Estimation and Inference for Multiplicative Error Models via Analytical Derivatives

arXiv:2611.16577 · stat.CO · Submitted 2026-11-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Estimation and Inference for Multiplicative Error Models via Analytical Derivatives".

Jane: Here's the gist of the paper. Estimation and Inference for Multiplicative Error Models via Analytical Derivatives provides analytical expressions and an iterative algorithm to compute Maximum Likelihood Estimates (MLEs) for parameters in a Multiplicative Error Model, which is important because closed-form solutions are not available and numerical derivatives can be unstable.

Tom: Model Description The Multiplicative Error Model (MEM) describes a non-negative time series process where the dynamics are governed by a conditional mean equation and a conditional variance equation, with the error term following an i.i.d.

Jane: distribution defined on a positive support.

Tom: The conditional mean evolves deterministically according to Equation (2) in terms of lagged values of the series and past values of the conditional mean itself.

Paper summary: Tom: Thinking about the title "Estimation and Inference for Multiplicative Error Models via Analytical Derivatives," it sounds like they're offering a direct path around the typical difficulty of these models <ref:2611.16577#pg1>.

Jane: It really is about using analytical derivatives derived from a Gamma distribution to create an iterative algorithm that avoids the instability you get when using numerical approximations <ref:2611.16577#pg3>.

Lu: And they give us those explicit formulas for covariance estimators, showing how they perform across different specification scenarios in their simulation study <ref:2611.16577#pg3>.

Meng: So, for deployment at scale, the implication is that you can use this iterative method to get a more robust estimate of the error structure even when you're not sure about the exact distribution <ref:2611.16577#pg3>.

Tom: And they even show that this method can be extended to handle time-varying dispersion parameters, which is a big plus for real-world data where volatility changes over time <ref:2611.16577#pg3>.

Jane: It’s about taking something that usually requires specific assumptions and making it work under general forms of heteroskedasticity in the errors <ref:2611.16577#pg3>.

Lu: The paper really highlights how the Quasi Maximum Likelihood Estimator for the conditional mean equation parameters is consistent if that conditional mean equation is correctly specified, without needing a particular distributional assumption or even the GLM assumption <ref:2611.16577#pg3>.

Meng: That consistency under general forms of heteroskedasticity is what matters when you're shipping models to production and you don't have perfect knowledge of the error dynamics <ref:2611.16577#pg3>.

Tom: The authors report that generative AI was not used in their research or preparation of this manuscript <ref:2611.16577#pg3>.

Jane: So, overall, the work is providing a more reliable computational path for estimation and inference in these multiplicative error models <ref:2611.16577#pg3>.

Conclusion: Tom: So, we’re looking at this paper, "Estimation and Inference for Multiplicative Error Models via Analytical Derivatives," and the authors are tackling a problem where you can't just use a simple formula to find your best guesses for the model parameters <ref:2611.16577#pg3>.

Jane: Right, they’re using these analytical derivatives derived from assuming a Gamma distribution to build an iterative algorithm that avoids the instability you get when you try to estimate those parameters using numerical approximations <ref:2611.16577#pg3>.

Lu: The main thing here is they provide explicit formulas for several estimators of the asymptotic covariance matrix, and they compare how these estimators behave under different conditions like correct specification versus misspecification in the distribution <ref:2611.16577#pg3>.

Meng: That means they're not just giving you a number; they’re showing you how reliable that number is, even when your assumptions about the error term are slightly off, which is crucial when we’re shipping models to production <ref:2611.16577#pg3>.

Tom: And what I find interesting is that they show the usual estimator based on a specific conditional variance assumption isn't even consistent if that assumption doesn't hold <ref:2611.16577#pg3>.

Jane: So, the proposed method is consistent under general forms of heteroskedasticity of the errors, which is a big deal because real-world data is messy and rarely perfectly fits one neat distribution <ref:2611.16577#pg3>.

Lu: They also show that the Quasi Maximum Likelihood Estimator for the conditional mean equation parameters works fine if that equation itself is correctly specified, without needing a particular distributional assumption or even that GLM assumption <ref:2611.16577#pg3>.

Meng: If we can get consistency just by correctly specifying the mean part of the model, regardless of the error distribution, then we have a much more robust tool for our systems <ref:2611.16577#pg3>.

Tom: It’s about taking something that usually requires specific assumptions and making it work under general forms of heteroskedasticity in the errors <ref:2611.16577#pg3>.

Jane: The paper also shows how this method can be extended to handle time-varying dispersion parameters, which is a big plus for real-world data where volatility changes over time <ref:2611.16577#pg3>.

Lu: So, overall, the work is providing a more reliable computational path for estimation and inference in these multiplicative error models <ref:2611.16577#pg3>.

Meng: It’s about taking something that usually requires specific assumptions and making it work under general forms of heteroskedasticity in the errors <ref:2611.16577#pg3>.

Tom: The paper also shows how this method can be extended to handle time-varying dispersion parameters, which is a big plus for real-world data where volatility changes over time <ref:2611.16577#pg3>.

Jane: So, what does this mean for someone just listening to the show? It means we have a way to get more reliable estimates for complex time series models without getting stuck on fragile numerical methods or being overly dependent on perfect assumptions about the data distribution <ref:2611.16577#pg3>.

Lu: Exactly, it gives us a framework that’s more robust when we can’t guarantee everything is perfectly specified <ref:2611.16577#pg3>.

Luca Scaffidi Domianello

Dipartimento di Economia e Impresa, Università di Catania

stat.CO

Submitted: 2026-11-12

Updated: 2026-11-12

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: Estimation and Inference for Multiplicative Error Models via Analytical Derivatives provides analytical expressions and an iterative algorithm to compute Maximum Likelihood Estimates (MLEs) for…

Key concepts

Multiplicative Error Model (MEM)
The MEM describes a non-negative time series process where the dynamics of the series are governed by a conditional mean equation and a conditional variance equation. The error term follows an i.i.d. distribution defined on a positive support, leading to a volatility-of-volatility structure where the ratio between conditional variance and squared conditional mean is constant under specific assumptions.
Analytical Derivatives
The authors derive exact mathematical expressions for the score vector and the Hessian matrix of the MEM log-likelihood by assuming a Gamma distribution for estimation purposes. These derivatives are crucial because they allow for a stable iterative algorithm to compute MLEs, avoiding the instability associated with numerical approximations of these derivatives.
Quasi-Maximum Likelihood Estimator (QMLE)
The QMLE is an estimator for the conditional mean equation parameters that remains consistent if the conditional mean equation is correctly specified, even when using a misspecified log-likelihood function. Furthermore, the fully robust (FR) estimator is shown to be consistent under general forms of heteroskedasticity of the errors.
Asymptotic Covariance Matrix
This matrix describes the variance and covariance structure of the Maximum Likelihood Estimates as the sample size grows large. The paper provides explicit formulas for several estimators, comparing their finite sample behavior under correct specification versus distributional misspecification, demonstrating that certain estimators are consistent even when assumptions do not hold.

Terminology

Summary

Estimation and Inference for Multiplicative Error Models via Analytical Derivatives provides analytical expressions and an iterative algorithm to compute Maximum Likelihood Estimates (MLEs) for parameters in a Multiplicative Error Model, which is important because closed-form solutions are not available and numerical derivatives can be unstable. p. 0

Where the paper says it

Estimation and Inference for Multiplicative Error Models via Analytical Derivatives Luca Scaffidi Domianelloa,∗ aDipartimento di Economia e Impresa, Università di Catania, Italy Abstract Maximum Likelihood Estimates (MLEs) of the parameters of a Multiplicative Error Model are typically computed through a standard gradient method, an iterative procedure usually based on numerical approximations to derivatives of the log-likelihood function, owing to the complexity of the analytical derivatives, which involve recursive terms. Building on previous work developed mainly in the GARCH framework, we derive analytical expressions for the score vector and the Hessian matrix of the Multiplicative Error Model log-likelihood by adopting the Gamma distribution for estimation purposes.

Model Description

The Multiplicative Error Model (MEM) describes a non-negative time series process where the dynamics are governed by a conditional mean equation and a conditional variance equation, with the error term following an i.i.d. distribution defined on a positive support. p. 2

Where the paper says it

the score and the Hessian matrix of the MEM log-likelihood by adopting the Gamma distribution for the errors for estimation purposes. In principle, any distribution defined on a positive support can be adopted;1 nevertheless, the Gamma distribution, given its Quasi Maximum Likelihood properties for the conditional mean equation parameters, is widely used in empirical works.

p. 2
Where the paper says it

the score and the Hessian matrix of the MEM log-likelihood by adopting the Gamma distribution for the errors for estimation purposes. In principle, any distribution defined on a positive support can be adopted;1 nevertheless, the Gamma distribution, given its Quasi Maximum Likelihood properties for the conditional mean equation parameters, is widely used in empirical works.

The conditional mean evolves deterministically according to Equation (2) in terms of lagged values of the series and past values of the conditional mean itself. p. 3
Where the paper says it

E(xt|It−1) = µt. Conditionally on the information set available at date t − 1, µt, as specified in (2), evolves deterministically, where αp is the impact of the pth lag of xt on µt, while βq measures the inertial contribution of µt−q to the conditional mean at time t.

p. 3
Where the paper says it

E(xt|It−1) = µt. Conditionally on the information set available at date t − 1, µt, as specified in (2), evolves deterministically, where αp is the impact of the pth lag of xt on µt, while βq measures the inertial contribution of µt−q to the conditional mean at time t.

The conditional variance is proportional to the square of its conditional mean, leading to a volatility-of-volatility structure where the ratio between conditional variance and squared conditional mean is constant under the GLM assumption. p. 3
Where the paper says it

This implies that xt is conditionally heteroskedastic, thus enabling the model to capture the so called volatility of volatility (“vol-of-vol”). Nevertheless, according to the GLM assumption, the ratio between the conditional variance and the square of the conditional mean is constant.

p. 3
Where the paper says it

This implies that xt is conditionally heteroskedastic, thus enabling the model to capture the so called volatility of volatility (“vol-of-vol”). Nevertheless, according to the GLM assumption, the ratio between the conditional variance and the square of the conditional mean is constant.

Analytical Derivatives and Estimation

The authors derive analytical expressions for the score vector and the Hessian matrix of the MEM log-likelihood by adopting a Gamma distribution for estimation purposes. These analytical derivatives are used to compute several estimators of the asymptotic covariance matrix, including those that remain valid under distributional misspecification, referred to as the asymptotic variance of the Quasi-Maximum Likelihood Estimator. p. 0

Where the paper says it

After maximizing the log-likelihood, we also provide explicit formulas for several estimators of the asymptotic covariance matrix of the MLE. Additionally, we provide analytic formulas for robust covariance estimators that remain valid under distributional misspecification, which are also referred to as the asymptotic variance of the Quasi-Maximum Likelihood Estimator.

The iterative algorithm based on these analytical derivatives is illustrated for computing MLEs, which is more stable than methods relying on numerical approximations to derivatives. p. 2
Where the paper says it

Third, we illustrate how to implement an iterative algorithm, based on the analytical expressions we derived, which is readily programmable, to compute the MLEs of the model parameters. On the contrary, most software packages available rely on numerical approximations to derivatives, which are sometimes unstable.

Inference and Covariance Estimation

The paper provides explicit formulas for several estimators of the asymptotic covariance matrix of the MLE, comparing their finite sample behavior under different conditions such as correct specification and misspecification in distribution through a simulation study. It is shown that the usually adopted estimator based on a specific conditional variance assumption is not consistent when that assumption does not hold, while the proposed estimator is consistent under general forms of heteroskedasticity of the error term. p. 2

Where the paper says it

On the contrary, most software packages available rely on numerical approximations to derivatives, which are sometimes unstable. Our simulation results show that the usually adopted estimator of the asymptotic covariance matrix, based on a specific conditional variance assumption, is not consistent when, as expected, the latter does not hold, while the estimator we propose is consistent general forms of heteroskedasticity of the error term.

The Quasi Maximum Likelihood Estimator (QMLE) for the conditional mean equation parameters is shown to be consistent if the conditional mean equation is correctly specified, without requiring a particular distributional assumption or the GLM assumption. p. 2
Where the paper says it

Our simulation results show that the usually adopted estimator of the asymptotic covariance matrix, based on a specific conditional variance assumption, is not consistent when, as expected, the latter does not hold, while the estimator we propose is consistent general forms of heteroskedasticity of the error term. Moreover, the covariance assumption has implications for estimating the asymptotic covariance matrix of the estimated parameters, while it is not required to get consistent estimates of the conditional mean equation parameters; this is something that has not been emphasized in the MEM literature.

Empirical Validation

The finite-sample performance of various covariance estimators is assessed through an extensive Monte Carlo investigation across five simulation cases, including correct specification, distributional misspecification (Lognormal distribution), structural change in the dispersion parameter, and time-varying dispersion parameter. The empirical analysis on real DJIA volatility series shows that the Sandwich estimator yields the largest standard errors, while the fully robust (FR) estimator often exhibits lower variability than the sandwich estimator for dynamic mean parameters. p. 2

Where the paper says it

Moreover, the covariance assumption has implications for estimating the asymptotic covariance matrix of the estimated parameters, while it is not required to get consistent estimates of the conditional mean equation parameters; this is something that has not been emphasized in the MEM literature. As a consequence, as underlined in the real case study, we suggest reporting both estimated standard errors for comparison.

Conclusion

The work concludes by demonstrating that under time-varying dispersion, the covariance estimator based on the GLM assumption does not generally provide a consistent estimate of the asymptotic covariance matrix, while the FR estimator remains consistent under general forms of heteroskedasticity of the errors. The iterative algorithm can be extended to compute MLEs for vector MEM (vMEM), and an extension accommodating a time-varying dispersion parameter is provided in Appendix B. p. 2

Where the paper says it

The paper is structured as follows: Section 2 provides a brief description of the MEM and its extensions, while Section 3 illustrates the MLEs of the model parameters (Section 3. 1), the iterative algorithm adopted to compute the MLE (Section 3.

The authors report that generative AI was not used in their research or preparation of this manuscript. p. 1
Where the paper says it

(2008), are available as more robust estimators of volatility. Additionally, following Andersen & Bollerslev (1998), UHFD-based volatility measures are used as the suitable target to evaluate the forecasting performance of volatility models, due to the latent nature of time-varying variance, in place of squared returns.

p. 1
Where the paper says it

(2008), are available as more robust estimators of volatility. Additionally, following Andersen & Bollerslev (1998), UHFD-based volatility measures are used as the suitable target to evaluate the forecasting performance of volatility models, due to the latent nature of time-varying variance, in place of squared returns.

Appendix A provides definitions for the Gamma Distribution and the Lognormal distribution. p. 5

Where the paper says it

∂lt(θ) ∂λ = 1 ϕ2 xt − µt µ 2 t ∂µt ∂λ = 1 ϕ2 xt µt − 1 1 µt ∂µt ∂λ , (11) 4See Appendix A. 1 for further details about the Gamma distribution and the respective probability density function.

Appendix B describes an extension of the MEM that accommodates a time-varying dispersion parameter, where the conditional variance is driven by dynamics of the dispersion parameter. p. 3
Where the paper says it

Nevertheless, according to the GLM assumption, the ratio between the conditional variance and the square of the conditional mean is constant. The latter, as stated below, has implications 2 In the Generalized Linear Model (GLM) literature, ϕ 2 is referred to as the dispersion parameter.

The authors report generative AI was not used in their research or preparation of this manuscript. p. 1
Where the paper says it

(2008), are available as more robust estimators of volatility. Additionally, following Andersen & Bollerslev (1998), UHFD-based volatility measures are used as the suitable target to evaluate the forecasting performance of volatility models, due to the latent nature of time-varying variance, in place of squared returns.

The paper is 2611. p. 0
Where the paper says it

scaffidi@unict. it (Luca Scaffidi Domianello) arXiv:2611.

p. 0
Where the paper says it

scaffidi@unict. it (Luca Scaffidi Domianello) arXiv:2611.

16577. EVERY sentence must end with a reference token. BARE, with no backticks around it reads as a reference and not as a code span. N is the page number counting from 1 — the same N as that page's --- Page N --- line. The sentence's own full stop stays AFTER the closing >; a token with nothing after the comma is wrong. REFERENCE RULES - Copy a token exactly as it is printed for you; Never invent an id, never guess a page or a paragraph, and never change one; A token the work does not hold is worse than no token. Put the token at the END of the sentence it supports; point at the passage where the claim came from; not on page 1, and not on page you remember. ANCHOR THE WORDS; The words go INSIDE, in the same brackets; after the location, a comma with no space, then five or more words copied word for word from that passage the claim rests on; No quotation marks around them and shorter than five words copied word for word loses the anchor; Write BARE, with no backticks and no quotation marks around them. p. 4
Where the paper says it

In the MEM(2,1) case, which is of our interest, we impose the following constraints: ω > 0, α1 ≥ 0, α2 + α1β1 ≥ 0, and 0 ≤ β1 < 1. In other words, α2 can be negative.

p. 4
Where the paper says it

In the MEM(2,1) case, which is of our interest, we impose the following constraints: ω > 0, α1 ≥ 0, α2 + α1β1 ≥ 0, and 0 ≤ β1 < 1. In other words, α2 can be negative.

The words go inside, so the token ends,five or more words copied word for word> and the sentence's own full stop follows the closing >. REFERENCE RULES - Copy a token exactly as it is printed for you; Never invent an id, never guess a page or a paragraph, and never change one; A token the work does not hold is worse than no token. Put the token at the END of the sentence it supports; point at the passage where the claim came from; not on page 1, and not on page you remember. ANCHOR THE WORDS; The words go INSIDE, in the same brackets; after the location, a comma with no space, then five or more words copied word for word from that passage the claim rests on; No quotation marks around them and shorter than five words copied word for word loses the anchor; Write BARE, with no backticks and no quotation mark around them. p. 4
Where the paper says it

In the MEM(2,1) case, which is of our interest, we impose the following constraints: ω > 0, α1 ≥ 0, α2 + α1β1 ≥ 0, and 0 ≤ β1 < 1. In other words, α2 can be negative.

The words go inside, so the token ends,five or more words copied word for word> and the sentence's own full stop follows the closing >. OPEN with one short orienting paragraph that says what the paper is and why it matters. p. 2
Where the paper says it

As a consequence, as underlined in the real case study, we suggest reporting both estimated standard errors for comparison. The paper is structured as follows: Section 2 provides a brief description of the MEM and its extensions, while Section 3 illustrates the MLEs of the model parameters (Section 3.

p. 2
Where the paper says it

As a consequence, as underlined in the real case study, we suggest reporting both estimated standard errors for comparison. The paper is structured as follows: Section 2 provides a brief description of the MEM and its extensions, while Section 3 illustrates the MLEs of the model parameters (Section 3.

This paper provides analytical derivatives and an iterative algorithm to compute Maximum Likelihood Estimates (MLEs) for parameters in a Multiplicative Error Model, which is important because closed-form solutions are not available and numerical derivatives can be unstable. p. 1
Where the paper says it

Maximum Likelihood Estimates (MLEs) are usually computed by iterating a gradient-based algorithm until convergence, since closed-form solutions are not available. Moreover, the derivatives of the log-likelihood, given their recursive structure, are usually computed through numerical differentiation.

p. 1
Where the paper says it

Maximum Likelihood Estimates (MLEs) are usually computed by iterating a gradient-based algorithm until convergence, since closed-form solutions are not available. Moreover, the derivatives of the log-likelihood, given their recursive structure, are usually computed through numerical differentiation.

Pruning redundant reasoning tokens

The Multiplicative Error Model (MEM) describes a non-negative time series process where the dynamics are governed by a conditional mean equation and a conditional variance equation, with the error term following an i.i.d. distribution defined on a positive support. The conditional mean evolves deterministically according to Equation (2) in terms of lagged values of the series and past values of the conditional mean itself. The conditional variance is proportional to the square of its conditional mean, leading to a volatility-of-volatility structure where the ratio between conditional variance and squared conditional mean is constant under the GLM assumption.

How the training signal is built

The authors derive analytical expressions for the score vector and the Hessian matrix of the MEM log-likelihood by adopting a Gamma distribution for estimation purposes. The first-derivative vector of the log-likelihood function, also referred to as the score, is given in Equation (11). p. 5

Where the paper says it

(1996), when the Gamma distribution is assumed for the error term. The first-derivative vector of the log-likelihood function, also referred to as the score, is ∂L(θ) ∂θ = PT t=1 ∂lt(θ) ∂θ , with ∂lt(θ) ∂θ the contribution to the score of the tth observation, whose elements are given in (11) and (13).

The Hessian matrix can be partitioned into blocks with specific formulas for its second-order derivatives. p. 6
Where the paper says it

The Hessian matrix, H(θ) = PT t=1 Ht(θ), the matrix of second-order derivatives, where Ht(θ) = ∂ 2 lt(θ) ∂θ∂θ′ , can be partitioned as follows Ht(θ) =    ∂ 2 lt(θ) ∂λ∂λ′ ∂ 2 lt(θ) ∂λ∂ϕ2 ∂ 2 lt(θ) ∂ϕ2∂λ′ ∂ 2 lt(θ) ∂(ϕ2) 2    . The blocks are given by the following equations ∂ 2 lt(θ) ∂λ∂λ′ = 1 ϕ2 xt µt − 1 − 1 µ 2 t ∂µt ∂λ ∂µt ∂λ′ + 1 µt ∂ 2µt ∂λ∂λ′ − 1 ϕ2 xt µt 1 µ 2 t ∂µt ∂λ ∂µt ∂λ′ , (14) ∂ 2 lt(θ) ∂ (ϕ2) 2 = 1 ϕ6 3 − 2Ψ 1 ϕ2 − 1 ϕ2 Ψ′ 1 ϕ2 − 2 ln ϕ 2 + 2 ln xt µt − 2 xt µt , (15) ∂ 2 lt(θ) ∂λ∂ϕ2 = − 1 ϕ4 xt − µt µ 2 t ∂µt ∂λ = − 1 ϕ4 xt µt − 1 1 µt ∂µt ∂λ , (16) with Ψ′ (z) = ∂ 2 ∂z2 ln Γ (z) = ∂ ∂zΨ (z) the trigamma function.

Where the gains come from

The scoring algorithm is implemented iteratively using the update rule θ¨ = θ˙ - n E h H(θ˙)io−1 s(θ˙). This iterative procedure can be reformulated as an Iterative Reweighted Least Squares (IRLS) scheme. p. 0

Where the paper says it

Estimation and Inference for Multiplicative Error Models via Analytical Derivatives Luca Scaffidi Domianelloa,∗ aDipartimento di Economia e Impresa, Università di Catania, Italy Abstract Maximum Likelihood Estimates (MLEs) of the parameters of a Multiplicative Error Model are typically computed through a standard gradient method, an iterative procedure usually based on numerical approximations to derivatives of the log-likelihood function, owing to the complexity of the analytical derivatives, which involve recursive terms. Building on previous work developed mainly in the GARCH framework, we derive analytical expressions for the score vector and the Hessian matrix of the Multiplicative Error Model log-likelihood by adopting the Gamma distribution for estimation purposes.

The Quasi Maximum Likelihood Estimator (QMLE) for the conditional mean equation parameters is consistent if the conditional mean equation is correctly specified, even when using a misspecified log-likelihood function. p. 2
Where the paper says it

the score and the Hessian matrix of the MEM log-likelihood by adopting the Gamma distribution for the errors for estimation purposes. In principle, any distribution defined on a positive support can be adopted;1 nevertheless, the Gamma distribution, given its Quasi Maximum Likelihood properties for the conditional mean equation parameters, is widely used in empirical works.

The fully robust (FR) estimator remains consistent under general forms of heteroskedasticity of the errors. p. 0
Where the paper says it

Additionally, we provide analytic formulas for robust covariance estimators that remain valid under distributional misspecification, which are also referred to as the asymptotic variance of the Quasi-Maximum Likelihood Estimator. We also propose a covariance estimator that is consistent under general forms of heteroskedasticity of the error term, while the usually adopted estimators rely on a specific variance assumption.

The simulation study evaluates several covariance matrix estimators, such as Hessian (H), outer product of gradients (OPG), information matrix (IM), sandwich (S), GLM, and fully robust (FR) across five cases. Empirical results show that under correct distributional specification, the mean of the standard errors based on the proposed estimators converges as T increases. p. 2

Where the paper says it

Our simulation results show that the usually adopted estimator of the asymptotic covariance matrix, based on a specific conditional variance assumption, is not consistent when, as expected, the latter does not hold, while the estimator we propose is consistent general forms of heteroskedasticity of the error term. Moreover, the covariance assumption has implications for estimating the asymptotic covariance matrix of the estimated parameters, while it is not required to get consistent estimates of the conditional mean equation parameters; this is something that has not been emphasized in the MEM literature.

When only the distributional assumption is misspecified, H, OPG, and IM estimators suffer from bias and tend to slightly underestimate standard errors. p. 2
Where the paper says it

Moreover, the covariance assumption has implications for estimating the asymptotic covariance matrix of the estimated parameters, while it is not required to get consistent estimates of the conditional mean equation parameters; this is something that has not been emphasized in the MEM literature. As a consequence, as underlined in the real case study, we suggest reporting both estimated standard errors for comparison.

The Sandwich estimator is consistent when the MLE of ϕ2’s cancel out. p. 0
Where the paper says it

Additionally, we provide analytic formulas for robust covariance estimators that remain valid under distributional misspecification, which are also referred to as the asymptotic variance of the Quasi-Maximum Likelihood Estimator. We also propose a covariance estimator that is consistent under general forms of heteroskedasticity of the error term, while the usually adopted estimators rely on a specific variance assumption.

The FR estimator often exhibits lower variability than the sandwich estimator for dynamic mean parameters. p. 2
Where the paper says it

Our simulation results show that the usually adopted estimator of the asymptotic covariance matrix, based on a specific conditional variance assumption, is not consistent when, as expected, the latter does not hold, while the estimator we propose is consistent general forms of heteroskedasticity of the error term. Moreover, the covariance assumption has implications for estimating the asymptotic covariance matrix of the estimated parameters, while it is not required to get consistent estimates of the conditional mean equation parameters; this is something that has not been emphasized in the MEM literature.

Concluding remarks

The authors provide analytical derivatives up to the second order and illustrate an iterative algorithm for MLE computation. The iterative algorithm can be extended to compute MLEs for vector MEM (vMEM) and a time-varying dispersion parameter is accommodated in Appendix B. p. 2

Where the paper says it

The paper is structured as follows: Section 2 provides a brief description of the MEM and its extensions, while Section 3 illustrates the MLEs of the model parameters (Section 3. 1), the iterative algorithm adopted to compute the MLE (Section 3.

The authors report that generative AI was not used in their research or preparation of this manuscript. The paper is 2611.16577. EVERY sentence must end with a reference token. BARE, with no backticks around it reads as a reference and not as a code span. N is the page number counting from 1 — the same N as that page's --- Page N --- line. The sentence's own full stop stays AFTER the closing >; a token with nothing after the comma is wrong. REFERENCE RULES - Copy a token exactly as it is printed for you; Never invent an id, never guess a page or a paragraph, and never change one; A token the work does not hold is worse than no token. Put the token at the END of the sentence it supports; point at the passage where the claim came from; not on page 1, and not on page you remember. ANCHOR THE WORDS; The words go INSIDE, in the same brackets; after the location, a comma with no space, then five or more words copied word for word from that passage the claim rests on; No quotation marks around them and shorter than five words copied word for word loses the anchor; Write BARE, with no backticks and no quotation marks around them. The words go inside, so the token ends,five or more words copied word for word> and the sentence's own full stop follows the closing >. OPEN with one short orienting paragraph that says what the paper is and why it matters. This paper provides analytical derivatives and an iterative algorithm to compute Maximum Likelihood Estimates (MLEs) for parameters in a Multiplicative Error Model, which is important because closed-form solutions are not available and numerical derivatives can be unstable.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Analytical derivatives for Multiplicative Error Model log-likelihood provide more accurate estimates than numerical approximations, leading to improved accuracy in Maximum Likelihood Estimates (MLEs).

  2. The iterative algorithm based on analytical derivatives should be computationally implemented, allowing for readily programmable computation of MLEs, unlike most software packages relying on numerical approximations to derivatives.

  3. Explicit formulas for several estimators of the asymptotic covariance matrix are provided, including analytic formulas for robust covariance estimators that remain valid under distributional misspecification, which are referred to as the asymptotic variance of the Quasi-Maximum Likelihood Estimator.

  4. A covariance estimator is proposed that is consistent under general forms of heteroskedasticity of the error term, unlike usually adopted estimators that rely on a specific variance assumption.

  5. The iterative algorithm can be extended to compute MLEs for vector Multiplicative Error Models (vMEM), taking into account contemporaneous correlations among error terms and lagged cross-effects.

  6. Simulation results show that under correct distributional specification, the mean of the standard errors based on the proposed estimators of the covariance matrix converges as T increases.

  7. When only distributional misspecification is present, the Hessian, Outer Product of Gradients (OPG), and Information Matrix (IM) estimators suffer from bias because they depend on the MLE of phi squared.

  8. The sandwich estimator is consistent even when only distributional misspecification is present, as the MLE of phi squared's cancel out in its formulation.

  9. The Fully Robust (FR) estimator does not require the GLM assumption, Var(xtIt-1) = phi squared µ2t, to be consistent under general forms of heteroskedasticity of the errors.

Abstract

Maximum Likelihood Estimates (MLEs) of the parameters of a Multiplicative Error Model are typically computed through a standard gradient method, an iterative procedure usually based on numerical approximations to derivatives of the log-likelihood function, owing to the complexity of the analytical derivatives, which involve recursive terms. Building on previous work developed mainly in the GARCH framework, we derive analytical expressions for the score vector and the Hessian matrix of the Multiplicative Error Model log-likelihood by adopting the Gamma distribution for estimation purposes. Then, we illustrate how the iterative algorithm based on the analytical derivatives should be computationally implemented. After maximizing the log-likelihood, we also provide explicit formulas for several estimators of the asymptotic covariance matrix of the MLE. Additionally, we provide analytic formulas for robust covariance estimators that remain valid under distributional misspecification, which are also referred to as the asymptotic variance of the Quasi-Maximum Likelihood Estimator. We also propose a covariance estimator that is consistent under general forms of heteroskedasticity of the error term, while the usually adopted estimators rely on a specific variance assumption. The finite-sample performance of the covariance estimators is assessed through a Monte Carlo investigation. An application on real data completes the analysis.

Sources

Related papers