SIDE Summer School, day 1

This morning, we start the SIdE (Italian Econometric Association) Summer School, on Machine Learning Algorithms for Econometricians. Emmanuel Flachaire will start with a presentation of nonparametric econometric techniques. I will then get back to the geometry of (standard) econometric techniques, to introduce kernels. The first series of slides are online.

I will then spend more time on the (popular) idea of “least squares” and mention other loss functions. Slides are online.

Optimal transport on large networks

With Alfred Galichon and Lucas Vernet, we recently uploaded a paper entitled optimal transport on large networks on arxiv.

This article presents a set of tools for the modeling of a spatial allocation problem in a large geographic market and gives examples of applications. In our settings, the market is described by a network that maps the cost of travel between each pair of adjacent locations. Two types of agents are located at the nodes of this network. The buyers choose the most competitive sellers depending on their prices and the cost to reach them. Their utility is assumed additive in both these quantities. Each seller, taking as given other sellers prices, sets her own price to have a demand equal to the one we observed. We give a linear programming formulation for the equilibrium conditions. After formally introducing our model we apply it on two examples: prices offered by petrol stations and quality of services provided by maternity wards (only the later is described here for privacy issues). These examples illustrate the applicability of our model to aggregate demand, rank prices and estimate cost structure over the network. We insist on the possibility of applications to large scale data sets using modern linear programming solvers such as Gurobi.

Demand for gas in gas stations in Britanny, and demand for maternity in France (with border correction)

In addition to this paper we released a R toolbox to implement our results and an online tutorial, optimalnetwork.github.io.

Pareto Models for Top Incomes

This week, The Society for the Study of Economic Inequality (ECINEQ) organised the Eighth ECINEQ Meeting 2019  in Paris, hosted by the Paris School of Economics and the World Inequality Lab. Emmanuel Flachaire was there to present our joint work on Pareto Models for Top Incomes. Slides are also available


The paper is still available on hal, and the package (TopIncomes) is also available from github,

library(devtools)
install_github("freakonometrics/TopIncomes")
library(TopIncomes)

On my way to Manizales (Colombia)

Next week, I will be in Manizales, Colombia, for the Third International Congress on Actuarial Science and Quantitative Finance. I will be giving a lecture on Wednesday with Jed Fress and Emilianos Valdez.

I will give my course on Algorithms for Predictive Modeling on Thursday morning (after Jed and Emil’s lectures). Unfortunately, my computer locked itself last week, and I could not unlock it (could not IT team at the university, who have the internal EFI password). So I will not be able to work further on the slides, so it will be based on the version as-at now (clearly in progress).

 

Quinze ans…

Il y a presque quinze ans, jour pour jour, en juin 2004, le premier tome de Mathématiques de l’Assurance Non-Vie  était disponible en librairie…

Quinze ans…! Déjà… Autant le second tome est un peu dépassé aujourd’hui (on reste sur les GLM et les GAM, mais de nombreuses autres techniques statistiques auraient pu être présentée si on avait pris le temps de faire un vraie nouvelle édition), autant le premier me semble toujours au goût du jour. Avec le recul, je mettrais la section plus économique dans le premier tome, en rajoutant une section sur la concurence… mais avec quinze ans de recul, je suis toujours fier de voir ce livre trôner dans la bibliothèque de collègues, ainsi que de nombreux praticiens….

Pareto Models for Top Incomes

With Emmanuel Flachaire, we uploaded on hal a paper on Pareto Models for Top Incomes,

Top incomes are often related to Pareto distribution. To date, economists have mostly used Pareto Type I distribution to model the upper tail of income and wealth distribution. It is a parametric distribution, with an attractive property, that can be easily linked to economic theory. In this paper, we first show that modelling top incomes with Pareto Type I distribution can lead to severe over-estimation of inequality, even with millions of observations. Then, we show that the Generalized Pareto distribution and, even more, the Extended Pareto distribution, are much less sensitive to the choice of the threshold. Thus, they provide more reliable results. We discuss different types of bias that could be encountered in empirical studies and, we provide some guidance for practice. To illustrate, two applications are investigated, on the distribution of income in South Africa in 2012 and on the distribution of wealth in the United States in 2013.

This paper was presented at and UCSB and in several workshops this spring, and this Summer, Emmanuel will present it at ECINEQ.

Note that a R package is also available on github, TopIncomes.

Données Agrégées et Variables Compositionnelles

Avec Enora Belz, nous venons de mettre en ligne une note méthodologique, Données Agrégées et Variables Compositionnelles, sur hal.

La réforme du droit sur les données personnelles en Europe rend difficile l’accès aux données individuelles (même souvent non-nominatives), surtout quand on cherche des données jugées sensibles (et souvent, le revenu entre dans cette catégorie). Une solution souvent envisagée est la mise à disposition de données agrégées spatialement. Ces données posent toutefois deux problèmes techniques. Le premier est que les données catégorielles deviennent des compositions. Le second est lié au paradoxe écologique qui dit qu’il est dangereux d’inférer des relations économétriques individuelles à partir de données agrégées. Nous verrons ici comment travailler avec des données compositionnelles (pour éventuellement juste valider une approche classique de régression linéaire-plus simple à interpréter). Et nous évoquerons le second, mais qui reste malheureusement trop général pour pouvoir être traité de manière satisfaisante.

Extended Scale Free Networks

With Emmanuel Flachaire, we recently uploaded a short article, Extended Scale-Free Networks, on arxiv.

Recently, Broido & Clauset (2019) mentioned that (strict) Scale-Free networks were rare, in real life. This might be related to the statement of Stumpf, Wiuf & May (2005), that sub-networks of scale-free networks are not scale-free. In the later, those sub-networks are asymptotically scale-free, but one should not forget about second-order deviation (possibly also third order actually). In this article, we introduce a concept of extended scale-free network, inspired by the extended Pareto distribution, that actually is maybe more realistic to describe real network than the strict scale free property. This property is consistent with Stumpf, Wiuf & May (2005): sub-network of scale-free larger networks are not strictly scale-free, but extended scale-free.

Estimates on training vs. validation samples

Before moving to cross-validation, it was natural to say “I will burn 50% (say) of my data to train a model, and then use the remaining to fit the model”. For instance, we can use training data for variable selection (e.g. using some stepwise procedure in a logistic regression), and then, once variable have been selected, fit the model on the remaining set of observations. A natural question is usually “does it really matter ?”.

In order to visualize this problem, consider my (simple) dataset

MYOCARDE=read.table(
  "http://freakonometrics.free.fr/saporta.csv",
  head=TRUE,sep=";")

Let us generate 100 training samples (where we keep about 50% of the observations). On each of them, we use a stepwise procedure, and we keep the estimates of the remaining variables (and their standard deviation actually)

n=nrow(MYOCARDE)
M=matrix(NA,100,ncol(MYOCARDE))
colnames(M)=c("(Intercept)",names(MYOCARDE)[1:7])
S1=S2=M1=M2=M
for(i in 1:100){
idx = which(sample(0:1,size=n, replace=TRUE)==1)
reg=step(glm(PRONO=="DECES"~.,data=MYOCARDE[idx,]))
nm=names(reg$coefficients)
M1[i,nm]=reg$coefficients
S1[i,nm]=summary(reg)$coefficients[,2]
f=paste("PRONO=='DECES'~",paste(nm[-1],collapse="+"),sep="")
reg=glm(f,data=MYOCARDE[-idx,])
M2[i,nm]=reg$coefficients
S2[i,nm]=summary(reg)$coefficients[,2]
}

Then, for the 7 covariates (and the constant) we can look at the value of the coefficient in the model fitted on the training sample, and the value on the model fitted on the validation sample (of course, only when they were remaining)

for(j in 1:8){
idx=which(!is.na(M1[,j]))
plot(M1[idx,j],M2[idx,j])
abline(a=0,b=1,lty=2,col="gray")
segments(M1[idx,j]-2*S1[idx,j],M2[idx,j],M1[idx,j]+2*S1[idx,j],M2[idx,j])  
segments(M1[idx,j],M2[idx,j]-2*S2[idx,j],M1[idx,j],M2[idx,j]+2*S2[idx,j])  
}

For instance, with the intercept, we have the following

 

where horizontal segments are confidence intervals of the parameter on the model fitted on the training sample, the vertical on the validation sample. The green part means some sort of consistency, while the red one means that actually, the coefficient was negative with one model, positive with the other one. Which is odd (but in that case, observe that coefficients are rarely significant).

We can also visualize the joint distribution of the two estimators,

for(j in 1:8){
library(ks)
idx = which(!is.na(M1[,j]))
Z = cbind(M1[idx,j],M2[idx,j])
H = Hpi(x=Z)
fhat = kde(x=Z, H=H)
image(fhat$eval.points[[1]],
fhat$eval.points[[2]],fhat$estimate)
abline(a=0,b=1,lty=2,col="gray")
abline(v=0,lty=2)
abline(h=0,lty=2)
}

which are here, almost on the diagonal,

meaning that the intercept on the two samples is (more or less) the same. We can then look at other parameters (which is actually more interesting).

On that variable, it seems that it is significant on the training dataset (somehow, it is consistent with the fact that it is remaining in the model after the stepwise procedure) but not on the validation sample (or hardly significant).

Others are much more consistent (with some possible outliers)

 

 

On the next one, we have again significance on the training sample, but not on the validation sample,

 

 

and probably more interesting

where the two are very consistent.

From betting to “prediction market”

This is the second part of a series on sports betting

Sports betting has long fascinated economists and statisticians. Griffith (1949) showed early on that horse race bettors put too much money on horses that have little chance of winning, and too little on those that have the best chance of winning. This tendency to underbid on the most likely event has been obaserved in all sports betting, where the “most likely event” is calculated on the basis of recent statistics. And it can be explained in a fundamental way by the mechanics of mutual betting: the bettor opposes his beliefs to those of the crowd, because the various bets will determine the odds.

Predictions, before surveys

Today, in the months leading up to each election, we find ourselves drowned under the polls, conducted every day (and commented on several times a day, as if estimation noise was worth exegesis). As Frédéric Dabi (Deputy Director General of Ifop) pointed out in a debate organised by Risques magazine in 2017, “surveys are an indication of the electoral balance of power, not a prediction”, but it is nevertheless often in the idea of having a prediction that they are used.

But if we go back in time, Rhode & Strumpf (2008) reminds us that other techniques were used, before the idea of surveys became necessary, in particular betting. In 1549, Matteo Dandolo (Ambassador of Veneto) noted that “it is therefore more than clear that the traders are very well informed of the state of the election, and that the employees of the cardinals in conclave (i conclavisti) participate with them in betting, which therefore leads to several tens of thousands of crowns changing hands” as Baumgartner (2003) tells us. Closer to home, betting markets during the elections were popular in the United States until the Second World War. Rhode & Strumpf (2008) suggests several reasons for the loss of interest in the second half of the 20th century: improvements in sampling techniques… and the legalization of horse betting. But online betting sites have revived interest in betting, whatever it may be. Because the sites we mentioned in a previous article are often not limited to sports betting, but also allow betting on a magnitude earthquake, an Oscar winner, or even the observation of the Higgs boson, as proposed by intrade.com, which was liquidated in 2015. As onlinebettingsites.com shows, we could bet on the French elections in 2017, or on the referendum on Brexit (even if for the latter, the predictive markets were not able to reflect the beliefs of the crowds, as an article in The Economist recalls).

Mathématiques du pari-mutuel

The “pari-mutual” theory is not unlike the mutualisation of risks, an important foundation of the insurance mechanism, dear to actuaries. Working in the horse betting markets, Edmund Eisenberg and David Gale obtained, in a short three-page article, Consensus of Subjective Probabilities, relatively general results, as long as the bet is static.

Supposons que I joueurs puissent parier sur J chevaux. Chaque joueur possède une somme totale bi, que l’on normalisera de telle sorte que bi désigne la part de la somme totale misée (et donc b1 +…+ bI =1). Le joueur i peut alors miser la somme bi,j sur le cheval j (avec ici bi,1+…+bi,J = bi). Lorsque les paris sont clôturés, on note pj le montant parié sur le cheval j, autrement dit b1,j+…+bI,j = pj). La contrainte de budget impose que la somme de ces montants soit égale a 1, ce qui donne aux pj une interprétation probabiliste. Nous reviendrons sur l’utilisation de ces « prix » par la suite. On peut aussi noter qj la cote de gain (payoff-odds) définie comme pj-1-1, de telle sorte que pj=(1+ qj) -1. Si on suppose qu’une portion 1-a est gardée par le bookmaker, alors pj= a(1+ qj) -1 et qj =( a -pj)/ pj. Les rendements espérés sur chacun des chevaux doivent être égaux, à l’équilibre, au rendement net attendu, où l’espérance est calculée sous la probabilité p, de manière à refléter les croyances de tous les parieur, soit ici

pjqj+ (1-pj)(-1)= a-1

The key result of the Eisenberg & Gale model is to show that there is a balance in this market. More precisely, the fraction bet on each horse must be equal to the probability of the horse market. To achieve this balance, it is often assumed that the equilibrium ratings are found by an auctioneer (this role will be played by the bookmaker). As Blough (2008) noted, the hypothesis that no wagering is made until the odds are balanced is a hypothesis that is indeed true in horse racing.

If we assume that each bettor is risk neutral (and seeks to maximize his expectation of winning) and that his beliefs are materialized by a probability vectors pi=(pi1,…,piJ) – in the sense that player i thinks that horse j will win with a probability pij – then at equilibrium, if bi,j >0,

pij=pj max{pis/ps}

where argmax{pis/ps}= argmax{pis(qs+1)}

s the horse on which bettor i must bet everything if he bets on a single horse. Blough (2008) elaborates at length on the description of this balance, and extends it to the case where agents potentially have risk aversion (but the same) and potentially different beliefs. This balance is then described as a consensus of belief.

In an article entitled Interpreting the Predictions of Prediction Markets, Charles Manski proposed using this theory to interpret these prices, in conjunction with more traditional approaches in economics, such as Arrow-Debreu prices.

To illustrate this consensus, let us consider a world cup final that should lead either to the victory of A or the victory of B. Let us imagine a contract offering $1 if A wins, and let this contract be offered at price pA. Si on n’autorise pas d’arbitrage, on a une loi du prix unique, et on en déduit que pB = 1-pA. Imaginons un joueur qui pense que la probabilité que A gagne est supérieure à pA, soit, avec les notations précédentes, piA > pA. Dans ce cas, le joueur a intérêt à parier tout son agent sur la victoire de A, c’est-à-dire acheter ce contrat. La demande agrégée pour ce titre sera alors

[b1P[p1A > pA]+…+ bI P[pIA > pA]] / pA

et on aura un équilibre si la demande agrégée pour les deux titres vérifie

[b1P[p1A > pA]+…+ bI P[pIA > pA]] / pA  

= [b1P[p1A < pA]+…+ bI P[pIA < pA]] / pB

 de telle sorte que

pA = b1P[p1A > pA]+…+ biP[piA > pA]] +…+ bIP[pIA > pA]

which allows the prize to be written as an average of the players’ beliefs.

It should be noted here that the balance is static, allowing the bookmaker to just set a rating. Recently, Agrawal et al (2014) proposed an algorithm to balance this market in continuous time. It may also be noted that this notion of equilibrium appears in many algorithms, such as in the so-called Fisher market.

The predictive power of prices

But this idea of seeing in the prices an aggregation of players’ beliefs is not new! In 1655, in Van Rekeningh in Spelen van Gelucken, (published in Latin under the title’De Ratiociniis in Aleæ Ludo’), Christiaan Huyghens proposed to extract information on beliefs from prices. In 1671, Wilhelmina de Witt noted that, as the price of a contract paying an annuity until death could be seen as a weighted average of annuities (with a fixed maturity), by observing the prices of the different insurance contracts, probabilities interpreted as probabilities of survival could be extracted.

These probabilities are “subjective” as Bruno de Finetti or Frank Ramsey called them. The latter did not see probabilities from a frequentist angle, but saw it as a measure of the degree of belief, which could be measured through bets, in Truth and Probability (1926). This is finally what the theory presented by Kenneth Arrow in 1953, and further developed by Gérard Debreu in 1959, introducing the “Arrow-Debreu prices”, says.

Many websites use odds to infer players’ beliefs, which are misrepresented as the probability that a team will win a competition. We can also note the work carried out last summer by doctoral students at the University of Rennes who had compared the odds on online betting sites, and the forecasts obtained by several algorithms (ranging from a naive Bayesian classifier to boosting, SVM or neural networks). A special issue of The Economist, published in 2007, entitled The Future of Futurology, noted that “the most heeded futurists these days are not individuals, but prediction markets, where the informed guesswork of many is consolidated into hard probability”. This idea has now largely returned to the forefront, as predicted in the article by Chen & Pennock (2010) published in AI Magazine.

Agrawal, Shipra, Delage, Erick, Peters, Mark, Wang, Zizhuo & Ye, Yinyu (2014). A Unified Framework for Dynamic Prediction Market Design. Operations Research.

Baron, Ken & Lange, Jeffrey (2006). Parimutuel Applications In Finance: New Markets for New Risks. Springer.

Baumgartner, Frederic (2003) Behind locked doors: a history of papal elections. Palgrave.

Blough, Stephen R. (2008) Differences of opinion at the racetrack. In Efficiency of Racetrack Betting Markets, 323-341, World Scientific.

Chen, Yiling & Pennock, David (2010). Designing Markets for Prediction. AI Magazine.

Decker, Wolfgang & Thuillier, Jean-Paul (2004). Le sport dans l’antiquité. Picard.

Eisenberg, Edmund & Gale, David (1959). Consensus of Subjective Probabilities: The Pari-Mutuel Method. Annals of Mathematical Statistics, 30:1, 165-168.

Griffith, RM (1949) Odds adjustments by American horse-race bettors. The American Journal of Psychology, 62, 290-294.

Manski, Charles (2005) Interpreting the Predictions of Prediction Markets. NBER 10359.

Rhode, Paul, W. & Strumpf, Koleman (2008) Historical Political Futures Markets: An International Perspective. NBER 14377.

[1] Baron & Lange (2006) discusses the comparison between so-called “risk-neutral” valuation in finance (based on the law of single price and arbitrage), and that relating to mutual betting. They thus speak of “self-hedging” because, in a bet, the bettors share the winnings in proportion to their initial bet. This is reminiscent of the way mutual insurance companies operate, where the money used to compensate victims must correspond to the total premiums charged.

 

An Open Lab-Notebook Experiment