14 Big Data in Economics: Evolution or

Revolution?

Christine De Mol, Eric Gautier, Domenico Giannone,

Sendhil Mullainathan, Lucrezia Reichlin, Herman van
Dijk and Jeffrey Wooldridge

Abstract
The Big Data Era creates a lot of exciting opportunities for new developments
in economics and econometrics. At the same time, however, the analysis of
large datasets poses difficult methodological problems that should be addressed
appropriately and are the subject of the present chapter.

14.1 Introduction
‘Big Data’ has become a buzzword both in academic and in business and policy
circles. It is used to cover a variety of data-driven phenomena that have very
different implications for empirical methods. This chapter discusses some of
these methodological challenges.1
In the simplest case, ‘Big Data’ means a large dataset that otherwise has
a standard structure. For example, Chapter 13 describes how researchers are
gaining increasing access to administrative datasets or business records cov-
ering entire populations rather than population samples. The size of these
datasets allows for better controls and more precise estimates and is a bonus
for researchers. This may raise challenges for data storage and handling, but
does not raise any distinct methodological issues.
However, ‘Big Data’ often means much more than just large versions of stan-
dard datasets. First, large numbers of units of observation often come with large
numbers of variables, that is, large numbers of possible covariates. To illustrate
with the same example, the possibility to link different administrative datasets
increases the number of variables attached to each statistical unit. Likewise,
business records typically contain all consumer interactions with the business.
This can create a tension in the estimation between the objective of ‘letting
the data speak’ and obtaining accurate (in a way to be specified later) coeffi-
cient estimates. Second, Big Data sets often have a very different structure from
those we are used to in economics. This includes web search queries, real-time
612

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 Big Data in Economics: Evolution or Revolution? 613

geolocational data or social media, to name a few. This type of data raises ques-
tions about how to structure and possibly re-aggregate them.
The chapter starts with a description of the ‘curse of dimensionality’, which
arises from the fact that both the number of units of observation and the number
of variables associated with each unit are large. This feature is present in many
of the Big Data applications of interest to economists. One extreme example of
this problem occurs when there are more parameters to estimate than observa-
tions. In this case, standard estimators (such as ordinary least squares) do not
yield a unique solution. The section, which borrows heavily from De Mol et al.
(2008), describes the econometric problems raised by the curse of dimension-
ality. It describes some of the methodological solutions called regularization
methods that have been proposed.
Section 14.3 then discusses recent research on recovering policy effects using
Big Data. In many fields of economics, we are interested in measuring a (causal)
relationship between some variable of interest (for example, a policy) and its
effects. In other words, although there might be many variables, some of them
(related to a specific policy) are of special interest to the researcher. The section
describes current efforts to develop methods that combine the ability of regu-
larization methods to harness the information contained in these richer datasets
with the possibility to identify the impact of specific policy relevant effects.
Section 14.4 turns to prediction problems. Here we are not interested in
specific coefficients per se but in our ability to forecast a variable of interest,
for example, inflation, growth or the probability of default. Forecasting has a
long tradition in macroeconomics and the greater availability of highly gran-
ular microdata is creating renewed interest in prediction problems also at the
microeconomic level. A priori, regularization methods are well-suited for this
type of problem. However, ‘off-the-shelf’ regularization methods are agnos-
tic regarding the data generation process. On the basis of the experience with
macro forecasting models, the section argues for the need to develop regulariza-
tion methods that account for the specificities of the data generation processes
in economics, such as serial correlation or mixed frequencies.
Recent progress in computing power and storage capacities has allowed
researchers to handle and analyse increasingly big datasets. For some of the
Big Data (e.g., high frequency trading data, browsing data), this may not be
enough. Section 14.5 discusses how simulation-based methods can be refined
to leverage the potential of parallel computing.
Section 14.6 concludes. The availability of unprecedented amounts of data
offers exciting research opportunities in economics. While researchers will be
able to exploit some of the methods developed in other fields, such as statistics
and computer science, it is essential that some of these methods be tailored to
the specificities of economic research questions and economic data. On these
fronts, there is still much to be done.

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 614 De Mol, Gautier, Giannone, Mullainathan, Reichlin et al.

14.2 The Curse of Dimensionality and Regularization

An early occurrence of the term ‘Big Data’ in economics is to be found in a
discussion by Diebold (2003, 2012). To quote, ‘I stumbled on the term Big Data
innocently enough, via discussion of two papers that took a new approach to
macro-econometric dynamic factor models (DFMs), Reichlin (2003) and Wat-
son (2003), presented back-to-back in an invited session of the 2000 World
Congress of the Econometric Society.’
The two authors referenced above were presenting their research on factor
models in high-dimensional time series (Forni et al., 2000, Stock and Watson,
2002), which mainly consisted in deriving asymptotic results for the case where
both the number of time samples and the cross-sectional dimension, that is, the
number of time series, tend to infinity. The approach relied on a factor model
dating back to Chamberlain and Rothschild (1983) in finance, but generalized
to take serial correlation into account. Stock and Watson (2002) considered so-
called ‘static’ factor models, whereas Forni et al. (2000) derived asymptotics
in the case of ‘dynamic’ factor models. The estimators they used are based
on a few principal components either in the time domain for the static case
or in the Fourier domain for the dynamic case. This factor-model literature
was probably the first in economics to address the difficulties arising from the
high dimensionality of the data, albeit under rather strong assumptions (namely
factor models) on the data generating process.
In statistics, the difficulties pertaining to the analysis of high-dimensional
data are well-known issues, often referred to as the ‘curse of dimensionality’.
Some of the facets of this curse can be explained using the familiar example of
the linear regression model. To introduce some background notation useful for
our discussion, let

Y = Xβ + U, (14.1)

where X is a n × p matrix containing the observed predictors (covariates), Y is

the outcome or n × 1 vector of the observed responses and U is an unobserved
zero-mean error or nuisance term. The p × 1 vector β contains the regression
coefficients. In the case of time series, n is the number of time samples and p
is the number of time series used for prediction. In the case of cross-section
data, n is the number of observations and p is the number of covariates. In the
discussion in this section, we will consider the matrix X as deterministic.
Depending on the application under study, two different problems can be
highlighted. The first is prediction (also referred to as ‘generalization’ by the
machine-learning community), in which case one is only interested in esti-
mating the outcome for future times or new examples to come. This requires
the estimation of the regression parameters, but only as an auxiliary step to
the estimation of the outcome. The second problem, the identification of the

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 Big Data in Economics: Evolution or Revolution? 615

model, pertains more to the vector β of regression coefficients itself, in the lin-
ear regression example in (14.1). This is essential for interpreting the estimated
coefficients in terms of their relevance in predicting the response. For exam-
ple, some coefficients can be zero, indicating that the corresponding predictors
are not relevant for this task. The determination of these zeroes, hence of the
relevant/irrelevant predictors, is usually referred to as ‘variable selection’.
As is well known, the most straightforward solution for the linear regression
problem is Ordinary Least Squares (OLS). The OLS estimator for β in (14.1)
minimizes the least-squares loss
(β ) = Y − Xβ 2
2 , (14.2)
n
where Y 2 = i=1 |Yi |2 is the L2 -norm of the vector Y. It is given by
β̂ols = (X X)−1 X Y (14.3)
(X denotes the transpose of the matrix X).
For the OLS estimator, expression (14.3), to make sense we need the p × p
matrix X X to be of full rank, hence invertible. This cannot be the case in high-
dimensional situations where the number of coefficients, p, is larger than the
number of observations, n.2 In that case, the minimizer of the least-squares
loss is nonunique, but uniqueness can be restored by selecting the so-called
‘minimum-norm least-squares solution’, orthogonal to the null-space, that is,
by ignoring the subspace corresponding to the zero eigenvalues.
Notice that although this remedy may work well for prediction, the identifica-
tion problem remains hindered by this nonuniqueness issue. An additional diffi-
culty arises when the matrix X X has eigenvalues that are close to zero, or more
precisely, when its ‘condition number’, that is, the ratio between the largest and
the smallest of its nonzero eigenvalues, becomes large. This situation prevents
a stable determination of the least-squares (or minimum-norm least-squares)
estimator: small fluctuations in the outcome vector Y will be amplified by the
effect of the small eigenvalues and will result in large uncontrolled fluctuations
(high variance/volatility) on the estimation of β, again preventing meaningful
identification.
The pathologies described above contribute to what is often referred to as the
‘curse of dimensionality’ or else the ‘large p, small n paradigm’ in high-dimen-
sional statistics. As early as in the 1950s, Stein (1956) introduced a ‘high-di-
mensional’ surprise in statistics by showing that the maximum-likelihood esti-
mator of the unknown mean vector of a multivariate Gaussian distribution is not
‘admissible’ in a dimension higher than three, that is, that it is outperformed by
‘shrinkage’ estimators. Heuristically, ‘shrinking’ means that a naive estimate
is improved by combining it with other information or priors.
Many remedies have been proposed to address these pathologies under the
common designation of ‘regularization methods’, which provide in one form or

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 616 De Mol, Gautier, Giannone, Mullainathan, Reichlin et al.

Box 14.1: Principal Component Regression (PCR)

The Principal Component Regression consists in estimating β by
k
X Y, Vi
β̂ pcr = Vi (14.4)
i=1
ξi2

where the Vi ’s are the eigenvectors of X X with eigenvalues ξi2 , and ·, ·

denotes the scalar product.

another the dimensionality reduction necessary to reduce the variance/volatility

of unstable estimators, or in other words, to avoid ‘overfitting’. Overfitting
refers to the fact that, when using a model with many free parameters (here
the p components of β), it is easy to get a good fit of the observed data, that is,
a small value for the residual (14.2), but that this does not imply that the corre-
sponding (unstable) value of β will have a good predictive power for responses
corresponding to new observations. For time series, good in-sample fit does not
imply good out-of-sample forecasts.
One of the simplest regularization methods is principal component regres-
sion (PCR), a statistical procedure that transforms the possibly correlated vari-
ables into a smaller number of orthogonal new variables (the components, see
Box 14.1). The truncation point k for the number of components, usually much
smaller than the true rank of X X, has to be carefully chosen to overcome insta-
bilities. In this method, also referred to as ‘Truncated Singular Value Decom-
position’ (TSVD), the truncation introduces a bias in order to reduce variance.
Until recently alternative estimators were less well known in econometrics.
Other regularization methods introduce constraints or penalties on the vector β
of the regression coefficients. Probably the oldest penalized regression method
is ‘Ridge regression’(see Box 14.2), due to Hoerl and Kennard (1970). This
method is also known in the applied mathematics literature as Tikhonov’s reg-
ularization. It consists in adding to the least-squares loss a penalty proportional
to the size of β, measured by its squared L2 -norm. As for the truncation point in
PCR, the regularization parameter has to be chosen carefully in order to provide
a proper balance between the bias introduced by shrinkage, and the variance of
the estimator and its value is usually determined by cross-validation.
Ridge regression introduces a form of linear ‘shrinkage’, where the compo-
nents of β̂ols are shrunk uniformly towards zero, as can be easily seen in the case
of orthonormal regressors (i.e., for X X = I), where β̂ridge = 1+λ 1
2
X Y. More
generally, quadratic penalties provide estimators which depend linearly on the
response Y but do not allow for variable selection, since typically all regression
coefficients are different from zero.

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 Big Data in Economics: Evolution or Revolution? 617

Box 14.2: The Ridge Regression Estimator

The ridge regression estimator is given by
β̂ridge = argminβ Y − Xβ 2
2 + λ2 β 2
2 (14.5)
= (X X + λ2 I)−1 X Y
where I is the identity matrix and λ2 > 0 is the so-called ‘regularization
parameter’, which, as seen from (14.5), reduces the impact of the smallest
eigenvalues of X X, at the origin of the instability of the OLS estimator.

An alternative to quadratic penalties that allows for variable selection by

enforcing sparsity, that is, the presence of zeroes in the vector β of the regres-
sion coefficients, has been popularized in the statistics and machine-learning
literature under the name of ‘Lasso regression’ by Tibshirani (1996). It con-
sists in replacing the L2 -norm penalty used in ridge regression by a penalty
proportional to the L1 -norm of β (see Box 14.3).
In the case of orthonormal regressors, it is easily seen that the Lasso penalty
provides a nonlinear shrinkage of the components of β̂ols , which are shrunk
differently according to their magnitude, as well as sparsity, since the jth coef-
ficient [β̂lasso] j = 0 if |[X Y] j | < λ1 /2. Unfortunately, there is no closed-form
expression for β̂lasso in the case of general matrices X, and the Lasso estima-
tor has to be computed numerically as the solution of a (nonsmooth) convex
optimization problem.
The previous estimators can be given a Bayesian interpretation, since β̂ols
can be viewed as the maximum (log-)likelihood estimator for a Gaussian error
term and the penalized maximum likelihood estimators β̂ridge and β̂lasso can be
interpreted as maximum a posteriori (MAP) estimators, the penalty resulting
from a prior distribution for the regression coefficients. In Ridge regression, it
corresponds to a Gaussian prior whereas in Lasso regression it is a Laplacian
or double-exponential prior.
The regularization techniques described above are paradigmatic since they
convey the essential ideas in dealing with high-dimensional settings. There are
however numerous extensions and generalizations. For example, more general
types of penalties can be used such as β γγ = pj=1 |β j |γ , i.e., the Lγ -norms
used in ‘bridge regression’ (Frank and Friedman, 1993). Notice that in this fam-
ily, though, only the choice γ = 1 yields both convexity and sparsity. Moreover,
weights or even a nondiagonal coupling matrix can be introduced in the penalty
to cover the case of non i.i.d. (independent and identically distributed) regres-
sion coefficients. Composite penalties are also used, for example, in elastic-net
or group-lasso regularization. Finally, different loss functions can be considered
such as those used in robust statistics, logistic regression, etc. A good pointer
to this variety of techniques is the book by Hastie et al. (2009).

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 618 De Mol, Gautier, Giannone, Mullainathan, Reichlin et al.

Box 14.3: The Lasso Regression Estimator

Lasso consists in replacing the L2 -norm penalty used in ridge regression
by a penalty proportional to the L1 -norm of β, that is, to the sum of the
absolute values of the regression coefficients, β 1 = pj=1 |β j |, yielding
the estimator
β̂lasso = argminβ Y − Xβ 2
2 + λ1 β 1 , (14.6)
In the special case of orthonormal regressors (X X = I), the Lasso estimator
is easily seen to be given by
[β̂lasso] j = Sλ1 ([X Y] j )
where Sλ1 (x) is the ‘soft-thresholder’ defined by
⎧
⎨ x + λ1 /2 if x ≤ −λ1 /2
Sλ1 (x) = 0 if |x| < λ1 /2
⎩
x − λ1 /2 if x ≥ λ1 /2 .

Let us remark that global variable selection methods, preferably convex to

facilitate computation, such as the Lasso and its relatives, are essential to deal
with high-dimensional situations. Indeed, considering all possible submodels
and selecting the best among them, for example according to the Akaike Infor-
mation Criterion (AIC) proposed by Akaike (1974) or the Bayesian Information
Criterion (BIC) proposed by Schwarz (1978), leads to a complexity growing
exponentially with the number of variables involved and renders the methods
totally unpractical. To paraphrase the title of a paper by Sala-I-Martin (1997):
‘You cannot just run two million regressions!’ (and, incidentally, two million
would not even suffice for p ≥ 22).
As concerns asymptotic and consistency results, the settings have to go
beyond the classical scheme of keeping the number of parameters p constant
(and usually small), while letting the number of observations n of the dependent
variable tend to infinity. In high-dimensional situations, both n and p may tend
to infinity, while assuming or not some relationship between their respective
growth rates. The theory is more subtle in this case and is still developing. This
question has been studied for principal component regression for time series
under a factor model assumption. Results in this line for the case of penalized
regression, and in particular of Ridge regression, have been derived by De Mol
et al. (2008, 2015). The first paper also contains an empirical part where pre-
dictive accuracy of PCR, Ridge and Lasso regression is evaluated based on
a dataset of about 100 time series. It is shown that all three methods perform
similarly and that results of Lasso are uninformative when used for applications

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 Big Data in Economics: Evolution or Revolution? 619

where, as is typically the case for macroeconomics, data are cross-correlated.

Moreover, in that case Lasso is unstable in selection.

14.3 Policy Analysis and Causal Inference

In the actual big data activity sphere, in parallel with the developments of pow-
erful machine-learning techniques, the emphasis is on predictive rather than
causal models. As we shall further discuss in the next section, successful pre-
dictive algorithms are rapidly developing in response to the increasing demand
coming from all kinds of applications. These algorithms convert large amounts
of unstructured data into predictive scores in an automatic way and often in real
time.
Whether this trend is desirable may be a matter of debate but it is clear
that it implies a significant shift from the single-covariate causal-effect frame-
work that has dominated much empirical research, especially in microe-
conomics. Being nonstructural, predictive models are subject to the Lucas
critique (Lucas, 1976) and their success should not obscure the fact that many
economic applications are about inference on a causal effect. In microeco-
nomics, for example, a successful literature has developed methods to assess
the effectiveness of a given policy or treatment.
In the case where the intervention is binary in nature, we define a binary
variable, W , equal to unity for the treatment group and zero for the control
group. We typically have in mind a counterfactual setting, where it makes sense
to think of potential outcomes for each unit in the control and treated states.
These outcomes are often denoted Y (0) and Y (1), and then we observe the treat-
ment status, W , and the outcome under the corresponding treatment status, Y =
(1 − W )Y (0) + WY (1) = Y (0) + W [Y (1) − Y (0)]. For unit i in the popula-
tion, the treatment effect is Yi (1) − Yi (0) – which is not observed. Instead, atten-
tion typically centres on the average treatment effect, τ = E [Y (1) − Y (0)], or
the average over an interesting subpopulation (such as those actually receiv-
ing the treatment). A special case is when the treatment effect is constant, in
which case we can write Y = τW + Y (0), and the Y (0) plays the role of the
unobserved factors affecting Y .
The potential outcomes setting can be extended to cases where the policy
variable, W , is not binary. If the policy effect is constant across units and across
levels of the treatment, we can write a simple regression equation

Y = τW + R , (14.7)

where Y, W and R are random variables and τ is a scalar coefficient of interest.

We (eventually) observe data on Y and W . The variable R includes unobserved
factors – Y (0) in the simplest setting – affecting Y .

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 620 De Mol, Gautier, Giannone, Mullainathan, Reichlin et al.

In medicine and the experimental sciences, truly randomized experiments

can be carried out, which means the treatment level W can be made independent
of R. For example, when W is binary, we can randomly assign individuals into
the treatment and control groups. In such cases, (14.7) can be estimated using
simple regression, which delivers an unbiased and consistent estimator of τ . In
economics, random assignment is much less common, and in general one has
access only to so-called observational – not experimental – data. Hence, several
strategies, when randomized assignment is not available have been developed.
Here we review a few of those strategies, highlighting how high-dimensional
regression methods can be applied to estimating causal effects. Good pointers
to part of the relevant work in this field are the review papers by Belloni et al.
(2013, 2014a).
A traditional and still commonly used method to handle nonrandom treat-
ment assignment is regression adjustment, where one assumes the availability
of covariates that render the policy assignment appropriately ‘exogenous’. Let
X be a 1 × p vector of covariates. Then, if X is thought to predict both Y and
the treatment assignment, W , we can ‘control for’ X in a multiple regression
analysis. This leads to a linear model,
Y = τW + Xβ + U, (14.8)
where now R = Xβ + U and, if the elements of X suitably control for the non-
random assignment, the treatment and covariates satisfy the exogeneity condi-
tions
E[WU] = 0, E[XU] = 0. (14.9)
If p, the number of control variables, is large and the p × 1 vector β is sparse,
the model can be estimated by means of a Lasso regression, as described in the
previous section. However, one has to know in advance the right vector X such
as, under the usual exogeneity conditions (14.9), there are no more confounding
variables and one recovers the marginal effect τ , holding fixed everything else.
One can relax the linearity assumption in X and just assume
E[R|W, X] = E[R|X], (14.10)
which yields
Y = τW + g(X ) + U, (14.11)
where g(X ) = E[R|X] and U = R − E[R|W, X] is such that E[U|W, X] = 0.
Model (14.11) is a so-called partially linear model and g is generally a non-
parametric function. Belloni et al. (2014b) use Lasso-type methods to nonpara-
metrically partial out X from both Y and W . They approximate the mean func-
tions E[Y |X] and E[W |X] using functions of the form pj=1 β j φ j (X ), for a large
dictionary of functions (φ j ) pj=1 , and build a confidence interval around τ . This

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 Big Data in Economics: Evolution or Revolution? 621

method is particularly appealing as it does not require one to chose a bandwidth

to estimate the nonparametric conditional mean functions. If the approxima-
tions pj=1 β j φ j (X ) are sparse, then the method selects the significant φ j (X )
for each of E[Y |X] and E[W |X]. As shown in Belloni et al. (2014a), using the
union of the functions selected from the methods in a standard regression anal-
ysis with Y as the response variable and W as the other regressor, the usual
heteroscedasticity-robust standard error produces valid t statistics and confi-
dence intervals. It should be emphasized that, while the approach works well
for selecting functions of X that appear in the conditional mean, it does not
select the variables X such as (14.10) holds; the researcher is assumed to have
already selected the appropriate controls.
When (14.9) or (14.10) do not hold, we can rely on instrumental variables,
namely, assume to have at our disposal a vector of random variables Z, called
instrumental variables, such as in (14.7),
Cov[Z, R] = 0. (14.12)
This yields the relation
Cov[Z, Y ] = τ Cov[Z, W ]. (14.13)
If Z is a scalar, (14.13) identifies τ when Cov[Z, W ] = 0. When we have more
than one instrumental variable for W , two stage least squares (2SLS) is a com-
mon estimation approach. However, 2SLS only uses the linear projection of
W on Z in forming instruments. If we strengthen the exogeneity requirement
to E[R|Z] = 0, 2SLS is asymptotically inefficient if E[W |Z] is nonlinear or
Var[R|Z] is not constant. If we assume homoscedasticity in (14.7), that is,
Var[R|Z] = Var[R], then the optimal instrument for W is given by E[W |Z], the
best mean square predictor of W . Belloni et al. (2012) propose to use Lasso-type
methods to estimate the regression function E[W |Z] using a large dictionary of
approximating functions, and they show how to conduct inference on τ using
a heteroscedastic-robust standard error.
Gautier and Tsybakov (2011) propose an instrumental variables method to
make inference for high-dimensional structural equations of the form
Y = Xβ + U (14.14)
where the dimension of X is large (and may include exogenous and endogenous
variables). This occurs, for example, in large demand systems, or when treat-
ment variables are interacted with (exogenous) group dummies. In the latter
case, the policy might have an effect on only certain groups, and the policy-
maker would like to determine for which group the policy has an effect. The
instrumental variables literature has identified various problems: (i) the instru-
mental variable candidates, Z, might not be exogenous; (ii) the instrumental
variables can be ‘weak’ and estimating in a first-stage a reduced form equation

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 622 De Mol, Gautier, Giannone, Mullainathan, Reichlin et al.

can yield multimodal and non-normal distributions of the parameter estimates,

even with very large sample size, so that asymptotic theory is not reliable;
(iii) in the presence of many instrumental variables, estimating in a first-stage
a reduced form equation can give rise to a large bias. Gautier and Tsybakov
(2011) rely on a new method which is robust to (ii) and (iii) in order to treat the
more challenging case of a high-dimensional structural equation. Confidence
sets can be obtained for arbitrary weak and numerous instrumental variables,
whether or not the condition Cov[Z, U] = 0 gives rise to a unique β. Therefore,
it is also possible to handle the case where the dimension of Z is smaller than
the dimension of X, which can yield the identification of β under sparsity of the
structural Equation (14.14) or other shape restrictions. To deal with the possi-
bility of (i), a high-dimensional extension of the Sargan and Hansen method is
developed.
There is much interest in the literature on heterogeneous treatment effects,
and variable selection methods can also be applied in such cases. For exam-
ple, variable selection methods can be applied to estimating the propensity
score when the treatment variable takes on a small number of levels. Moreover,
variable selection methods can be applied to both propensity score estimation
and regression function estimation to obtain so-called doubly robust estima-
tors. Unlike the linear, additive Equation (14.11), methods that weigh by the
inverse of the propensity score allow for heterogeneous treatment effects. See,
for example, the papers by Farrell (2015) and Athey and Imbens (2015).
Besides these high-dimensional problems, let us mention another important
issue which arises in connection with the availability of big datasets, namely,
to determine whether the accumulation of data, say, over an entire population,
affects the precision of estimates. Abadie et al. (2014) analyse how to compute
uncertainty in empirical situations where the sample is the entire population
and where the regression function is intended to capture causal effects. Other
contributions on Causal Inference in a big data setting use machine-learning
methods. A recent example is the work by Athey and Imbens (2015). There
are many open challenges in this area, pointers to recent progress are available
from the site of the Sackler Colloquium on ‘Drawing Causal Inference from
Big Data’, organized in March 2015 at the US National Academy of Science
in Washington.3
The previous discussion has focused on cross-sectional data, but empiri-
cal researchers attempting to estimate causal effects often rely on panel data
that exploit changes in policies over time. An important component of panel
data models is allowing for time-constant, unobserved heterogeneity. Belloni
et al. (2014a) propose first differencing a linear unobserved effects equation to
remove additive heterogeneity, and then using variable selection methods, such
as Lasso, to allow for correlation between unobserved heterogeneous trends and
unknown functions of observed covariates – including the policy variable or

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 Big Data in Economics: Evolution or Revolution? 623

variables being studied. The approach seems promising. So far, such methods
have been applied to linear models with relatively few sources of heterogeneity.

14.4 Prediction
Despite recent advances in identification and causality in big data settings,
which we have just reviewed, it is fair to say that the literature in the field is
mainly focused on prediction. Using the same notation as above, the problem
consists in computing the conditional expectation
E(Y |W, X ) . (14.15)
Forecasting has a long tradition in economics, especially in macroeconomics.
Indeed, many economists in the private sector and policy institutions are
employed for this task. In forecasting, robustness is typically tested in out-of-
sample validation studies, a perspective typically ignored in empirical microe-
conomics. For desirable out-of-sample performance, models must respect the
principle of parsimony (i.e., contain a rather small number of free parameters)
to avoid overfitting. However, the curse of dimensionality problem naturally
arises from lags, nonlinearities, and the presence of many potentially relevant
predictors.
The recent literature has suggested methods to deal with the curse of dimen-
sionality issue in dynamic models. Here we should mention dynamic factor
models cited earlier and, more recently, large Bayesian vector autoregressive
models. Following the work of De Mol et al. (2008), Banbura et al. (2010) have
shown empirically how to set priors to estimate a vector autoregressive model
with large datasets. The idea is to set the degree of ‘shrinkage’ in relation to
the dimension of the data. Intuitively this implies to set priors so as to avoid
overfitting, but still let the data be informative. Giannone et al. (2015) have
developed a formal procedure to conduct inference for the degree of shrinkage.
These models have many applications in economics beyond pure forecasting
and can be used to design counterfactuals for policy analysis and identification
of exogenous shocks and dynamic propagation mechanisms. Large data allow
to better identify exogenous shocks since they can accommodate for realistic
assumptions on agents’ information set (for an analysis on this point, see Forni
et al. (2009)).
One very successful application of the large models described above (if mea-
sured by impact on modelling in policy institutions and the financial indus-
try) has been ‘now-casting’. Now-casting is the forecast of the present (present
quarter or present month) based on data available at different frequencies (daily,
weekly, monthly and quarterly). A now-cast produces a sequence of updates
of predictions in relation to the publication calendar of the data. This allows to
exploit the timeliness of certain data releases to obtain an early estimate of those

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 624 De Mol, Gautier, Giannone, Mullainathan, Reichlin et al.

series which are published with a delay with respect to the reference quarter
such as GDP or the reference month such as employment (see Giannone et al.,
2008 and subsequent literature). Empirical results show that exploiting survey
data, which are published earlier than hard data, allows to obtain an accurate
early estimate at the beginning of the quarter and, as new data are released
through time, the estimates become more accurate (see Banbura et al., 2013
for a review of the literature). In principle, nonstandard data such as Google
queries or twitters, due to their timeliness, could be exploited in this context.
However, once the details of the problem (mixed frequency, nonsynchronous
data releases) are appropriately modelled and relevant timely indicators con-
sidered, there is no evidence that Google indexes used successfully in a sim-
pler setup by Choi and Varian (2012) and Scott and Varian (2014) have any
additional predictive value (see Li, 2016), but more research is needed on this
topic.
It has to be noted that most of the applied work on the methods mentioned
have concerned traditional time series (macroeconomic variables, possibly dis-
aggregated by sectors or regions, financial variables and surveys) and rarely
with dimension above 150. Empirical results show that, in general, forecasts
of macroeconomic variables based on datasets of medium dimension (of the
order of 20) are not outperformed by forecasts based on 100 or more vari-
ables although the dimension helps especially in now-casting where success-
ful results rely on the use of timely information. Moreover, as mentioned in
Section 14.2, Lasso regressions provide unstable variable selection due to the
near-collinear feature of macroeconomic data. Important empirical issues are
also related to robustness with respect to variable transformation such as de-
seasonalization or detrending as well as nonlinearity. Potentially, machine-
learning type of techniques could be useful in this setup but this is open to future
research. The literature is at too early stage to provide a definitive answer on
the potentials of new data and new methods in this context but it is our view that
any successful applications have to incorporate the detailed micro-structure of
the problem. In now-casting, for example, this implies taking care of mixed
frequency, the nonsynchronicity of releases and other details.
In microeconometrics the emphasis on predictions and out-of-sample is
newer than in macro but studies using big data are more numerous. Predic-
tions based on a large cross-section of data have been successfully obtained for
various problems. Examples can be found in papers by Varian (2014), by Einav
and Levin (2014) and by Kleinberg et al. (2015b), as well as in the references
therein. The last paper discusses a problem of health economics, namely the
prediction of whether replacement surgery for patients with osteoarthritis will
be beneficial for a given patient, based on more than 3000 variables recorded for
about 100,000 patients. Another policy decision based on prediction is studied
by Kleinberg et al. (2015a), who show that machine-learning algorithms can be

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 Big Data in Economics: Evolution or Revolution? 625

more efficient than a judge in deciding who has to be released or go to jail while
waiting for trial because of the danger of committing a crime in the meanwhile.
Another application would be to predict the risk of unemployment for a given
individual based on a detailed personal profile.
It should be remarked that machine-learning algorithms present several
advantages: they focus on a best-fit function for prediction, possibly handling
very rich functional forms, and have built-in safeguards against overfitting so
that they can handle more variables than data points. Moreover, they do not
require too many assumptions about the data generating process as it is the case
in classical econometrics. We should be aware, however, that precisely because
of their great generality and versatility, they may not be optimally tailored for
the specificities of a given problem.
Another trend is to make use not only of more data, but also of new types
of data. Many types of data are nowadays passively collected and are largely
unexploited, such as those provided by social networks, scanner data, credit
card records, web search queries, electronic medical records, insurance claim
data, etc. They could complement more traditional and actively collected data
or even be a substitute for them. The mining of language data, such as online
customer reviews, is also a challenge and can be used for so-called ‘sentiment
analysis’ (see e.g., Pang et al., 2002).
Returning to the issue of causality discussed in the previous section, it should
be noted that prediction algorithms also provide new ways to test theories.
Indeed, we can see how well we can predict the output Y with all variables
but X and/or how much the inclusion of a given variable (or a group of vari-
ables) X helps improving the prediction. We should be cautious, however, in
drawing conclusions: the fact that a variable is not among the best predictors
does not necessarily mean that it is not ‘important’. For example, when Var-
ian (2014) discusses differences in race in mortgage applications, saying that
race is not among the best predictors is not the same as saying that evidence of
discrimination does not exist.
In addition, the completeness of a given theory could be tested by confronting
its prediction abilities against an atheoretical benchmark provided by machine
learning.

14.5 Computational Issues

The collection and analysis of bigger and bigger datasets obviously pose
methodological as well as computational challenges. Nevertheless, since there
has been at the same time a tremendous increase in computing capabili-
ties, researchers can handle larger and larger datasets using standard software
and desktop computers. For example, up to the late 90s maximum-likelihood
estimation of dynamic factor models could be performed only with a small set

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 626 De Mol, Gautier, Giannone, Mullainathan, Reichlin et al.

of variables (Stock and Watson, 1989), while recent research has shown how
these models can be easily estimated in a high-dimensional context (Doz et al.,
2012, Jungbacker and Koopman, 2015). In parallel with this increase in com-
puting power, significant progress has been made in the development of fast
and reliable numerical algorithms which scale well with dimension. In particu-
lar, considerable research effort has been dedicated to improving the speed and
performance of algorithms for Lasso regression.
In many situations, however, computational capability still represents an
important constraint on our ability to handle and analyse big datasets. Meth-
ods that can handle thousands of variables may become inappropriate when
moving to millions of variables. Moreover, some procedures can be particu-
larly demanding in terms of computational complexity when applied to more
than a handful of data. This is the case, for example, for complex latent variable
models for which closed-form solutions are not available. In this context there
is a demand for extra computing power. Unfortunately, the growth rate in com-
putational capability of integrated circuits (CPU microchips) seems to be slow-
ing down. However, thanks to technological progress driven by the video-game
industry, new and fast growing computational power is coming from so-called
graphics processing units (GPU), which allow for parallel computation and are
easy to program. The general idea is that it is often possible to divide large
problems into smaller independent tasks, which are then carried out simultane-
ously.
Splitting large problems into small ones is particularly natural in simulation-
based Bayesian methods, which have recently attracted growing interest (see
e.g., Hoogerheide et al., 2009, Lee et al., 2010, Durham and Geweke, 2013). In
Bayesian methods, the reduction in dimensionality is made by assuming prior
distributions for the unknown parameters to infer and, whereas the computa-
tion of the so-called MAP (Maximum a Posteriori) estimator requires solving
an optimization problem, the computation of conditional means and covari-
ances only requires integration, but in a high-dimensional space. For this task
stochastic simulation methods and artificially generated random variables are
used. Since the early days of Monte Carlo methods, there has been substantial
development of new more sophisticated Sequential Monte Carlo and Particle
Filter methods, allowing us to deal with complex posterior distributions and
more flexible econometric models.
Examples of successful applications of simulation-based Bayesian methods
are reported by Billio et al. (2013a,b) and Casarin et al. (2013, 2015). The paper
by Casarin et al. (2015) deals with the problem of conducting inference on
latent time-varying weights used to combine a large set of predictive densities
for 3712 individual stock prices, quoted in NYSE and NASDAQ, using 2034
daily observations from 18 March 2002 to 31 December 2009. The authors

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 Big Data in Economics: Evolution or Revolution? 627

find substantial forecast and economic gains and also document improvement
in computation time achieved by using parallel computing compared to tradi-
tional sequential computation. Another application to nowcasting is discussed
by Aastveit et al. (2014), who show that a combined density now-cast model
works particularly well in a situation of early data releases with relatively large
data uncertainty and model incompleteness. Empirical results, based on US
real-time data of 120 leading indicators, suggest that combined density now-
casting gives more accurate density now-casts of US GDP growth than a model
selection strategy and other combination strategies throughout the quarter, with
relatively large gains for the first two months of the quarter. The model also
provides informative signals on model incompleteness during recent recessions
and, by focusing on the tails, delivers probabilities of negative growth, that pro-
vide good signals for calling recessions and ending economic slumps in real
time.

14.6 Conclusions
Data are essential for research and policy. Definitely there is a trend towards
empirical economics, and from this perspective, the advent of big data offers an
extraordinary opportunity to take advantage of the availability of unprecedented
amounts of data, as well as of new types of data, provided that there is easy
access to them, in particular for academic research.
In this chapter, we have focused on some methodological aspects of the anal-
ysis of large datasets. We have argued that many of the issues raised by big data
are not entirely new and have their roots in ideas and work over the past decades.
On the applied side, applications with truly big data are still rare in economics
although in recent years more research has been devoted to the use of relatively
large but traditional datasets.
While in many problems the focus is shifting from identification towards
prediction, which is a more ‘revolutionary trend’, causality is still considered
important and this duality is a matter for interesting debates in econometrics.
As concerns algorithmic and computational issues, the field of ‘machine
learning’, a popular heading covering very different topics, is and will remain
helpful in providing efficient methods for mining large datasets. However, we
should be careful rather than blindly import methodologies from other fields,
since economic data structures have their own specificities and need appropri-
ately designed research tools.
Undoubtedly, this research area calls for a lot of new, exciting and perhaps
unexpected developments within and outside the framework sketched here,
and if the datasets are big, the challenges ahead are even bigger, in optimally
exploiting the information they contain.

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 628 De Mol, Gautier, Giannone, Mullainathan, Reichlin et al.

Notes
1. This chapter is based on the presentations given by the authors at the COEURE
workshop on ‘Developments in Data and Methods for Economic Research’ held in
Brussels in July 2015. The presentations took place in two sessions of the work-
shop: ‘Big Data: Definition, challenges and opportunities’, chaired by Christine De
Mol, and ‘How will Big Data change econometrics?’ chaired by Domenico Gian-
none. Christine De Mol coordinated and integrated the authors’ presentations to the
chapter.
2. Since this implies the existence of a nontrivial null-space for X X, with at least p − n
zero eigenvalues.
3. See http://www.nasonline.org/programs/sackler-colloquia/completed_colloquia/
Big-data.html.

References
Aastveit, K. A., Ravazzolo, F., and van Dijk, H. K. 2014 (Dec). Combined Density Now-
casting in an Uncertain Economic Environment. Tinbergen Institute Discussion
Papers 14-152/III. Tinbergen Institute. Accepted for publication in Journal of Busi-
ness Economics and Statistics. http://tandfonline.com/doi/abs/10.1080/07350015
.2015.1137760
Abadie, A., Athey, S., Imbens, G. W., and Wooldridge, J. M. 2014 (July). Finite Popula-
tion Causal Standard Errors. Working Paper 20325. National Bureau of Economic
Research.
Akaike, H. 1974. A New Look at the Statistical Model Identification. IEEE Transactions
on Automatic Control, AC-19(6), 716–723.
Athey, S., and Imbens, G. W. 2015. Machine Learning Methods for Estimating Hetero-
geneous Causal Effects. ArXiv e-prints, Apr.
Banbura, M., Giannone, D., and Reichlin, L. 2010. Large Bayesian Vector Auto Regres-
sions. Journal of Applied Econometrics, 25(1), 71–92.
Banbura, M., Giannone, D., Modugno, M., and Reichlin, L. 2013. Now-Casting and
the Real-Time Data-Flow. In: Elliott, G., and Timmermann, A. (eds), Handbook
of Economic Forecasting, Volume 2, Part A. Elsevier-North Holland, pp. 195–
237.
Belloni, A., Chen, D., Chernozhukov, V., and Hansen, C. 2012. Sparse Models and
Methods for Instrumental Regression, with an Application to Eminent Domain.
Econometrica, 80(6), 2369–2429.
Belloni, A., Chernozhukov, V., and Hansen, C. 2013. Inference for High-Dimensional
Sparse Econometric Modelling. In: Acemoglu, D., Arellano, M., and Dekel, E.
(eds), Advances in Economics and Econometrics, Tenth World Congress of the
Econometric Society, vol. 3. Cambridge University Press, pp. 245–295.
Belloni, A., Chernozhukov, V., and Hansen, C. 2014a. High-Dimensional Methods and
Inference on Structural and Treatment Effects. Journal of Economic Perspectives,
28(2), 29–50.
Belloni, A., Chernozhukov, V., and Hansen, C. 2014b. Inference on Treatment Effects
after Selection among High-Dimensional Controls. The Review of Economic Stud-
ies, 81(2), 608–650.

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 Big Data in Economics: Evolution or Revolution? 629

Billio, M., Casarin, R., Ravazzolo, F., and van Dijk, H. K. 2013a. Interactions between
Eurozone and US Booms and Busts: A Bayesian Panel Markov-switching VAR
Model. Working Papers 2013:17. Department of Economics, University of Venice
‘Ca’ Foscari’.
Billio, M., Casarin, R., Ravazzolo, F., and van Dijk, H. K. 2013b. Time-varying combi-
nations of predictive densities using nonlinear filtering. Journal of Econometrics,
177(2), 213–232.
Casarin, R., Grassi, S., Ravazzolo, F., and van Dijk, H. K. 2013. Parallel Sequential
Monte Carlo for Efficient Density Combination: The DeCo Matlab Toolbox. Work-
ing Papers 2013:08. Department of Economics, University of Venice ‘Ca’ Foscari’.
Casarin, R., Grassi, S., Ravazzolo, F., and van Dijk, H. K. 2015 (July). Dynamic Pre-
dictive Density Combinations for Large Data Sets in Economics and Finance. Tin-
bergen Institute Discussion Papers 15-084/III. Tinbergen Institute.
Chamberlain, G., and Rothschild, M. 1983. Arbitrage, Factor Structure, and Mean-
Variance Analysis on Large Asset Markets. Econometrica, 51(5), 1281–304.
Choi, H., and Varian, H. 2012. Predicting the Present with Google Trends. The Economic
Record, 88(s1), 2–9.
De Mol, C., Giannone, D., and Reichlin, L. 2008. Forecasting Using a Large Number
of Predictors: Is Bayesian Shrinkage a Valid Alternative to Principal Components?
Journal of Econometrics, 146(2), 318–328.
De Mol, C., Giannone, D., and Reichlin, L. 2015. Forecasting with High-dimensional
Time Series. Oberwolfach Reports, No. 38/2015. European Mathematical Society.
Diebold, F. X. 2003. Big Data Dynamic Factor Models for Macroeconomic Measure-
ment and Forecasting: A Discussion of the Papers by Reichlin and Watson. In:
Dewatripont, M., Hansen, L., and Turnovsky, S. (eds), Advances in Economics and
Econometrics: Theory and Applications, Eighth World Congress of the Economet-
ric Society, vol. 3. Cambridge University Press, pp. 115–122.
Diebold, F. X. 2012 (Sep). On the Origin(s) and Development of the Term ‘Big Data’.
PIER Working Paper Archive 12-037. Penn Institute for Economic Research,
Department of Economics, University of Pennsylvania.
Doz, C., Giannone, D., and Reichlin, L. 2012. A Quasi-Maximum Likelihood Approach
for Large, Approximate Dynamic Factor Models. The Review of Economics and
Statistics, 94(4), 1014–1024.
Durham, G., and Geweke, J. 2013 (Apr.). Adaptive Sequential Posterior Simulators for
Massively Parallel Computing Environments. Working Paper Series 9. Economics
Discipline Group, UTS Business School, University of Technology, Sydney.
Einav, L., and Levin, J. 2014. Economics in the Age of Big Data. Science, 346(6210).
Farrell, M. H. 2015. Robust Inference on Average Treatment Effects with Possibly More
Covariates than Observations. Journal of Econometrics, 189(1), 1 – 23.
Forni, M., Hallin, M., Lippi, M., and Reichlin, L. 2000. The Generalized Dynamic Fac-
tor Model: Identification and Estimation. Review of Economics and Statistics, 82,
540–554.
Forni, M., Giannone, D., Lippi, M., and Reichlin, L. 2009. Opening the Black Box:
Structural Factor Models with Large Cross Sections. Econometric Theory, 25(10),
1319–1347.
Frank, I. E., and Friedman, J. H. 1993. A Statistical View of Some Chemometrics
Regression Tools. Technometrics, 35(2), pp. 109–135.

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 630 De Mol, Gautier, Giannone, Mullainathan, Reichlin et al.

Gautier, E., and Tsybakov, A. 2011. High-dimensional Instrumental Variables Regres-

sion and Confidence Sets. ArXiv e-prints, May.
Giannone, D., Reichlin, L., and Small, D. 2008. Nowcasting: The Real-time Informa-
tional Content of Macroeconomic Data. Journal of Monetary Economics, 55(4),
665–676.
Giannone, D., Lenza, M., and Primiceri, G. E. 2015. Prior Selection for Vector Autore-
gressions. Review of Economics and Statistics, 97(2), 436–451.
Hastie, T., Tibshirani, R., and Friedman, J. 2009. The Elements of Statistical Learning.
2 edn. Springer-Verlag.
Hoerl, A. E., and Kennard, R. W. 1970. Ridge Regression: Biased Estimation for
Nonorthogonal Problems. Technometrics, 12(1), pp. 55–67.
Hoogerheide, L. F., van Dijk, H. K., and van Oest, R. D. 2009. Simulation-Based
Bayesian Econometric Inference: Principles and Some Recent Computational
Advances. In: Handbook of Computational Econometrics. John Wiley and Sons,
Ltd, pp. 215–280.
Jungbacker, B., and Koopman, S. J. 2015. Likelihood-based Dynamic Factor Analysis
for Measurement and Forecasting. Econometrics Journal, 18, C1–C21.
Kleinberg, J., Lakkaraju, H., Leskovec, J., Ludwig, J., and Mullainathan, S. 2015a.
Human Decisions and Machine Predictions. Harvard mimeo.
Kleinberg, J., Ludwig, J., Mullainathan, S., and Obermeyer, Z. 2015b. Prediction Policy
Problems. American Economic Review, 105(5), 491–495.
Lee, A., Yau, C., Giles, M. B., Doucet, A., and Holmes, C. C. 2010. On the Utility
of Graphics Cards to Perform Massively Parallel Simulation of Advanced Monte
Carlo Methods. Journal of Computational and Graphical Statistics, 19(4), 769–
789.
Li, Xinyuan. 2016. Nowcasting with Big Data: Is Google Useful in the Presence of Other
Information? London Business School Mimeo.
Lucas, R. J. 1976. Econometric Policy Evaluation: A Critique. Carnegie-Rochester Con-
ference Series on Public Policy, 1(1), 19–46.
Pang, B., Lee, L., and Vaithyanathan, S. 2002. Thumbs up? Sentiment Classifica-
tion Using Machine Learning Techniques. In: Proceedings of EMNLP, pp. 79–
86.
Reichlin, L. 2003. Factor Models in Large Cross Sections of Time Series. In: Dewa-
tripont, M., Hansen, L., and Turnovsky, S. (eds), Advances in Economics and
Econometrics: Theory and Applications, Eighth World Congress of the Economet-
ric Society, vol. 3. Cambridge University Press, pp. 47–86.
Sala-I-Martin, X. X. 1997. I Just Ran Two Million Regressions. The American Economic
Review, 87(2), pp. 178–183.
Schwarz, G. 1978. Estimating the Dimension of a Model. The Annals of Statistics, 6(2),
461–464.
Scott, S. L., and Varian, H. R. 2014. Predicting the Present with Bayesian Structural
Time Series. International Journal of Mathematical Modelling and Numerical
Optimisation, 5(1/2), 4–23.
Stein, C. 1956. Inadmissibility of the Usual Estimator for the Mean of a Multivariate
Normal Distribution. In: Proceedings of the Third Berkeley Symposium on Mathe-
matical Statistics and Probability, Volume 1: Contributions to the Theory of Statis-
tics. Berkeley, Calif.: University of California Press, pp. 197–206.

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
 Big Data in Economics: Evolution or Revolution? 631

Stock, J. H., and Watson, M. W. 1989. New Indexes of Coincident and Leading Eco-
nomic Indicators. In: NBER Macroeconomics Annual 1989, Volume 4. NBER
Chapters. National Bureau of Economic Research, Inc, pp. 351–409.
Stock, J. H., and Watson, M. W. 2002. Forecasting Using Principal Components from a
Large Number of Predictors. Journal of the American Statistical Association, 97,
147–162.
Tibshirani, R. 1996. Regression Shrinkage and Selection via the Lasso. Journal of the
Royal Statistical Society. Series B (Methodological), 58, 267–288.
Varian, H. R. 2014. Big Data: New Tricks for Econometrics. Journal of Economic Per-
spectives, 28(2), 3–28.
Watson, M. W. 2003. Macroeconomic Forecasting Using Many Predictors. In: Dewa-
tripont, M., Hansen, L., and Turnovsky, S. (eds), Advances in Economics and
Econometrics: Theory and Applications, Eighth World Congress of the Economet-
ric Society, vol. 3. Cambridge University Press, pp. 87–115.

Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016
Downloaded from https://www.cambridge.org/core. London Business School Library, on 19 Dec 2017 at 14:36:48, subject to the Cambridge
Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/9781316636404.016