Vincent Zoonekynd's Blog

Sun, 09 Jan 2022: 2021 in Machine Learning

As every year, here are some of the papers I read this year, covering topics such as copulas, convex and non-convex optimization, reinforcement learning, Shapley score, loss landscape, time series with missing data, neural ODEs, graphs, monotonic neural nets, linear algebra, variance matrices, etc.

If you want a shorter reading list:

Reinforcement learning
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3554486

ADMM
https://arxiv.org/abs/1909.10233

Copulas
https://link.springer.com/book/10.1007/978-3-030-13785-4

If you want a longer list:

http://zoonek.free.fr/Ecrits/articles.pdf

2022-01-09_2021_in_ML.png

Copulas

Vine copulas

Copulas measure the dependence between random variables, but, contrary to correlation, they do not depend on the marginal distributions, and can describe arbitrary dependence structures.

Pairwise copulas are already widely used, and many models are available (Gaussian, Student, Gumbel, Clayton, etc.), but their generalizations to higher dimensions are not flexible enough: the Gaussian and Student copulas are only parametrized by a correlation matrix, and the Archimedean copulas are exchangeable.

It is however possible to combine pairwise copulas to build more complex models. The two main constructions are hierarchical Archimedean copulas (HAC), and vine copulas (aka the pair copula construction, PCC).

Vine copulas are surprisingly easy to interpret; R and Python software is readily available to fit them.

Analyzing Dependent Data with Vine Copulas
C. Czado (2019)
https://link.springer.com/book/10.1007/978-3-030-13785-4

Simulating copulas: stochastic models, sampling algorithms and applications
J.F. Mai and M. Scherer (2017)
https://www.worldscientific.com/worldscibooks/10.1142/10265

Applications of copulas

Applications include time series models (AR-like models, but using copulas to measure the dependence over time), factor models, risk measurement, etc.

https://www.mdpi.com/journal/econometrics/special_issues/copula_models

COPAR: multivariate time series modeling using copulas autoregressive model
E.C. Brechmann amd C. Czado (2012)
https://arxiv.org/abs/1203.3328

Dynamic copula methods in finance
U. Cherubini et al. (2012)
https://www.wiley.com/en-sg/Dynamic+Copula+Methods+in+Finance-p-9780470683071

We start to see applications of vine copulas in finance, but they often boil down to "here is the minimum spanning tree estimated from the rank correlation instead of the linear correlation" (they tend to use Kendall's τ instead of the rank correlation, and the edges of the MST are labeled with pairwise copulas instead of just correlations).

Risk management with high-dimensional vine copulas: an analysis of the Euro Stoxx 50
E.C. Brechmann and C. Czado (2013)
https://mediatum.ub.tum.de/doc/1079276/1079276.pdf

ESG, risk and (tail) dependence
R. Bax  et al. (2021)
https://arxiv.org/abs/2105.07248

Copulas and neural networks

If the parametric copulas are not enough, you can always use neural networks.

Implicit generative copulas
T. Janke et al. (2021)
https://arxiv.org/abs/2109.14567
Copulas as high-dimensional generative models: vine copula autoencoders
N. Tagasovska et al. (2019)
https://arxiv.org/abs/1906.05423
 

Local Gaussian correlation

We sometimes have the impression that correlation is not constant, and becomes larger for extreme observations.

This is often measured with the tail dependence coefficient (TDC): the probability that one variable has an extreme value given that the other has.

The local Gaussian correlation goes further, and lets the correlation vary depending on where you are in the distribution: it is not limited to the tails and can show, for instance, a different behaviour for positive and negative values. It is estimated by locally fitting non-centered Gaussian distributions to the data.

Introducing localgauss, an R package for estimating and visualizing local Gaussian correlation
C.D. Berensten et al. (2014)
https://www.jstatsoft.org/article/view/v056i12

Recognizing and visualizing copulas: an approach using local Gaussian approximation
G.D. Berentsen et al. (2014)
https://fam.tuwien.ac.at/events/eaj2014/c/stove_bard.pdf

Optimization

Convex optimization

Convex, constrained optimization has become very easy, thanks to packages such as CVXR, cvxpy, ROI. The following papers present many examples, showing how useful those problems are.

CVXR an R package for disciplined convex optimization
A. Fu et al. (2020)
https://stanford.edu/~boyd/papers/cvxr_paper.html

ROI: an extensible R optimization infrastructure
S. Theußl et al. (2020)
https://www.jstatsoft.org/article/view/v094i15

Those packages transform your optimization problem to a form amenable to solvers. This is not limited to simple mathematical operations (matrix multiplication, quadratic forms, logarithm, exponential, largest eigenvalue, etc.): some can also, automatically, discretize optimization problems involving functions, probability distributions, differential operators, integrals, and expectations.

A unifying modeling abstraction for infinite-dimensional optimization
J.L. Pulsipher et al. (2021)
https://arxiv.org/abs/2106.12689

Convex optimization and neural networks

As always, it is possible to use a neural network to help solve optimization problems (train the network to generate a partial solution, complete it, and finish with a few gradient descent steps).

DC3: a learning method for optimization with hard constraints
P.L. Donti et al. (2021)
https://arxiv.org/abs/2104.12225

This also works for combinatorial problems.

Solving mixed integer programs using neural networks
V. Nair et al.
https://arxiv.org/abs/2012.13349

Exact combinatorial optimization with graph convolutional neural networks
M. Gasse et al. (2019)
https://arxiv.org/abs/1906.01629

Conversely, it is possible (with cvxpylayers) to put an optimization layer inside your neural network, in particular if you had a 2-step procedure, in which you were fitting some model, and using its output in an optimization problem.

Automatically learning compact quality aware surrogates for optimization problems
K. Wang et al. (2020)
https://arxiv.org/abs/2006.10815

Non-convex optimization and ADMM

Non-convex optimization is becoming more widespread: in particular, you may want a non-convex penalty in your optimization problems – the traditional L¹ penalty is an approximation of the L⁰ penalty, but there are better (but non-convex) ones.

Those optimization problems can be solved with the CCCP (convex concave procedure, or difference of convex (DC) algorithm)

A unified algorithm for the non-convex penalized estimation: the ncpen package
D. Kim et al. (2020)
https://journal.r-project.org/archive/2021/RJ-2021-003/index.html

or the ADMM algorithm, which is now very popular.

Machine learning optimization algorithms and portfolio allocation
S. Perrin and T. Roncalli (2019)
https://arxiv.org/abs/1909.10233

Optimization on a manifold

Constrained optimization on a Euclidean space can often be reformulated as unconstrained optimization on a manifold (e.g., the Steifel manifold if you are looking for an orthogonal matrix): the optimization can be performed entirely inside the manifold – for instance, the gradient steps are performed with parallel transport along a geodesic.

ManifoldOptim: an R interface to the ROPTLIB library for Riemannian manifold optimization
S.R. Martin et al. (2020)
https://arxiv.org/abs/1612.03930

Optimization pathologies

Some of the pleasant properties of convex optimization are actually... not true.

Curiosities and counterexamples in smooth convex optimization
J. Bolte and E. Pauwels (2020)
https://arxiv.org/abs/2001.07999

Reinforcement learning in finance

Reinforcement learning seems to have become very popular in finance, but most of the applications are a bit suspicious: they use too little data for the conclusions to be convincing – only a few assets, often just one, and features from prices alone.

There are a few exceptions, though. For instance, we can combine many financial ratios, for thousands of stocks, to directly compute portfolio weights, maximizing some measure of performance, such as the information ratio. (I developed a similar model a few years ago.)

AlphaPortfolio: direct construction through deep reinforcement learning and interpretable AI
L.W. Cong (2019)
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3554486

This description may not sound like reinforcement learning, but this end-to-end approach is actually a "policy gradient" algorithm.

A universal end-to-end approach to portfolio optimization via deep learning
C. Zhang et al. (2021)
https://arxiv.org/abs/2111.09170

End-to-end risk budgeting portfolio optimization with neural networks
A.S. Uysal et al. (2021)
https://arxiv.org/abs/2107.04636?context=q-fin.CP

Integrating prediction in mean-variance portfolio optimization
A. Butler and R.H. Kwon (2021)
https://arxiv.org/abs/2102.09287

The other credible application of reinforcement learning in finance uses high-frequency data and limit-order-book features.

Multi-horizon forecasting for limit order books: novel deep learning approaches and hardware acceleration using intelligent processing units
Z. Zhang and S. Zohren (2021)
https://arxiv.org/abs/2105.10430

Deep reinforcement learning for active high-frequency trading
A. Briola et al.
https://arxiv.org/abs/2101.07107

There are also applications in (long-term) financial planning. (I did something similar in 2013, without knowing it was called "reinforcement learning").

Embracing advanced AI/ML to help investors achieve success: RL for financial goal planning
S. Mohammed et al.
https://arxiv.org/abs/2110.12003

Several books on reinforcement learning in finance have appeared: here is one.

Foundations of reinforcement learning with applications in finance
A. Rao and T. Jelvis (2021)
https://stanford.edu/~ashlearn/RLForFinanceBook/book.pdf

But my recommendation, for a (non-finance) reinforcement learning book, is still:

Deep Reinforcement Learning hands-on
M. Lapan (2018)

Explainable AI (XAI)

Shapley score

The Shapley score (I prefer to speak of Shapley "contributions") has become mainstream: it is now (one of) the standard way(s) of explaining any type of model. It decomposes the output (of a neural network, or anything) into a sum of contributions of the inputs.

Explaining by removing: a unified framework for model explanation
I.C. Covert et al. (2020)
https://arxiv.org/abs/2011.14878

As always, there is neural network approach, useful if you want to repeatedly compute Shapley scores for the same model (but even if you want the contributions for a single sample, learning a neural network with many other samples may produce a better result than the traditional Monte Carlo approximation).

FastSHAP: Real-Time Shapley Value Estimation
N. Jethani et al. (2021)
https://arxiv.org/abs/2107.07436

Plotting the Shapley contributions versus the values of the features may be insightful.

Understanding machine learning for diversified portfolio construction by explainable AI
M. Jaeger et al. (2020)

For a list of software packages in R and Python to explain machine learning models, check:

Landscape of R packages for explainable artificial intelligence
S. Maksymiuk et al. (2021)
https://arxiv.org/abs/2009.13248

The underlying theory is interesting: it is related to non-additive measures (Choquet integral) and submodular functions (Lovász extension).

Non-additive measures
V. Torra et al. (2014)
https://link.springer.com/book/10.1007/978-3-319-03155-2

Learning with submodular functions: a convex optimization perspective
F. Bach (2013)
https://arxiv.org/abs/1111.6453

Shapley score in finance

Since the Shapley approach can decompose anything into a sum of contributions, this applies to measures of performance – and we can use the same approach for all measures, regardless of how they are defined (we do not need a separate approach for each of them). We can compute the contributions of the assets in the portfolio,

The Shapley decomposition for portfolio risk
S. Mussard and V. Terraza (2008)
http://gredi.recherche.usherbrooke.ca/wpapers/GREDI-0609.pdf

or of features that can be turned on or off (inputs, constraints, penalties in the optimization objective, etc.)

Porfolio performance attribution via Shapley value
N. Moehle et al. (2021)
https://arxiv.org/abs/2102.05799

Explainable models

Instead of interpreting models as an afterthought, some prefer to stick to models that are inherently interpretable, such as generalized models with interactions (GA²M, aka explainable boosting machines, EBM)

Accurate intelligible models with pairwise interactions
Y. Lou et al. (2013)

GAMI-Net: An Explainable Neural Network based on Generalized Additive Models with Structured Interactions
Z. Yang et al. (2020)
https://arxiv.org/abs/2003.07132

GAM boosting,

Boosting Algorithms: Regularization, Prediction and Model Fitting
P. Bühlmann and T. Hothorn (2007)
https://projecteuclid.org/journals/statistical-science/volume-22/issue-4/Boosting-Algorithms-Regularization-Prediction-and-Model-Fitting/10.1214/07-STS242.full

Model-based Boosting in R
B. Hofner et al. (2014)
https://cran.r-project.org/web/packages/mboost/vignettes/mboost_tutorial.pdf

or GAM-like neural networks.

Enhancing Explainability of Neural Networks through Architecture Constraints
Z. Yang et al. (2019)
https://arxiv.org/abs/1901.03838

Monotonic neural nets

An easy way of making neural networks interpretable by construction, with a prescribed interpretation, is to impose monotonicity constraints: this ensures that each variable is used in the direction we expect.

The easiest way of doing this is to add a penalty whenever the sign of a gradient (on a training sample) is incorrect.

Monotonic trends in deep neural networks
A. Gupta at al. (2019)
https://arxiv.org/abs/1909.10662

This only guarantees that the function learned by the neural network is monotonic in a neighbourbood of each training sample. To actually ensure that the function is monotonic, everywhere, we can train a neural network to look for an adversarial example or, even better, actually prove (mathematically) that the function is monotinic: if the neural network only uses ReLU activations, this is a satisfiability problem, which can be solved with MILP or SMT solvers.

Certified Monotonic Neural Networks
X. Liu et al. (2020)
https://arxiv.org/abs/2011.10219

Counterexample-guided learning of monotonic neural networks
A. Sivaraman et al. (2020)
https://arxiv.org/abs/2006.08852

An older approach to monotonic neural network was to build them from monotonic blocks, but they were either too restrictive (positive weights and ReLU activations produce an increasing function, but it is always convex as well) or complex to implement (e.g., lattice networks).

Deep Lattice Networks and Partial Monotonic Functions
S. You (2017)
https://arxiv.org/abs/1709.06680

Training neural networks

Loss landscape

In the loss landscape of deep learning models, global minima are believed to form submanifolds: looking for a submanifold minimizing the loss function (a segment, a simplex, a Bézier curve, etc.) instead of just a point, helps generalization.

Loss surface simplexes for mode-connecting volumes and fast ensembling
G.W. Benton et al. (2021)
https://arxiv.org/abs/2102.13042
Learning neural network subspaces
M. Wortsman et al. (2021)
https://arxiv.org/abs/2102.10472

Double descent (interpolation regime)

It was once believed that, as the complexity of the model increases, the (out-of-sample) performance first improved, as the model becomes complex enough to fit the data, and then degraded, as the model becomes complex enough to overfit the training data. But there is a third regime: when the model becomes complex enough to perfectly fit the training data (the interpolation regime), making it more complex will let the optimization find smoother models interpolating the data, and these models generalize better: this is the "double descent phenomenon".

Fit without fear: remarkable mathematical phenomena of deep learning through the prism of interpolation
M. Belkin (2021)
https://arxiv.org/abs/2105.14368

Surprises in high-dimensional ridgeless least squares interpolation
T. Hastie et al. (2020)
https://arxiv.org/abs/1903.08560

Missing data

Multiple imputation

The two naive approaches to missing data are either to remove the observations with at least one missing value (potentially discarding a lot of valuable data) or to "impute" the missing values, by replacing them with their "most likely value" (e.g., the median of the observed values, or the prediction of a model to predict that variable given the other ones).

These approaches are suboptimal and potentially very biased: imagine estimating the volatility of a time series of returns after replacing the missing returns with their median...

Instead, one can replace the missing values, not with a value, but with a distribution. Since dealing with distributions instead of values is a bit tricky, one can use samples from those distributions instead; 5 samples, i.e., generating 5 completed datasets, is often considered enough (multiple imputation).

mice: multiple imputation by chained equations in R
S. van Buuren and K. Groothuis-Oudshooin (2011)
https://www.jstatsoft.org/article/view/v045i03

Amelia II: a program for missing data
J. Honaker et al. (2011)
https://www.jstatsoft.org/article/view/v045i07

Mixtures, EM and missing data
B. Stewart (2017)
https://scholar.princeton.edu/sites/default/files/bstewart/files/lecture5_missing_slides.pdf

Time series with missing values

For time series, it is easy to impute univariate data

imputeTS: time series missing value imputation in R
S. Moritz and T. Batrz-Beielstein (2017)
https://journal.r-project.org/archive/2017/RJ-2017-009/index.html

Comparison of different methods for univariate time series imputation in R
S. Moritz et al. (2015)
https://arxiv.org/abs/1510.03924

but mutivariate data complicates things.

Imputation, estimation and missing data in finance
G. DiCesare (2006)

What to do about missing values in time series cross-section data
J. Honaker and G. King (2010)
https://gking.harvard.edu/files/abs/pr-abs.shtml

A simple approach is to assume that the data is Gaussian (e.g., log-prices, assumed to follow a multivariate Gaussian random walk): the conditional Gaussian distribution formula then gives the distribution of the missing data given the observed data. This may not look scalable (the matrix to invert is very large), but it can be implemented efficiently with a Kalman filter.

Equivalently, one can use Gaussian processes.

Gaussian process imputation of multiple financial time series
T. de Wolff et al. (2020)
https://arxiv.org/abs/2002.05789

As always, there are also deep learning approaches.

BRITS: bidirectional recurrent imputation for time series
W. Cao et al. (2018)
https://arxiv.org/abs/1805.10572

NAOMI: Non-autoregressive multiresolution sequence imputation
Y. Liu et al. (2019)
https://arxiv.org/abs/1901.10946

Multivariate time series imputation with generative adversarial networks
Y. Luo et al. (2018)
https://papers.nips.cc/paper/2018/hash/96b9bff013acedfb1d140579e2fbeb63-Abstract.html

Time series imputation and prediction with bidirectional generative adversarial networks
M. Gupta and R. Beheshti (2020)
https://arxiv.org/abs/2009.08900

Deep learning with missing values

Deep learning usually assumes there are no missing values, but we can relax that assumption.

Neumiss networks: differentiable programming for supervised learning with missing values
M. Le Morvan et al. (2020)
https://arxiv.org/abs/2007.01627

Ordinary differential equations (ODE) and neural networks

Invertible layers

As data moves through a neural network, irrelevant information is progressively discarded. This is sometimes intentional (if we want a yes/no answer), sometimes not (if we want to transform an image).

Judicious use of skip-layer connections can make a network invertible, by construction. This is related to second-order differential equations (residual networks can be seen as a discretization of first-order ODEs).

Momentum residual networks
M.E. Sander et al. (2021)
https://arxiv.org/abs/2102.07870

Fully hyperbolic convolutional neural networks
K. Lensink et al. (2019)
https://arxiv.org/abs/1905.10484

Normalizing flows are another way of making layers invertible.

Self-normalizing flows
T.A. Keller et al. (2020)
https://arxiv.org/abs/2011.07248

GNNs and ODEs

Graph neural networks (GNN) aggregate the information from neighbouring nodes: they can be seen as the Euler discretization of a diffusion PDE (partial differential equation). Other discretizations (Runge-Kutta, implicit schemes, etc.) may perform better.

GRAND: graph neural diffusion
B.P. Chamberlain et al. (2021)
https://arxiv.org/abs/2106.10934

Identifiability

Neural networks allow us to estimate a differential equation from observations but, with a single trajectory, the problem is ill-posed (but regularization helps).

Beyond prediction in neural ODEs: identification and interventions
H. Aliee et al. (2021)
https://arxiv.org/abs/2106.12430

Neural ODEs

There are many variants of neural ODEs: NODE, CT-RNN, LTC, etc.

Liquid time-constant networks
R. Hasani et al. (2021)
https://arxiv.org/abs/2006.04439

Graphs

DAG constraint

It is not straightforward to process graphs with neural networks. For your model to output a graph, it suffices to output an adjacency matrix, but things become more complicated if you want to impose properties on that graph, for instance, to make it acyclic – those properties look combinatorial.

It turns out to be surprisingly easy (both to state and prove): a directed graph is acyclic iff its adjacency matrix A satisfies trace( exp(A) ) = d. Since this is differentiable, it can be used as a constraint in the loss function.

DAGs with NO TEARS: continuous optimization for structure learning
X. Zheng et al. (2018)
https://arxiv.org/abs/1803.01422

Dynotears: structure learning from time series data
R. Pamfil et al. (2020)
https://arxiv.org/abs/2002.00498

Graph features: network portrait

When processing complicated objects (time series, correlation matrices, graphs, etc.) with machine learning algorithms or deep learning models, we first transform them into a set of numbers of numbers ("features"), summarizing one aspect of the object.

For instance, for graph, the "network portrait" counts how many nodes have k n-hop neighbours,

An information-theoretic all-scales approach to comparing networks
J.P. Bagrow and E.M. Bollt (2019)
https://arxiv.org/abs/1804.

Portraits of complex networks
J.P. Bagrow et al. (2007)
https://arxiv.org/abs/cond-mat/0703470

Comparing methods for comparing networks
T. Tantardini et al. (2019)
https://www.nature.com/articles/s41598-019-53708-y

PGM

Graphs (more precisely, probabilistic graphical models (PGMs), Bayesian networks) can be used to model the relations between a large number of variables, if we have some prior knowledge that each relation only involves a small numberr of variables.

A probabilistic graphical model approach to model interconnectedness
A. Denev et al. (2017)  
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3078021

Which variables are involved in which relations can be automatically inferred by processing text (news, etc.).

Building probabilistic causal models using collective intelligence
O. Laudy et al. (2021)
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3808233

Networks

If you need a reference for the study of networks, the following (700-page) book looks up-to-date and exhaustive; it is easy to read, and nicely illustrated.

The atlas for the aspiring network scientist
M. Coscia (2020)
https://arxiv.org/abs/2101.00863

Linear algebra

PCA

Principal component analysis (PCA) is a well-known topic, and we would not expect any new development around it.

Principal component analysis: a review and recent developments
I.T. Jolliffe and J. Cadima (2016)
https://royalsocietypublishing.org/doi/pdf/10.1098/rsta.2015.0202

When computing PCA with time series, on a moving window, we would like the eigenvectors to change progressively. Even if the eigenvalues remain distinct, this can be problematic, because only the eigenspaces are well-defined: the eigenvectors can freely flip sign. Computing the PCA with an iterative procedure can help avoid that problem.

Iterated and exponentially weighted moving principal component analysis
P. Bilokon and D. Finkelstein (2021)
https://arxiv.org/abs/2108.13072

Iterative refinement for symmetric eigenvalue decomposition
T. Ogita and K. Aishima (2018)
https://www.keisu.t.u-tokyo.ac.jp/data/2016/METR16-11.pdf

There are several generalizations of PCA adapted to time series. For instance, one can diagonalize, not just the covariance matrix, but (jointly) several cross-covariance matrices, for several lags. The first dynamic principal component of a set of time series is the time series that allows the best reconstruction of all those time series, if we also use its lags.

Dimension reduction for time series in a blind source separation context using R
K. Nordhausen et al. (2021)
https://www.jstatsoft.org/article/view/v098i15

gdpc: an R package for generalized dynamic principal components
D. Peña et al. (2020)
https://www.jstatsoft.org/article/view/v092c02

Generalized dynamic principal components
D. Peña and V.J. Yohai (2015)
http://halweb.uc3m.es/esp/Personal/personas/dpena/publications/ingles/2016JASA_yohai.pdf

Even though PCA is an archetypal unsupervised problem, there are supervised variants...

Dimension reduction in R
S. Weisberg (2002, 2015)
https://www.jstatsoft.org/article/view/v007i01

There are sparse variants (adding an L¹ penalty is not quite enough),

Principal component analysis with sparse fused loading
J. Guo et al. (2010)
http://dept.stat.lsa.umich.edu/~jizhu/pubs/Guo-JCGS10.pdf

and higher-dimensional variants (tensor decompositions).

TensorLy: tensor learning for Python
J. Kossaifi et al. (2019)
https://jmlr.org/papers/v20/18-277.html
https://github.com/tensorly/tensorly

Tensor learning for regression
W. Guo and I. Patras (2012)
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.676.870

Sparse and parsimonious matrices

To reduce the number of parameters needed to describe a matrix, we often assume it is approximately low-rank and use a truncated singular value decomposition (SVD), or that it is diagonal or block diagonal. We can combine those two ideas, and look for a block decomposition, with dense diagonal blocks and low-rank off-diagonal blocks.

Efficient scalable algorithms for hierarchically semiseparable matrices
S. Wang et al.
https://www.math.purdue.edu/~xiaj/parhss.pdf

Random projections may provide good enough an approximation (Johnson-Lindenstrauss lemma).

Fast and accurate network embeddings via very sparse random projection
H. Chen et al. (2019)
https://arxiv.org/abs/1908.11512

Database-friendly random projections: Johnson-Lindenstrauss with binary coins
D. Achlioptas (2002)
http://cgi.di.uoa.gr/~optas/papers/jl.pdf

Fused lasso

The lasso penalty can be seen as forcing a group of predictors to have the same coefficient, zero. This can be generalized to several groups, all variables in the same group sharing the same coefficient.

Simultaneous supervised clustering and feature selection over a graph
X. Shen et al. (2012)
https://europepmc.org/articles/pmc3629856?pdf=render

Variance matrices

Finding good estimators of large variance or correlation matrices is a perennial problem in finance.

Parametrizations

When estimating a correlation matrix, as part of some optimization problem, it may be difficult to impose the "correlation matrix" constraint. Fortunately, there are (many) parametrizations of correlation matrices, which turn those problems into unconstrained optimization problems.

The most general methodology to create a valid correlation matrix for risk management and option pricing
R. Rebonato and P. Jäckel (1999)
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1969689

Unconstrained parametrization for variance covariance matrices
J.C. Pinheiro and D.M. Bates (1996)
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.31.494

The factor decomposition of variance matrices can be applied recursively, with smaller and smaller variance matrices: Vₙ₊₁ = βₙ vₙ βₙ' + Δₙ.

Heterotic risk models
Z. Kakushadze (2015)
https://arxiv.org/abs/1508.04883

Graphs

Graphs are often used to understand or simplify correlation matrices, for instance by considering the mimimum spanning tree (add edges, one by one, starting with the most important ones, but only if they do not introduce any cycle) and discarding all the correlations not associated with an edge. This can be generalized to other properties: planar graphs, triangulated graphs, etc.

Portfolio optimization with sparse multivariate modelling
P.F. Procacci and T. Aste (2021)
https://arxiv.org/abs/2103.15232

The graph associated to a correlation matrix can also be used to define features which may help in forecasting future correlation.

Forecasting financial market structure from network features using machine learning
D. Castilho et al. (2021)
https://arxiv.org/abs/2110.11751

Constraints

For more reliable estimates of (large) correlation matrices, one can add constraints, e.g., on the sparsity patterns of the precison matrix, or the sign of the partial correlations.

The resulting optimization problems may no longer be convex, but they are amenable to ADMM.

Algorithms for learning graphs in financial markets
J.V.M. Cardoso et al. (2020)
https://arxiv.org/abs/2012.15410

Learning undirected graphs in financial markets
J.V.M. Cardodo and D.P. Palomar (2020)
https://arxiv.org/abs/2005.09958

Learning high-dimensional Gaussian graphical models under total positivity without adjustment of tuning parameters
Y. Wang et al. (2020)
https://arxiv.org/abs/1906.05159

Asynchronous estimators

Estimating a covariance matrix from asynchronous data is much trickier.

Under Gaussian assumptions, we can model the data as a state space process (log-return time series form a multivariate Brownian motion) and use a Kalman filter.

There are other approaches, which also try to remove microstructure noise.

High frequency covariance: a Julia package for estimating covariance matrices using high-frequency financial data
S. Baumann and M. Klymak (2021)
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3786912

On covariation estimation for multivariate continuous Itō semimartingales with noise in non-synchronous observation schemes
K. Christensen et al. (2011)
https://econ.au.dk/fileadmin/site_files/filer_oekonomi/Working_Papers/CREATES/2011/rp11_53.pdf

Estimating the quadratic covariation matrix from observations: local method of moments and efficiency
M. Bibinget et al. (2014)
https://arxiv.org/abs/1303.6146

Neural networks

Neural networks can help estimate variance matrices, e.g., to compute factor returns in a factor model (X=βF+ε), or, directly, to replace the factor model (X=ϕ(F)+ε).

Deep risk models: a deep learning solution for mining latent risk factors to improve covariance matrix estimation
H. Lin et al.
https://arxiv.org/abs/2107.05201
 
Deep fundamental factor models
M.F. Dixon and N.G. Polson (2020)
https://arxiv.org/abs/1903.07677

Deep factor model
K. Nakagawa et al. (2018)  
https://arxiv.org/abs/1810.01278

Bayesian estimators

The Black-Litterman framework combines two sources of information about a Gaussian distribution: a prior, X∼N(μ,V), and some additional but incomplete information ("views"), w'X∼N(m,v). It has been extended in many ways, for instance with views on volatilities, correlations, or other moments.

Conditional distribution in portfolio theory
E. Qian and S. Gorman (2001)
(paywall)

Fully flexible views: theory and practice
A. Meucci (2008)
https://arxiv.org/abs/1012.2848

Miscellaneous

Generative models in Finance

Generative models, in particular GANs and conditional GANs, are now used to generate synthetic data in finance, in particular, (very low-dimensional) time series, correlation matrices, or limit-order books.

Towards realistic market simulations: a generative adversarial networks approach
A. Coletta et al. (2021)
https://arxiv.org/abs/2110.13287

Improving the robustness of trading strategy backtesting with Boltzmann machines and generative adversarial networks
E. Lezmi et al. (2020)
https://arxiv.org/abs/2007.04838

Matrix evolutions: synthetic correlations and explainable machine learning for constructing robust investment portfolios
J. Papenbrock et al. (2021)
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3663220

Generating realistic stock market order streams
J. Li et al. (2020)
https://arxiv.org/abs/2006.04212

Style transfer with time series: generating synthetic financial data
B. da Silva and S.S. Shi (2019)
https://arxiv.org/abs/1906.03232

Time series generative adversarial networks
J. Yoon et al. (2019)
https://arxiv.org/abs/2107.11098

(Conditional) GANs can also be used for forecasting, in a Bayesian way: use the predictors as conditioning variables and generate several samples, to approximate a posterior distribution.

Risk measures

There are many generalizations of value-at-risk (VaR) and expected shortfall (ES, or CVaR), allowing for several scenarios, i.e., several distributions for future asset prices (they appear, informally, in the Basel regulations).

Scenario-based risk evaluation
R. Wang and J.F. Ziegel (2018)
https://arxiv.org/abs/1808.07339

A framework for measures of risk under uncertainty
T. Fadina et al. (2021)
https://arxiv.org/abs/2110.10792

Many risk measures can be defined as the minimum amount to invest in some reference asset (e.g., cash) to make the strategy considered "acceptable", in some sense (e.g., so that its value, at some point in the future, be positive, with probability above 99%).

This can be generalized to assets other than cash, and even to several assets.

Risk measures beyond frictionless markets
M. Arduca and C. Munari (2021)
https://arxiv.org/abs/2111.08294

Online estimation and optimization of utility-based shortfall risk
A.S. Menon et al. (2021)
https://arxiv.org/abs/2111.08805

k-nearest neighbours

The k-nearest neighbour classifier can be generalized (a weighted average of the k nearest neighbours), and extrapolated to k=0.

Extrapolation towards imaginary 0-nearest neighbour and its improved convergence rate
A. Okuno and H. Shimodaira (2020)
https://arxiv.org/abs/2002.03054

Mode and quantiles

The mode tree is a simple plot to help assess the presence of several modes in univariate data.

multimode: an R package for mode assessment
J. Ameijeiras-Alonso et al. (2021)
https://www.jstatsoft.org/article/view/v097i09

There are many multidimensional generalizations of the median.

Computing the Oja median in R: the package OjaNP
D. Fischer et al. (2020)
https://www.jstatsoft.org/article/view/v092i08

Even in dimension 1, the notion of quantile can be generalized.

An axiomatization of Λ-quantiles
F. Bellini and I. Peri (2021)
https://arxiv.org/abs/2109.02360

Multipole methods

The Barnes-Hut algorithm speeds up n-body simulations by aggregating long-distance interactions. The Fast Multipole Method (FMM) generalizes that to arbitrary interactions (arbitrary kernels), but requires approximations specific to the kernel (a low-rank factorization of the off-diagonal blocks of the kernel matrix, i.e., of long-range interactions).

A short course on fast multipole methods
R. Beatson and L. Greengard (1997)
https://math.nyu.edu/~greengar/shortcourse_fmm.pdf

With automatic differentiation, the computer can do the computations, and use the FMM for arbitrary kernels.

The fast kernel transform
J.P. Ryan et al.
https://arxiv.org/abs/2106.04487

Data envelopment analysis (DEA)

Data envelopment analysis (DEA) is a way of measuring the ``distance'' to the efficient frontier. It is a very old topic, but it seems to re-emerge from time to time.

A data envelopment analysis toolbox for Matlab
I.C. Álvarez et al. (2020)
https://www.jstatsoft.org/article/view/v095i03

Decentralized finance (DeFi) and cryptocurrencies

Cryptocurrencies do not seem to lose any of their popularity, and blockchain technologies make decentralized exchanges (automated market makers) possible: anyone can become a liquidity provider.

The adoption of blockchain-based decentralized exchanges
A. Capponi and R. Jia (2021)
https://arxiv.org/abs/2103.08842

Compression with neural nets

To compress and image, one can simply overfit a neural net (pixel coordinates ↦ RGB) to it.

COIN: Compression with implicit neural representation
E. Dupont et al. (2021)
https://arxiv.org/abs/2103.03123

Matrix profile of a time series

When studying complicated objects, such as images, sounds, text, graphs, correlation matrices, we often summarize the information they contain with a few numbers: features. For time series, the matrix profile is a recent addition to the long list of features: it is another time series, which keeps track, for each observation, of the distance to the nearest observation.

tsmp: an R package for time series with matrix profile
F. Bischoff and P.P. Rodrigues (2020)
https://journal.r-project.org/archive/2020/RJ-2020-021/index.html

Causality

Transfer entropy is a non-linear generalization of the Granger causality test, but it is difficult to estimate: there are better estimators than the naive ones.

NlinTS: an R package for causality detection in time series
Y. Hmamouche (2020)
https://journal.r-project.org/archive/2020/RJ-2020-016/index.html

posted at: 11:16 | path: /ML | permanent link to this entry