can explain the vast majority of the variation in six different bond yields.
Trend-Cycle Decomposition: In U.S. GDP analysis, the OTF produces a stationary, zero-mean cycle that accurately marks NBER recession dates, whereas the Kalman Filter produces a biased, non-stationary cycle due to trend misspecification.

5. Significance

This paper represents a significant advancement in semiparametric structural estimation. By shifting the focus from "fitting the data to the model" (which causes distortion) to "mapping the data into a model-consistent space," the authors provide a way to conduct structural analysis that is both mathematically valid and economically interpretable, even when the underlying theory is imperfect. It bridges the gap between flexible reduced-form modeling and rigid structural modeling.