Constrained mixtures of generalized normal distributions

Alfred Kume; Pierdomenico Duttilo; Stefano Antonio Gattone

REVIEW 5 major objections 5 minor 34 references

Constrained mixtures of generalized normal distributions

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Imposing equality constraints on location, scale, or shape parameters across subsets of generalized-normal mixture components improves parameter estimation when the constraints are correct, and the resulting CMGND models beat constrained…

desk verdict A real but incremental extension: subset-level equality constraints for generalized normal mixtures, with a working R package and plausible simulation evidence; send it to a serious referee, but ask for convergence checks and a cleaner manuscript. read the letter →

arxiv 2506.03285 v1 pith:RY4EP4YA submitted 2025-06-03 stat.ME stat.CO

Pierdomenico Duttilo , Stefano Antonio Gattone , Alfred Kume This is my paper

classification stat.MEstat.CO

keywords generalizednormaldistributionmixturemodelsequalityconstraintsECMalgorithmNewton-RaphsonstockindexreturnskurtosisBayesianinformationcriterion

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a family of univariate mixtures of generalized normal distributions in which the location, scale, and shape parameters may be forced to be equal across any chosen subset of components, giving a six-model family for two-component mixtures. The central claim is that when these equality constraints match the true data-generating process, maximum-likelihood estimation via a constrained ECM algorithm yields lower root mean square error than the unconstrained mixture, with the largest gains when components overlap heavily. A second claim is that constrained generalized-normal mixtures fit daily Euro Stoxx 50 stock returns better than constrained normal or Student-t mixtures by the Bayesian information criterion, and that the shape constraint is especially effective at capturing marginal kurtosis. If true, the approach offers a way to gain parsimony and interpretability without sacrificing distributional flexibility, which matters for financial return data with heavy tails and for mixture modelling generally.

What carries the argument

The engine of the paper is a constrained expectation conditional maximisation (ECM) algorithm built on Newton-Raphson updates. For each constrained parameter group r, the update for location μ_r, scale σ_r, and shape ν_r is obtained by setting the derivative of the Q-function with respect to the shared parameter to zero, summing contributions from all components in the constrained subset; the resulting nonlinear equations are solved by Newton-Raphson steps (Eqs. (6) and (7)), while the shape update (Eq. (8)) includes an adaptive step size α(ν_r)=$e^{{-ν_r}}$ to dampen instability. The generalized normal density f(x|μ,σ,ν)=ν/(2σΓ(1/ν)) exp(-|(x-μ)/σ|^ν) is the working object, with ν=2 giving the normal, ν=1 the Laplace, and extremes approaching uniform or impulsive limits. The machinery's key work is to turn a hypothesis about shared tail behaviour or shared scale between components into a smaller parameter space whose estimates are then compared across models by BIC.

What would settle it

Run the constrained ECM on simulated data in which the imposed constraints are deliberately wrong (for example, forcing equal shapes when the true shapes differ) with random k-means starts; if the log-likelihood does not increase monotonically, or if the BIC-selected constrained model produces higher RMSE for the marginal mean, variance, skewness, and kurtosis than the unconstrained fit, the central claim that constraints improve estimation would fail.

Watch

Extended reading notes

Core claim

The paper's central discovery is that equality constraints on component parameters can act as a stabiliser and an efficiency booster, not just a parsimony device. In simulations with K=3 components, fixing the scale or shape of two components to be equal reduced RMSE for the constrained parameters—for example, the constrained scale σ₂ fell from 0.51 to 0.20 with N=400 in the low-overlap UCU model—and the benefit often spilled over to unconstrained parameters in the same component because estimates are interdependent. Under high overlap and small samples, the most constrained model (UCC) produced more accurate estimates for nearly every parameter, including parameters in unconstrained components. On real data, the BIC selected the two-component CMGND model for 72% of the 50 Euro Stoxx 50 constituents, with the model imposing a common mean and common scale (CCU) the most frequent choice, outperforming constrained normal and Student-t mixtures. The paper presents this as evidence that constraints on arbitrary subsets of components, including location constraints unavailable in standard packages, can reduce complexity without compromising fit.

Load-bearing premise

Every RMSE and BIC advantage reported here assumes the constrained ECM algorithm, with its Newton-Raphson updates and adaptive step size, actually converges to a genuine local maximum of the constrained log-likelihood; the paper gives no convergence or monotonicity theorem, and the simulations test only correctly specified constraints, leaving misspecified or badly initialised fits untested.

Editorial extensions

If this is right

Correctly specified constraints reduce estimation error most when mixture components overlap substantially, so constrained fits are most valuable in noisy, hard-to-separate settings.
The BIC will often prefer the most parsimonious constrained model (UCC) even when the true model is unconstrained, especially with small samples and high overlap, so model selection must be read as a trade-off rather than a truth detector.
Imposing a common shape parameter across components is a practical way to model marginal kurtosis, because the BIC-selected constrained models matched or beat the true model's moment estimates, particularly for kurtosis.
Constrained generalized-normal mixtures can beat constrained Student-t mixtures on financial return data, suggesting that the shape parameter's flexible tail control is competitive with the t distribution's degrees of freedom.
The ability to constrain location parameters—not offered by standard normal or t mixture packages—lets a model express a common mean across regimes, which the paper exploits with the CCU model for stock returns.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the reported gains suggest the constraints act as a regulariser, so one testable extension is whether CMGND with misspecified constraints still yields lower predictive loss on holdout data than the unconstrained model; if so, constraints are buying stability rather than only efficiency.
Editorial inference: the CCU model's interpretation—one stable component and one turmoil component distinguished mainly by tail shape—implies a time-varying version of the shape parameter could connect these mixtures to score-driven volatility models, an extension the paper itself mentions.
Editorial inference: because the shape parameter controls kurtosis directly, the UUC and UCC constraints give a natural way to compare whether different market regimes share tail heaviness; a direct test would compare BIC between UCC and UUC on subsets of stocks sorted by volatility regimes.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The genuinely new thing here is the constraint scheme: unlike mclust or teigen, which constrain all components at once, this paper lets you impose equality constraints on location, scale, and shape across any subset of components in a generalized normal mixture. The update equations (6)–(8) are derived correctly from the Q-function, and they reduce to known MGND updates when the constraints are removed. The R package cmgnd is available, which counts for something. The simulation study is honest and reasonably thorough: multiple overlap levels, two sample sizes, and the RMSE reductions are consistent and go in the direction you would expect. The BIC model selection section is also candid about the tendency of BIC to prefer parsimonious constrained models even when the true model is unconstrained.

The soft spots are real but not fatal. Most important: no convergence guarantee for the constrained ECM. The mu and sigma updates are plain Newton–Raphson steps with no line search, the shape update uses a heuristic step size, and for nu < 1 the score function has singularities. The paper never checks whether the algorithm actually increases the Q-function or compares against a generic optimizer. That leaves a gap, but the simulation evidence is broad enough to suggest the algorithm is not silently stalling in those settings. A referee should ask for empirical convergence diagnostics or a partial theoretical statement, not reject the paper.

Two other issues. First, the real-data comparison is not apples-to-apples: CMGND can constrain the mean, while mclust and teigen cannot, so part of the BIC advantage comes from a larger model space. The paper admits this, but the headline claim that CMGND outperforms is stronger than the benchmark allows. Second, the manuscript has avoidable artifacts: Eq. (10) defines log-returns as ln(P_t - P_{t-1}) instead of ln(P_t / P_{t-1}), and there is a leftover authorial comment after Eq. (6): “I removed the further derivation of eq. (6) see the commented code.” These are easy fixes but they make the paper look rushed.

Who is this for? Mixture modelers and applied statisticians working with fat-tailed or heavy-tailed data. It is not a landmark, but it is a useful, clearly presented extension with working software. I would send it to peer review and ask for a revision that (a) provides convergence diagnostics or a monotonicity argument, (b) makes the real-data comparison symmetric by adding a constrained-normal or constrained-t model that can also share means, and (c) cleans up the text.

Referee Report

5 major / 5 minor

Summary. The paper introduces a family of univariate constrained mixtures of generalized normal distributions (CMGND) in which location, scale, and shape parameters may be equated across arbitrary subsets of mixture components. An ECM algorithm with Newton-Raphson updates is proposed (Eqs. 5–8), and the method is evaluated through simulations comparing RMSE of constrained versus unconstrained fits under known constraints, through BIC-based model selection over a set of constrained models, and through a real-data application to daily returns of the 50 Euro Stoxx 50 constituents, comparing CMGND to constrained normal and Student-t mixtures. The central claims are that imposing correct constraints improves parameter estimation, especially under component overlap, and that CMGND offers competitive or superior performance on financial data.

Significance. The constrained parametrization is a useful and apparently novel extension of MGND, and the paper connects it to an existing R package (cmgnd), which is a concrete reproducibility asset. The simulation design, in which data are generated under a known constraint structure and then fitted under the same structure, demonstrates a parametric-efficiency effect that is plausible and directionally expected. The BIC selection study is thorough in varying sample size and overlap and reveals realistic limits of model selection in overlapping mixtures. However, the lack of a convergence or monotonicity guarantee for the proposed ECM updates is a serious gap, because all empirical results are reported as properties of the model class while being computed by an unverified algorithm. In addition, the real-data analysis compares CMGND only to constrained normal and Student-t mixtures, not to the unconstrained MGND, so it does not isolate the effect of the constraints themselves.

major comments (5)

[Section 2.1, Eq. (4)] The Q-function is misstated. The expression Q(θ,θ^{(m-1)}) = Σ_n log( Σ_k z^{(m-1)}_{nk} π^{(m-1)}_k f_k(x_n|θ^{(m-1)}) ) is not the standard EM Q-function, and it is not the function whose derivatives lead to the updates in Eqs. (5)–(8). The standard form is Q(θ,θ^{(m-1)}) = Σ_n Σ_k z^{(m-1)}_{nk} log(π_k f_k(x_n;θ)). As written, Eq. (4) cannot generate the given mixture-weight update or the conditional maximization equations. Please correct Eq. (4) and re-derive the updates from the correct Q-function.
[Section 2.1, Eqs. (6)–(8)] No convergence or ascent guarantee is provided for the constrained ECM algorithm. The Newton-Raphson updates for μ_r and σ_r are not safeguarded: there is no line search, no positivity enforcement for σ, and the shape update uses the heuristic step α(ν)=e^{-ν} together with a stopping rule that can skip shape updates. Because Newton-Raphson steps do not automatically increase the Q-function, the usual EM/ECM monotonicity argument does not apply, so the algorithm may terminate at non-stationary or spurious points. All RMSE and BIC results in Sections 3 and 4 are computed from this algorithm, so the empirical comparisons do not necessarily describe the CMGND model class. Please provide a monotonicity/convergence theorem or replace the updates with safeguarded versions, and report convergence diagnostics (e.g., fraction of runs reaching a stationary point, iteration counts, and sensitivity of results to the shape-update skip threshold).
[Section 4, Table A2] The real-data analysis does not support the claim that constraints are beneficial. The comparison is only between CMGND, CMND, and CMSTD; since CMGND nests CMND (ν=2) and has a more flexible tail-shape parameter than CMSTD, the BIC wins reported for CMGND can be attributed to the added flexibility of the GND family rather than to the constraint structure. To demonstrate that constraints help in practice, the paper should report BIC for the unconstrained MGND on the same data, or provide a real-data comparison in which the constraint structure is systematically varied while the distribution family is fixed.
[Section 2.1, after Eq. (6)] The manuscript contains an editorial artifact: the sentence 'I removed the further derivation of eq. (6) see the commented code' appears verbatim in the text. This must be removed, and the derivation of Eq. (6), including the expression for the second derivative g′(μ_r), should be provided in the paper so that the update is self-contained and reproducible.
[Section 2.1, Eq. (6)] For ν_k < 1, the derivative terms in the location update contain (x_n − μ_r)^{ν_k−2}, which is singular when x_n = μ_r. This case arises in the simulations, where ν_1 = 0.5. The paper does not state how ties are handled numerically or why the Newton-Raphson step remains well-defined at such points. Please clarify the treatment of x_n = μ_r and, if necessary, the conditions under which g′(μ_r) is finite.

minor comments (5)

[Section 4, Eq. (10)] The return definition in Eq. (10) is written as r_t = ln(P_t − P_{t−1})100, which is not a log-return. It should be r_t = ln(P_t/P_{t−1})×100.
[Section 3.3, Table 6] Table 6 is difficult to read: the scenario labels in the first column are ambiguous (e.g., 'Low UUU 400' and 'Low UUU 1000' are not clearly separated), and the table would benefit from clearer row groupings and a note describing which model corresponds to the BIC-selected model in each block.
[Section 3.1] The simulation study reports RMSE point estimates from 250 replications but no standard errors or confidence intervals; adding these would help assess whether the reported RMSE differences are statistically reliable, particularly where the gaps are small.
[Section 4] The text states that 'All algorithms employ 5 starting points with k-means initialisation' for CMND and CMSTD, but it does not explicitly state whether the same initialization scheme is used for the CMGND model; please clarify.
[Section 2.1] The stopping rule from Duttilo and Gattone (2025), which skips shape updates when the derivative is below a threshold, is referenced but not described; a brief description is needed for the paper to be self-contained.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the constrained ECM updates are derived from the Q-function, the simulation is a favorable validation rather than a forced prediction, and the real-data BIC comparisons use external benchmarks.

full rationale

I walked the derivation chain from the log-likelihood (Eq. 3) and Q-function (Eq. 4) through the constrained update equations (Eqs. 6-8). The updates are derived by differentiating the Q-function with respect to the constrained location, scale, and shape parameters and then applying Newton-Raphson steps; they do not assume the values they are used to estimate. The simulation studies generate data from CMGND models and then fit the same constrained models, so the lower RMSE is a validation of the estimator in a favorable, correctly-specified setting. This is a simulation design choice, not a 'prediction' forced by definition: the paper does not rename fitted values as out-of-sample predictions. The real-data comparison is externally grounded: CMGND is benchmarked against constrained normal mixtures fitted with the mclust package and constrained Student-t mixtures fitted with the teigen package, with per-stock BIC values reported in Table A2. The paper does rely on the authors' earlier ECM variant (Duttilo and Gattone 2025) and on their adaptive step size (Duttilo et al. 2024b), but these are algorithmic building blocks accompanied by code, not a uniqueness theorem used to forbid alternatives, and the novel constrained updates in Eqs. (6)-(8) are derived in this paper. A manuscript artifact in Section 2.1, 'I removed the further derivation of eq. (6) see the commented code', flags an omitted derivation; an omission is a completeness or correctness concern, not circularity. Similarly, the absence of a convergence or monotonicity theorem for the Newton-Raphson updates is a correctness risk, not a circularity issue, because the update derivation does not assume the target result. The score reflects only the minor, non-load-bearing self-citations for the inherited ECM machinery.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard mixture-modeling assumptions plus the unproved convergence of the constrained ECM algorithm. No new theoretical entity is introduced; the model family combines the existing generalized normal distribution with equality-constraint machinery. The adaptive step size and stopping rules are inherited from prior work and are not derived in this paper.

free parameters (3)

Adaptive step size alpha(nu)=exp(-nu) in shape update (Eq. 8) = exp(-nu), not estimated
Design choice from Duttilo et al. (2024b) to stabilise Newton-Raphson steps for the shape parameter; the exponential form is not derived from the likelihood.
Shape-update skip threshold in ECMs stopping rule = not reported in this paper
The ECMs algorithm inherited from Duttilo and Gattone (2025) skips shape updates when the first derivative of the Q-function falls below a threshold; the threshold value is not stated here, which affects reproducibility.
Number of starting points and initialization scheme = 5 starting points, k-means
Used in Section 4 for all algorithms; the choice can influence whether the reported BIC values correspond to global or local maxima.

assumptions (4)

domain assumption The constrained ECM iterations converge to a local maximum of the constrained log-likelihood; no monotonicity or convergence theorem is provided.
Invoked in Section 2.1 for all parameter updates and in Sections 3 and 4 for the validity of RMSE and BIC comparisons.
domain assumption The partition of components into constrained and unconstrained subsets is fixed before estimation and is correctly specified in the simulation study.
The simulation designs in Section 3.1 generate data from UCU, UUC, and UCC models and fit those same constraints; the practical case where constraints are wrong is handled only informally via BIC.
domain assumption BIC is a suitable criterion for selecting among these non-nested mixture models.
Used throughout Section 3.2 and Section 4; no check of BIC regularity for non-identifiable or boundary parameter cases is reported.
domain assumption The constrained MGND parameter space remains identifiable and the likelihood bounded under the imposed constraints.
Discussed only in passing in the introduction and Section 3.2; there is no formal identifiability analysis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Constrained mixtures of generalized normal distributions." pith.science (2026). https://pith.science/paper/RY4EP4YA

@misc{pith2026250603285,
  author       = {Pith},
  title        = {Pith review of: Constrained mixtures of generalized normal distributions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RY4EP4YA}},
  note         = {Machine review of arXiv:2506.03285}
}

read the original abstract

This work introduces a family of univariate constrained mixtures of generalized normal distributions (CMGND) where the location, scale, and shape parameters can be constrained to be equal across any subset of mixture components. An expectation conditional maximisation (ECM) algorithm with Newton-Raphson updates is used to estimate the model parameters under the constraints. Simulation studies demonstrate that imposing correct constraints leads to more accurate parameter estimation compared to unconstrained mixtures, especially when components substantially overlap. Constrained models also exhibit competitive performance in capturing key characteristics of the marginal distribution, such as kurtosis. On a real dataset of daily stock index returns, CMGND models outperform constrained mixtures of normals and Student's t distributions based on the BIC criterion, highlighting their flexibility in modelling nonnormal features. The proposed constrained approach enhances interpretability and can improve parametric efficiency without compromising distributional flexibility for complex data.

Figures

Figures reproduced from arXiv: 2506.03285 by the authors.

**Figure 1.** The GND densities for µ = 1, σ = 1 and different shape values. A finite mixture of GND with K components is given by the marginal distribution of the random variable X F(x|θ) = X K k=1 πk fk (x|µk ,σk ,νk ) (2) where fk is defined as in Eq.(1), and the set of all mixture parameters is given by θ = {πk ,µk ,σk , νk , k = 1, ..., K} belonging to the parameter space Θ = {θ : 0 < πk < 1,PK k=1 πk = 1,µk ∈ R,σk > 0,νk > … view at source ↗

**Figure 2.** Simulated MGND and CMGND models: low overlap scenario. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Simulated MGND and CMGND models: medium overlap scenario. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: Simulated MGND and CMGND models: high overlap scenario. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 5.** Figure 5: Proportion of times models selected by BIC: low overlap scenario. Sample size [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

**Figure 6.** Figure 6: Proportion of times models selected by BIC: low overlap scenario. Sample size [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

**Figure 7.** Figure 7: Proportion of times models selected by BIC: medium overlap scenario. Sample size [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]

**Figure 8.** Figure 8: Proportion of times models selected by BIC: medium overlap scenario. Sample size [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]

**Figure 9.** Figure 9: Proportion of times models selected by BIC: high overlap scenario. Sample size [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]

**Figure 10.** Figure 10: Proportion of times models selected by BIC: high overlap scenario. Sample size [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]

**Figure 11.** Figure 11: Proportion of times models selected by BIC for both CMGND and CMSTD models. [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]

**Figure 12.** Figure 12: Log-returns (%) of ABI.BE panel a, and estimated densities panel b. [PITH_FULL_IMAGE:figures/full_fig_p027_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

[1]

Allili, M. S. (2012). Wavelet modeling using finite mixtures of generalized gaussian distributions: Application to texture discrimination and retrieval. IEEE Transactions on Image Processing , 21(4):1452--1464

work page 2012
[2]

S., Bouguila, N., and Ziou, D

Allili, M. S., Bouguila, N., and Ziou, D. (2008). Finite general Gaussian mixture modeling and application to image and video foreground segmentation . Journal of Electronic Imaging , 17(1):013005

work page 2008
[3]

L., McNicholas, P

Andrews, J. L., McNicholas, P. D., and Subedi, S. (2011). Model-based classification via mixtures of multivariate t-distributions. Computational Statistics & Data Analysis , 55(1):520--529

work page 2011
[4]

L., Wickins, J

Andrews, J. L., Wickins, J. R., Boers, N. M., and McNicholas, P. D. (2018). teigen: An r package for model-based clustering and classification via the multivariate t distribution. Journal of Statistical Software , 83(7):1--32

work page 2018
[5]

Banfield, J. D. and Raftery, A. E. (1993). Model-based gaussian and non-gaussian clustering. Biometrics , 49(3):803--821

work page 1993
[6]

Bazi, Y., Bruzzone, L., and Melgani, F. (2006). Image thresholding based on the em algorithm and the generalized gaussian distribution. Pattern Recognition , 40(2):619--634

work page 2006
[7]

and P\" o tter, U

Behr, A. and P\" o tter, U. (2009). Alternatives to the normal model of stock returns: Gaussian mixture, generalised logf and generalised hyperbolic models. Annals of Finance , 5:49--68

work page 2009
[8]

and Chr\'etien, S

Biernacki, C. and Chr\'etien, S. (2003). Degeneracy in the maximum likelihood estimation of univariate gaussian mixtures with em. Statistics & Probability Letters , 61(4):373--382

work page 2003

Show all 34 references

[9]

Bishop, C. M. (2006). Pattern Recognition and Machine Learning (Information Science and Statistics) . Springer-Verlag, Berlin, Heidelberg

2006
[10]

and Govaert, G

Celeux, G. and Govaert, G. (1995). Gaussian parsimonious clustering models. Pattern Recognition , 28(5):781--793

1995
[11]

and Hunter, D

Chauveau, D. and Hunter, D. R. (2013). ECM and MM algorithms for normal mixtures with constrained parameters . working paper or preprint

2013
[12]

Deledalle, C.-A., Parameswaran, S., and Nguyen, T. Q. (2018). Image denoising with generalized gaussian mixture model patch priors. SIAM Journal on Imaging Sciences , 11(4):2568--2609

2018
[13]

and Gattone, S

Duttilo, P. and Gattone, S. A. (2025). Enhancing parameter estimation in finite mixture of generalized normal distributions. Computational Statistics , pages 1--28. DOI:10.1007/s00180-025-01638-x

2025 doi
[14]

A., and Iannone, B

Duttilo, P., Gattone, S. A., and Iannone, B. (2024a). Mixtures of generalized normal distributions and egarch models to analyse returns and volatility of esg and traditional investments. AStA Advances in Statistical Analysis , 108(4):755--775

2024
[15]

A., and Kume, A

Duttilo, P., Gattone, S. A., and Kume, A. (2024b). cmgnd: constrained mixture of generalized normal distributions. R package . https://github.com/pierdutt/cmgnd

2024
[16]

Dytso, A., Bustin, R., Poor, H., and Shamai, S. (2018). Analytical properties of generalized gaussian distributions. Journal of Statistical Distributions and Applications , 5(6):1--40

2018
[17]

C., editor (2013)

Harvey, A. C., editor (2013). Dynamic Models for Volatility and Heavy Tails: With Applications to Financial and Economic Time Series . Cambridge University Press

2013
[18]

Hathaway, R. J. (1986). A constrained em algorithm for univariate normal mixtures. Journal of Statistical Computation and Simulation , 23(3):211--230

1986
[19]

Iannone, B., Duttilo, P., and Gattone, S. A. (2025). Evaluating the resilience of esg investments in european markets during turmoil periods. Corporate Social Responsibility and Environmental Management , n/a(n/a)

2025
[20]

Kon, S. J. (1984). Models of stock returns--a comparison. The Journal of Finance , 39(1):147--165

1984
[21]

and Mclachlan, G

Lee, S. and Mclachlan, G. (2013). On mixtures of skew normal and skew t-distributions. Advances in Data Analysis and Classification , 7

2013
[22]

LSEG Workspace

LSEG Data & Analytics (2025). LSEG Workspace

2025
[23]

and Ramos, A

Massing, T. and Ramos, A. (2021). Student's t mixture models for stock indices. A comparative study . Physica A: Statistical Mechanics and its Applications , 580(C)

2021
[24]

and Peel, D

McLachlan, G. and Peel, D. (2000). Finite Mixture Models . Wiley Series in Probability and Statistics. John Wiley & Sons, New York

2000
[25]

J., Lee, S

McLachlan, G. J., Lee, S. X., and Rathnayake, S. I. (2019). Finite mixture models. Annual Review of Statistics and Its Application , 6(1):355--378

2019
[26]

idane-Sa\

Mohamed, O. M. M. and Ja\"idane-Sa\"idane, M. (2009). Generalized gaussian mixture model. In 2009 17th European Signal Processing Conference , pages 2273--2277

2009
[27]

Nadarajah, S. (2005). A generalized normal distribution. Journal of Applied Statistics , 32(7):685--694

2005
[28]

M., Jonathan Wu , Q., and Zhang, H

Nguyen, T. M., Jonathan Wu , Q., and Zhang, H. (2014). Bounded generalized gaussian mixture model. Pattern Recognition , 47(9):3132--3142

2014
[29]

Pearson, K. (1894). Contributions to the mathematical theory of evolution. Philosophical Transactions of the Royal Society of London. A , 185:71--110
[30]

Quandt, R. E. and Ramsey, J. B. (1978). Estimating mixtures of normal distributions and switching regressions. Journal of the American Statistical Association , 73(364):730--738

1978
[31]

A., and Di Mari, R

Rocci, R., Gattone, S. A., and Di Mari, R. (2018). A data driven equivariant approach to constrained gaussian mixture modeling. Advances in Data Analysis and Classification , 12(2):235--260

2018
[32]

A., Lukin, V

Roenko, A. A., Lukin, V. V., Djurovic, I., and Simeunovi\'c, M. (2014). Estimation of parameters for generalized gaussian distribution. In 2014 6th International Symposium on Communications, Control and Signal Processing (ISCCSP) , pages 376--379

2014
[33]

B., and Raftery, A

Scrucca, L., Fraley, C., Murphy, T. B., and Raftery, A. E. (2023). Model-Based Clustering, Classification, and Density Estimation Using mclust in R . Chapman and Hall/CRC

2023
[34]

Wen, L., Qiu, Y., Wang, M., Yin, J., and Chen, P. (2020). Numerical characteristics and parameter estimation of finite mixed generalized normal distribution. Communications in Statistics - Simulation and Computation , 0(0):1--25

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Allili, M. S. (2012). Wavelet modeling using finite mixtures of generalized gaussian distributions: Application to texture discrimination and retrieval. IEEE Transactions on Image Processing , 21(4):1452--1464

work page 2012

[2] [2]

S., Bouguila, N., and Ziou, D

Allili, M. S., Bouguila, N., and Ziou, D. (2008). Finite general Gaussian mixture modeling and application to image and video foreground segmentation . Journal of Electronic Imaging , 17(1):013005

work page 2008

[3] [3]

L., McNicholas, P

Andrews, J. L., McNicholas, P. D., and Subedi, S. (2011). Model-based classification via mixtures of multivariate t-distributions. Computational Statistics & Data Analysis , 55(1):520--529

work page 2011

[4] [4]

L., Wickins, J

Andrews, J. L., Wickins, J. R., Boers, N. M., and McNicholas, P. D. (2018). teigen: An r package for model-based clustering and classification via the multivariate t distribution. Journal of Statistical Software , 83(7):1--32

work page 2018

[5] [5]

Banfield, J. D. and Raftery, A. E. (1993). Model-based gaussian and non-gaussian clustering. Biometrics , 49(3):803--821

work page 1993

[6] [6]

Bazi, Y., Bruzzone, L., and Melgani, F. (2006). Image thresholding based on the em algorithm and the generalized gaussian distribution. Pattern Recognition , 40(2):619--634

work page 2006

[7] [7]

and P\" o tter, U

Behr, A. and P\" o tter, U. (2009). Alternatives to the normal model of stock returns: Gaussian mixture, generalised logf and generalised hyperbolic models. Annals of Finance , 5:49--68

work page 2009

[8] [8]

and Chr\'etien, S

Biernacki, C. and Chr\'etien, S. (2003). Degeneracy in the maximum likelihood estimation of univariate gaussian mixtures with em. Statistics & Probability Letters , 61(4):373--382

work page 2003