Mixed Frequency Data Sampling Regression Models: The R Package midasr

When modeling economic relationships it is increasingly common to encounter data sampled at diﬀerent frequencies. We introduce the R package midasr which enables esti-mating regression models with variables sampled at diﬀerent frequencies within a MIDAS regression framework put forward in work by Ghysels, Santa-Clara, and Valkanov (2002). In this article we deﬁne a general autoregressive MIDAS regression model with multiple variables of diﬀerent frequencies and show how it can be speciﬁed using the familiar R formula interface and estimated using various optimization methods chosen by the re-searcher. We discuss how to check the validity of the estimated model both in terms of numerical convergence and statistical adequacy of a chosen regression speciﬁcation, how to perform model selection based on a information criterion, how to assess forecasting accuracy of the MIDAS regression model and how to obtain a forecast aggregation of diﬀerent MIDAS regression models. We illustrate the capabilities of the package with a simulated MIDAS regression model and give two empirical examples of application of MIDAS regression.


Introduction
Regression models involving data sampled at different frequencies are of general interest. In this document we introduce the R (R Core Team 2016) package midasr (Ghysels, Kvedaras, and Zemlys 2016) for regression modeling with mixed frequency data based on a framework put forward in recent work by Ghysels et al. (2002), Ghysels, Santa-Clara, and Valkanov (2006a) and Andreou, Ghysels, and Kourtellos (2010) using so called MIDAS, meaning Mi(xed) Da(ta) S(ampling), regressions. The package is available from the Comprehensive R Archive Network (CRAN) at http://CRAN.R-project.org/package=midasr.
In a general framework of regressions with functional constraints on parameters, the midasr package not only provides similar functionality within a standard R framework of the model specification comparable to that available in the usual functions lm or nls, but also deals with an extended model specification analysis for MIDAS regressions.
Several recent surveys on the topic of MIDAS are worth mentioning at the outset. They are: Andreou, Ghysels, and Kourtellos (2011) who review more extensively some of the material summarized in this document, Armesto, Engemann, and Owyang (2010) who provide a very simple introduction to MIDAS regressions and finally Ghysels and Valkanov (2012) who discuss volatility models and mixed data sampling.
MIDAS regression can also be viewed as a reduced form representation of the linear projection which emerges from a state space model approach -by reduced form we mean that the MIDAS regression does not require the specification of a full state space system of equations. Bai et al. (2013) show that in some cases the MIDAS regression is an exact representation of the Kalman filter, in other cases it involves approximation errors which are typically small. The Kalman filter, while clearly optimal as far as linear projection goes, has several disadvantages: (1) it is more prone to specification errors as a full system of measurement and state equations is required and as a consequence (2) requires a lot more parameters, which in turn results in (3) computational complexities which often limit the scope of applications. In contrast, MIDAS regressions -combined with forecast combination schemes if large data sets are involved (see Andreou, Ghysels, and Kourtellos 2013) -are computationally easy to implement and are less prone to specification errors.
The key feature of the package is its flexibility in terms of the model formulation and estimation, which allows for 1 : • estimation of regression models with their parameters defined (restricted) by certain functional constraints using the familiar R formula interface allowing any choice of a constraint, which can be selected from a standard list or can be customized using user-defined R functions; • parsimonious aggregation-linked restrictions (as, e.g., in Ghysels 2013) as a special case; • estimation of MIDAS models with many variables and (numerous) different frequencies; • constrained, partially constrained, or unconstrained estimation of the model; • various mixtures of restrictions/weighting schemes and also lag orders as they can be specific to each series; • statistical testing for the adequacy of the model specification and the imposed functional constraint; 1 Ghysels (2013) also developed a package for MATLAB (The MathWorks Inc. 2014) which deals with the estimation and information criteria-based specification of MIDAS regressions as well as forecasting and nowcasting of low frequency series. All of the midasr features replicate or extend features provided by this package. The key extensions are: the specification of any user-defined functional constraint, the inclusion of multiple variables of different frequency and different functional constraints, the testing of the adequacy of a chosen model specification, and the option to calculate standard errors of the parameters robust to serial correlation and heteroscedasticity in regression disturbances.
• information criteria and testing-based selection of models; • forecasting and nowcasting functionality, including various forecast combinations.
Suppose {y t , t ∈ Z} is a univariate process observed at low frequency. Lags of the process are denoted by By t = y t−1 , where B is the low frequency lag operator. A MIDAS regression involves linear projections using stochastic processes {x (i) τ , τ ∈ Z}, i = 0, . . . , k, observed at a higher frequency, i.e., for each low frequency period t = t 0 we observe the process x (i) τ at m i ∈ N high frequency periods τ = (t 0 − 1)m i + 1, . . . , t 0 m i . Throughout the article we represent ith high frequency period τ in terms of low frequency period t as τ = (t − 1)m i + j, j = 1, . . . , m i . Note that this notation does not exclude the case m i = 1. In that case the high frequency process x (i) τ is observed at the same frequency as the low frequency process y t . However we require that m i ≥ 1, such that the process y t is observed at the lowest frequency. Lags of the processes x (i) τ are denoted by Lx where L is the high frequency lag operator, which operates on the lag irrespective of the frequency of the process.
The package deals with any specification of mixed frequency regression model which can be represented as where we require E(ε t |y t−1 , . . . , y t−p , x The model stated in the Equation 1 can be estimated in the usual time series regression fashion or using a Bayesian approach. However, the number of parameters in this model d = p + k i=0 l i can be very large in terms of the number n of available observations of y t 2 . Since the estimation of the model can easily become infeasible, whenever either larger differences in frequencies or more variables and/or higher lag orders prevail, Ghysels et al. (2002) introduced a sufficiently flexible parametric restriction to be imposed on the original parameters, This approach greatly reduces the number of parameters to be estimated, from d to q = h i i=0 q i , which is assumed to be always considerably less than the number of observations available at the lowest frequency. This gain is offset by the fact that (1) becomes a non-linear model; however, if the parameters of an underlying data generating process followed a certain functional constraint which is perfectly or well approximated by a constraint function chosen by a researcher, then significant efficiency gains could be achieved from the imposed constraints. Figure 1 plots the out-of-sample prediction precision (on the left) and the parameter estimation precision (on the right) in an unconstrained and constrained simple model with correct and approximate restrictions (see Appendix A for details). As can be seen, even an incorrect constraint might be useful whenever the number of degrees of freedom in an unconstrained model is low and, consequently, one cannot rely on the large sample properties of unconstrained estimators. Furthermore, this approach seems to be necessary whenever estimation is simply infeasible because of the lack of degrees of freedom.

Theory
The model (1) can be rewritten in a more compact form: where α(z) = 1 − p j=1 α j z j and In order to simplify notation, without loss of generality, a single order of the lag polynomials is used with l being the maximum lag order. If the orders of some components of β(z) are smaller, it is easy to set some coefficients of the polynomial equal to zero.
We require the existence of the continuous second derivative of the functional constraint with respect to its parameters, i.e., ∂ 2 f i ∂γ i ∂γ i . The functional constraints can vary with each variable and/or frequency, and therefore we use γ to represent a vector of all the parameters of a restricted model with q = dim(γ) their total number.
As will be shown in the next section, all variants of the usual linear (in terms of variables) MIDAS regression model are covered by regression (3) via the specification of functional constraints. When each restriction function is an identity mapping, one obtains an unrestricted MIDAS regression model. 3

Frequency alignment
It is instructive to rewrite the model (1) in matrix notation. We start with a few examples. Suppose y t is observed quarterly and we want to explain its variation with the variable x τ , which is observed monthly. Since each quarter has three months, the frequency m is 3 in this example. Suppose we assume that the monthly data in the current and the previous quarter has explanatory power. This means that for each quarter t we want to model y t as a linear combination of variables x 3t , x 3t−1 , x 3t−2 observed in the quarter t and variables y t−1 and x 3(t−1) , x 3(t−1)−1 , x 3(t−1)−2 observed in the previous quarter t − 1. In matrix notation the MIDAS model (1) for this example is: By writing the model in matrix notation we transform the high-frequency variable x τ into a low-frequency vector (x 3t , . . . , x 3t−5 ) . We call this transformation the frequency alignment. Note that we require that the number of observations of x τ is exactly 3n.
Let us examine another example. Suppose we have another variable z t observed weekly which we want to add to the model. The model (1) does not allow varying frequency ratios, so we need to assume that each month has exactly 4 weeks. If months do not always have four weeks, as they do in practice, one can simply think of this model as taking a fixed set of weekly lags. The frequency m for the variable z τ is then 12. We use again the current and previous quarter data for explaining variation in y t . This means that for quarter t we model y t as a linear combination of variables x 3t , x 3t−1 , x 3t−2 and z 12t , z 12t−1 , . . . , z 12t−11 observed in the quarter t, and variables y t−1 , x 3(t−1) , . . . , x 3(t−1)−2 and z 12(t−1) , z 12(t−1)−1 , . . . , z 12(t−1)−11 observed in the quarter t − 1. The model in matrix form is then: In this example we aligned x τ into a vector (x 3t , . . . , x 3t−5 ) and z τ into a vector (z 12t , . . . , z 12t−23 ) . Again we require that the number of observations of high frequency variables are multiples of n, with multiplication factor being the corresponding frequencies. This is not a restrictive assumption in practical applications as will be further explained in Section 3.
Let us return to the general case of the model (1). We align the frequency of high-frequency variable x τ by transforming it to the low-frequency vector (x (i) (1) is then expressed in matrix notation as follows: and u is the smallest integer such that um i − l > 0 and u > p.
The purpose of this subsection was to show how the frequency alignment procedure turns a MIDAS regression into a classical time series regression where all the variables are observed at the same frequency.

Estimation
Equation 3 can be estimated directly via ordinary least squares (OLS), without restrictions on the parameters. This is a so called U-MIDAS regression model, see Foroni et al. (2015). Furthermore, a consistent non-parametric approach could be used to estimate the underlying parameters of a function as, e.g., in Breitung, Roling, and Elengikal (2013). Since, none of these approaches use a parametric functional constraint, they can be estimated using already available R packages. The midasr package aims at the estimation of mixed frequency models with some parametric functional constraints.
While model (3) is a linear model in terms of variables, any non-linear functional constraints will result in non-linearities with respect to the parameters γ. Therefore, in the general case, we use in the function midas_r the non-linear least squares (NLS) estimator of parameters γ of a restricted model (3) as defined by where the lag polynomial of constrained parameters is defined by for each (i, j) ∈ {0, 1, . . . , k} × {0, 1, . . . , l}. A number of numerical algorithms are readily available in R. By default, the optim optimization function is used with optional choices of optimization algorithms in it. However, a user can also choose within the function midas_r other procedures available in R such as nls, thus customizing the desired algorithm which is suitable for the problem at hand.
The efficiency of the estimator and consistency of the standard errors depend on whether the errors of the model are spherical. We leave the aspect of efficiency of estimation to be considered by a user, however the implementation of heteroscedasticity and autocorrelation (HAC) robust standard errors is an option available in the package sandwich (see Zeileis 2004, Zeileis 2006).
If all the functional relations f i (·) were non-constraining identity mappings, then the NLS estimator would be equivalent to the ordinary least squares (OLS) problem in terms of the original parameters. For convenience, such a U-MIDAS version can be dealt with directly using a different function midas_u of the package (see an illustration in the Section 3) or a standard lm function, provided the alignment of data frequencies is performed as discussed in the previous section.

Taxonomy of aggregates-based MIDAS regression models
Based on the parsimony of representation argument, the higher-frequency part of conditional expectation of MIDAS regressions is often formulated in terms of aggregates as follows with some low-frequency number of lags q ∈ N and parameter-driven low-frequency aggregates which depend on a weighting (aggregating within a low-frequency period) function w r (δ i,r ; s) with its parameter vector δ i,r , which, in the general case, can vary with each variable/frequency and/or the low-frequency lag order r ∈ N. Here the aggregation weights are usually nonnegative and, for identification of parameters {λ , satisfy the normalization constraint such as m i −1 s=0 w r (δ i,r ; s) = 1. To have the weights add to one, it is convenient to define a weighting function in the following form given some underlying function ψ (i) r (·). Provided that the latter function is non-negativelyvalued (and the denominator is positive), the resulting weights in Equation 7 are also nonnegative. Table 1 provides a list of some underlying functions producing, within the context of Equation 7, the usual weighting schemes with non-negative weights (whenever the parameter Resulting (normalized) ψ(δ; s) := ψ (i) r (δ i,r ; s) Related midasr weighting scheme function Exponential Almon lag polynomial nealmon Beta (analogue of probability density function) , with some marginally small quantity ξ > 0, and δ = (δ 1 , δ 2 ) ∈ R 2 + .
nbeta Gompertz (analogue of probability density function) gompertzp Log-Cauchy (analogue of probability density function) lcauchyp Nakagami (analogue of probability density function) Table 1: A sample of weighting schemes in aggregation-based MIDAS specifications.
space of underlying functions is appropriately bounded, which in some cases is also needed for identification of parameters). In order to avoid heavy notation, indices i and r -which are connected respectively with frequency/variable and the lag order -are dropped in the table.
Some other weighting functions which do not have a representation as in Equation 7 are also available in the package such as (non-normalized) almonp and the polynomial specification with step functions polystep (see Ghysels et al. 2006b for further discussion of step functions).
However, the choice of a particular weighting function in the MIDAS regression with aggregates represents only one restriction imposed on β(L) out of many other choices to be made. To see this, let us note that aggregates-based MIDAS regressions can be connected with the following restrictions on the conditional expectation of model (3): where y t,1 = (y t−1 , . . . , y t−p ) .
As can be seen -and leaving aside other less intuitive restrictions -depending on the choice of a particular MIDAS specification with aggregates, one can impose restrictions on the equality of • the applied weighting scheme/function across variables and/or frequencies (∀ i, w • the applied weighting scheme/function across all low-frequency lags r = 0, 1, . . . , q of aggregates (∀ r, w r (·) = w(·)); • parameters of the weighting functions in each lag (∀ r, δ i,r = δ i ); • impact of contemporaneous and lagged aggregates for all lags (∀ r, λ Furthermore, let s i stand for an enumerator of ith higher-frequency periods within a lowfrequency period. Then, noting that, given a frequency ratio m i , there is a one-to-one mapping between higher-frequency index j ∈ N and a pair (r, Hence, it is easy to see that the aggregates-based MIDAS induces a certain periodicity of the functional constraint f i in Equation 3 as illustrated bellow using a stylized case where all the restrictions are imposed in Equation 8: for any i ∈ {0, 1, . . . , h}. From Equation 9 it is clear that any specification of MIDAS regression models which relies on aggregates is a special case of representation (3) with just a specific functional constraint on parameters. On the other hand, not every general constraint β(L) can be represented using periodic aggregates. For instance, in the above characterized example the correspondence necessarily breaches whenever there exists at least one frequency i, for which none of q ∈ N satisfies l = qm i − 1.

Specification selection and adequacy testing
Besides the usual considerations about the properties of the error term, there are two main questions about the specification of the MIDAS regression models. First, suitable functional constraints need to be selected, since their choice will affect the precision of the model. Second, the appropriate maximum lag orders need to be chosen.
One way to address both issues together is to use some information criterion to select the best model in terms of the parameter restriction and the lag orders using either in-or out-ofsample precision measures. Functions midas_r_ic_table and amidas_table of the package allow the user to make an in-sample choice using some usual information criteria, such as AIC and BIC, and a user-specified list of functional constraints. 4 Another way is to test the adequacy of the chosen functional constraints. For instance, whenever the autoregressive terms in model (3) are present (p > 0), it was pointed out by Ghysels et al. (2006b) that, in the general case, φ(L) = β(L)/α(B) will have seasonal patterns thus corresponding to some seasonal impact of explanatory variables on the dependent one in a pure distributed lag model (i.e., without autoregressive terms). To avoid such an effect whenever it is not (or is believed to be not) relevant, Clements and Galvão (2008) proposed to us a common factor restriction which can be formulated as a common polynomial restriction with a constraint on the polynomial β(L) to satisfy a factorization β(L) = α(B)φ(L), so that inverting Equation 3 where A is a suitable normalization matrix (see Kvedaras and Zemlys 2012 for a standard and Kvedaras and Zemlys 2013 for a HAC-robust version of the test), and q = dim(γ) and d = dim(θ) denotes the number of parameters in a restricted and unrestricted model, respectively. Functions hAh_test and hAhr_test of the package implement the described testing as will be illustrated later.

Forecasting
Let us write model (3) for period t + 1 as where y t,0 = (y t , . . . , y t−p+1 ) and α = (α 1 , α 2 , . . . , α p ) is a vector of parameters of the autoregressive terms. This representation is well suited for (one step ahead) conditional forecasting of y t+1 , provided that the information on the explanatory variables is available.
If it were absent, forecasts of x t+1,0 would be also necessary from a joint process of {y t , x t,0 } which might be difficult to specify and estimate correctly, especially, bearing in mind the presence of data with mixed frequencies. Instead, a direct approach to forecasting is often applied in the MIDAS framework. Namely, given an information set available up to a moment t defined by an h-step ahead direct forecast can be formed leaning on a model linked to a corresponding conditional expectation where α h and β h (L) are the respective horizon h-specific parameters. Note that, in principle, these conditional expectations have a form of representation (3) with certain restrictions on the original lag polynomials of coefficients. Hence, in the general case, the suitable restrictions for each h will have a different form.
Given periods h = 1, 2, . . . , and a selected model or a list of specifications to be considered, package midasr provides the point forecasts corresponding to the estimated analogue of Equation 11 evaluates the precision of different specifications, and performs weighted forecasting using the framework defined in Ghysels (2013).

Alternative representations of MIDAS regressions
The model (1) represents a very general MIDAS regression representation. We give below a sample of other popular MIDAS regression specifications from Andreou et al. (2011). These specifications assume that only one high frequency variable is available. We denote its frequency by m.
1. DL-MIDAS(p X ): where F t is a factor derived from additional data.
4. ADL-MIDAS-M(p X ,p Y ), or multiplicative MIDAS: 5. MIDAS with leads, where it is assumed that we have only J < m observations of high frequency variable x τ available for period t + 1 The parametric restriction is usually placed on coefficients β j in these representations in the usual manner of (2), i.e., it is assumed that β j = f (γ, j), for chosen function f with parameters γ.
If we compare these representations with (1) the notable differences are the specification of y t+1 instead of y t as a response variable and a preference to make the maximum high frequency lag be a multiple of the corresponding frequency. The multiplicative MIDAS is the special case of aggregates-based MIDAS specification. It is evident then that all these representations are special cases of (1).
Various examples illustrating MIDAS regressions specifications together with corresponding R code are also given in the Table 3.

Data handling
From a data handling point of view, the key specificity of the MIDAS regression model is that the length of observations of variables observed at various frequencies differs and needs to be aligned as described in Section 2. There is no existing R function which performs such a transformation and the package midasr provides a solution to these challenges. The basic functionality of data handling is summarized in Table 2.
Function fmls(x, k, m) performs exactly the transformation defined in Equation 4, converting an observation vector x of a given (potentially) higher-frequency series into the corresponding stacked matrix of observations of (k + 1) low-frequency series (contemporaneous with k lags) as defined by the maximum lag order k and the frequency ratio m. For instance, given a series of twelve observations R> x <-1:12 we get the following result Stacks a HF data vector x into a corresponding matrix of observations at LF of size dim x m × dim k: from the first to the last HF lag defined by vector k.
mls(x, 2:3, 3) dim x m must be an integer (NA are allowed). For m = 1, the function produces lags of x that are defined by vector argument k, e.g., mls(x, 2:3, 1) yields a data set containing the lags x t−2 and x t−3 of x t . fmls(x, k, m) Same as mls, except that k is a scalar and the k + 1 lags are produced starting from 0 up to k.
dmls(x, k, m) Same as fmls, only the resulting matrix contains k + 1 first-order HF differences of x.
Function mls is slightly more flexible as the lags included can start from a given order rather than from zero, whereas the function fmls uses a full lag structure. dmls performs in addition a first-order differencing of the data which is convenient when working with integrated series.
A couple of issues should be taken into account when working with series of different frequencies: • It is assumed that the numbers of observations of different frequencies match exactly through the frequency ratio (n i = nm i ), and the first and last observations of each series of different frequency are correspondingly aligned (possibly using NA to account for some missing observations for series of higher frequency).
• Because of different lengths of series of various frequencies, the data for the model cannot be kept in one data.frame. It is expected that variables for the model are either vectors residing in the R global environment, or are passed as elements of a list.
Variables of different frequency can be in the same data.frame, which in turn should be an element of a list.

An example of simulated MIDAS regression
Using the above data handling functions, it is straightforward to simulate a response series from the MIDAS regression as a data generating process (DGP). For instance, suppose one is willing to generate a low-frequency response variable y in the MIDAS with two higherfrequency series x and z where the impact parameters satisfy the exponential Almon lag polynomials of different orders as follows: where (x τ 1 , z τ 2 , ε t ) are independent for any (τ 1 , τ 2 , t) ∈ Z 3 , and where d 1 = k 1 + 1 = 8 is a multiple of the frequency ratio m 1 = 4, whereas d 2 = k 2 + 1 = 17 is not a multiple of m 2 = 12. Here q 1 = 2, q 2 = 3 with parameterizations which yield the shapes of functional constraints as plotted in Figure 2.

Some specification examples of MIDAS regressions
Suppose now that we have (only) observations of y, x, and z which are stored as vectors, matrices, time series, or list objects in R, and our intention is to estimate a MIDAS regression model as in Equation 12: (a) without restricting the parameters (as in U-MIDAS) and using OLS; (b) with the exponential Almon lag polynomial constraint on parameters (as in the function nealmon) and using NLS.
The following R code estimates the constrained case (b) using the function midas_r and reports the NLS estimates γ of parameters with the related summary statistics.
The model is specified via the familiar formula interface. The lags included and functional restriction used can be individual to each variable and are specified within the respective mls, fmls, or dmls function used with midas_r. It is necessary to provide a list of starting values for each variable with restricted coefficients, since it implicitly defines the number of parameters of the constraint functions to be used for each series.
The main difference to the function nls is that there is a greater choice of numerical optimization algorithms. The function midas_r is written in a way that in theory it can use any R optimization function. The choice is controlled via the Ofunction argument. Currently it is possible to use functions optim and nls which are present in a standard R installation and function optimx from the package optimx (Nash and Varadhan 2011; Nash 2014). The additional arguments to the aforementioned functions can be specified directly in the call to midas_r. So for example if we want to use the optimization algorithm of Nelder and Mead, which is the default option in the function optim we use the following code R> eq_r2 <-midas_r(y~trend + mls(x, 0:7, 4, nealmon) + + mls(z, 0:16, 12, nealmon), start = list(x = c(1, -0.5), + z = c(2, 0.5, -0.1)), Ofunction = "optim", method = "Nelder-Mead") If we want to use the Golub-Pereyra algorithm for partially linear least-squares models implemented in the function nls we use the following code R> eq_r2 <-midas_r(y~trend + mls(x, 0:7, 4, nealmon) + + mls(z, 0:16, 12, nealmon), start = list(x = c(1, -0.5), + z = c(2, 0.5, -0.1)), Ofunction = "nls", method = "plinear") It is possible to re-estimate the NLS problem with the different algorithm using as starting values the final solution of the previous algorithm. For example it is known, that the default algorithm in nls is sensitive to starting values. So first we can use the standard Nelder-Mead algorithm to find "more feasible" starting values and then use nls to get the final result: R> eq_r2 <-midas_r(y~trend + mls(x, 0:7, 4, nealmon) + + mls(z, 0:16, 12, nealmon), start = list(x = c(1, -0.5), + z = c(2, 0.5, -0.1)), Ofunction = "optim", method = "Nelder-Mead") R> eq_r2 <-update(eq_r2, Ofunction = "nls") The output of the optimization function used can be found by inspecting the element opt of the midas_r output.

$message NULL
Here we observe that the Nelder-Mead algorithm evaluated the cost function 502 times.
The optimization functions in R report the status of the convergence of the optimization algorithm by a numeric constant with 0 indicating successful convergence. This numeric constant is reported as the element convergence of the midas_r output.

R> eq_r2$convergence
[1] 1 In this case the convergence was not successful. The help page of the function optim indicates that convergence code 1 means that the iteration limit was reached.
In order to improve the convergence it is possible to use user defined gradient functions. To use them it is necessary to define the gradient function of the restriction. For example for the nealmon restriction the gradient function is defined in the following way: R> nealmon_gradient <-function(p, d, m) { + i <-1:d + pl <-poly(i, degree = length(p) -1, raw = TRUE) + eplc <-exp(pl %*% p[-1])[, , drop = TRUE] + ds <-colSums(pl * eplc) + s <-sum(eplc) + cbind(eplc / s, p[1] * (pl * eplc / s -eplc %*% t(ds) / s^2)) + } The gradient functions are passed as named list elements via argument weight_gradients: R> eq_r2 <-midas_r(y~trend + mls(x, 0:7, 4, nealmon) + + mls(z, 0:16, 12, nealmon), start = list(x = c(1, -0.5), z = c(2, 0.5, + -0.1)), weight_gradients = list(nealmon = nealmon_gradient)) This way midas_r calculates the exact gradient of the NLS problem (5) using the specified gradient function of the restriction. For all the types of the restrictions referenced in Table 3 their gradient functions are specified in the package midasr. The naming convention for gradient functions is restriction_name_gradient. It is not necessary to explicitly pass gradient functions named according to this convention. If weight_gradients is not NULL and does not contain the appropriately named element, it is assumed that there exists a gradient function conforming to the gradient naming convention which is then subsequently used.
The gradient and the Hessian of the NLS problem are supplied as the output of midas_r.
The gradient is calculated exactly if appropriate gradients for weight functions are supplied as explained above, otherwise the numerical approximation of the gradient is calculated using the package numDeriv (Gilbert and Varadhan 2015). For Hessian the numerical approximation is always used. Having the gradient and Hessian calculated allows to check whether the necessary and sufficient conditions for convergence are satisfied. This is performed by the function deriv_test which calculates the Euclidean norm of the gradient and the eigenvalues of the Hessian. It then tests whether the norm of the gradient is close to zero and whether the eigenvalues are positive.
R> deriv_tests(eq_r, tol = 1e-06) In the example provided above, a functional constraint was imposed directly on β(L) terms corresponding to each series without the usage of aggregates. Relying on the relationship (9), it is always possible to write such an explicit general constraint from an aggregates-based one. For convenience of use, the function amweights can be used to form several standard periodic functional constraints with "typical" restrictions explicated in Equation 7. For instance, R> amweights(p = c(1, -0.5), d = 8, m = 4, weight = nealmon, type = "C") twice (d/m = 2), as implied by the number of periods at higher-frequency (d = 8) and the frequency ratio (m = 4). In this way, the function amweights can be used to define explicitly a new functional constraint relying on the relationship (9). Alternatively, one can indicate directly within the function midas_r that the aggregates-based restriction must be used as follows R> eq_r2 <-midas_r(y~trend + mls(x, 0:7, 4, amweights, nealmon, "C") + + mls(z, 0:16, 12, nealmon), start = list(x = c(1, -0.5), + z = c(2, 0.5, -0.1))) where the first variable follows an aggregates-based MIDAS restriction scheme. Note that the selection of alternative types "A" and "B" are connected with specifications having a larger number of parameters (see Table 3), hence the list of starting values needs to be adjusted to account for an increase in the number of (potentially unequal) impact parameters.
It should be also noted that, whenever the aggregates-connected restrictions are used, the number of periods must be a multiple of the frequency ratio. For instance, the current lag specification for variable z is not consistent with this requirement and cannot be represented through the (periodic) aggregates, but either mls(z, 0:11, 12, amweights, nealmon, "C") Constraints on β (i) j , i = 1, 2 are given by different func-tions.

Adequacy testing of restrictions
Given a MIDAS regression model estimated with midas_r, the empirical adequacy of the functional restrictions can be tested under quite standard assumptions (see Zemlys 2012 and using functions hAh_test and hAhr_test of the package. In the case of a stationary series {y t } they can be applied directly, whereas whenever {y t } is cointegrated with explanatory variables, a special transformation needs to be applied before the testing (see, e.g., . The hAh_test can be used whenever errors of the process are independently and identically distributed, whereas the hAhr_test uses a HAC-robust version of the test. We should just point out that, whenever no significant HAC in the residuals are observed, we would suggest using hAh_test which would then have more precise test sizes in small samples. In the case of an integrated series {y t } which is co-integrated with explanatory variables, some other alternatives are available (see Bilinskas, Kvedaras, and Zemlys 2013).
For illustration, let us use eq_r to represent an estimated model as in the previous subsections. Then the functions produce, respectively,

R> hAhr_test(eq_r)
hAh restriction test (robust version) data: hAhr = 14.854, df = 20, p-value = 0.7847 Here the value of a test statistic, the degrees of freedom (the number of binding constraints on parameters in Equation 3), and the empirical significance of the null hypothesis that a functional constraint is adequate are reported.

R> hAhr_test(eq_rb)
hAh restriction test (robust version) data: hAhr = 32.879, df = 17, p-value = 0.01168 Whenever the empirical adequacy cannot be rejected at some appropriate level of significance for a couple of models, we could further rely on information criteria to make the selection of the best candidate(s).
The potential lag structures are given by the following ranges of high-frequency lags: from [from; m * min(to)] to [from; m * max(to)]. When aggregates-based modeling is involved using amweights in midas_r, m can be set to the frequency ratio which ensures that the considered models (lag structures) are multiples of it. Otherwise, we would recommend to operate with high-frequency lag structures without changing the default value m = 1. Then, the set of potential models is defined as all possible different combinations of functions and lag structures with a corresponding set of starting values. A simple example bellow illustrates the result in order to reveal the underlying structure, which, besides the understanding of it, is otherwise not needed for a user.
The summary table is a data.frame where each row corresponds to a candidate model; so this table can be manipulated in the usual R way. The table can be accessed as table element of the list returned by midas_r_ic_table. The list of fitted 'midas_r' objects of all candidate models can be accessed as candlist element. It is possible to inspect each candidate model and fine-tune its convergence if necessary.

R> eqs_ic <-update(eqs_ic)
It should be pointed out that there is no need to provide the weighting function nor a specific lag order in the mls functions in a call to midas_r_ic_table, since they are defined by the respective potential sets of models under option table. Any provided values with mls (or other similar functions) are over-written by those defined in table.
Finally, the best model in terms of a selected information criterion in a restricted or unrestricted model then is simply obtained by using R> modsel(eqs_ic, IC = "AIC", type = "restricted") which also prints the usual summary statistics as well as the testing of adequacy of the applied functional restriction using, by default, hAh_test. A word of caution is needed here to remind that, as is typical, the empirical size of a test corresponding to a complex model-selection procedure might not correspond directly to a nominal one of a single-step estimation.

Forecasting
Conditional forecasting (with prediction intervals, etc.) using unrestricted U-MIDAS regression models which are estimated using lm can be performed using standard R functions, e.g., the predict method for 'lm' objects. Conditional point prediction given a specific model is also possible relying on a standard predict function.
The predict method implemented in package midasr works similarly to the predict method for 'lm' objects. It takes the new data, transforms it into an appropriate matrix and multiplies it with the coefficients. Suppose we want to produce the forecastŷ T +1|T for model (12). To produce this forecast we need the data x 4(T +1) , . . . , x 4T −3 and z 12(T +1) , . . . , z 12T −4 . It would be tedious to calculate precisely the required data each time we want to perform a forecasting exercise. To alleviate this problem package midasr provides the function forecast. This function assumes that the model was estimated with the data up to low frequency index T. It is then assumed that the new data is the data after the low frequency T and then calculates the appropriate forecast. For example suppose that we have new data for one low frequency period for model (12). Here is how the forecast for one period would look like: R> newx <-rnorm(4) R> newz <-rnorm (12) R> forecast(eq_rb, newdata = list(x = newx, z = newz, trend = 251)) Point Forecast 1 28.28557 It is also common to estimate models which do not require new data for forecasting where is the desired forecasting horizon. This model can be rewritten as and can be estimated using midas_r. For such a model we can get forecastsŷ T + |T , . . . ,ŷ T +1|T using the explanatory variable data up to low frequency index T. To obtain these forecasts using the function forecast we need to supply NA values for explanatory variables. An example for = 1 is as follows: R> eq_f <-midas_r(y~trend + mls(x, 4 + 0:7, 4, nealmon) + + mls(z, 12 + 0:16, 12, nealmon), start = list(x = c(1, -0.5), + z = c(2, 0.5, -0.1))) R> forecast(eq_f, newdata = list(x = rep(NA, 4), z = rep(NA, 12), + trend = 251)) Point Forecast 1 27.20452 Note that we still need to specify a value for the trend.
In addition, the package midasr provides a general flexible environment for out-of-sample prediction, forecast combination, and evaluation of restricted MIDAS regression models using the function select_and_forecast. If exact models were known for different forecasting horizons, it can also be used just to report various in-and out-of-sample prediction characteristics of the models. In the general case, it also performs an automatic selection of the best models for each forecasting horizon from a set of potential specifications defined by all combinations of functional restrictions and lag orders to be considered, and produces forecast combinations according to a specified forecast weighting scheme.
In general, the definition of potential models in the function select_and_forecast is similar to that one used in the model selection analysis described in the previous section. However, different best performing specifications are most likely related with each low-frequency forecasting horizon = 0, 1, 2, . . . . Therefore the set of potential models (parameter restriction functions and lag orders) to be considered for each horizon needs to be defined.
Suppose that, as in the previous examples, we have variables x and z with frequency ratios m 1 = 4 and m 2 = 12, respectively. Suppose that we intend to consider forecasting of y up to three low-frequency periods ∈ {1, 2, 3} ahead. It should be noted that, in terms of highfrequency periods, they correspond to m 1 ∈ {4, 8, 12} for variable x, and m 2 ∈ {12, 24, 36} for variable z. Thus these variable-specific vectors define the lowest lags 6 of high-frequency period to be considered for each variable in the respective forecasting model (option from in the function select_and_forecast). Suppose further that in all the models we want to consider specifications having not less than 10 high-frequency lags and not more than 15 for each variable. This defines the maximum high-frequency lag of all potential models considered for each low-frequency horizon period ∈ {1, 2, 3}. Hence, for each variable, three corresponding pairs ( m 1 + 10, m 1 + 15), ∈ {1, 2, 3} will define the upper bounds of ranges to be considered (option to in the function select_and_forecast). For instance, for variable x, three pairs (14, 19), (18, 23), and (22, 27) correspond to = 1, 2, and 3 and together with that defined in option from (see x = (4, 8, 12)) imply that the following ranges of potential models will be under consideration for variable x: The other options of the function select_and_forecast are options of functions midas_r_ic_table, modsel and average_forecast.
Then, among others, R> cbfc$accuracy$individual R> cbfc$accuracy$average report, respectively: • the best forecasting equations (in terms of a specified criterion out of the above-defined potential specifications), and their in-and out-of-sample forecasting precision measures for each forecasting horizon; • the out-of-sample precision of forecast combinations for each forecasting horizon.
The above example illustrated a general usage of the function select_and_forecast including selection of best models. Now suppose that a user is only interested in evaluating a one step ahead forecasting performance of a given model. Suppose further that he/she a priori knows that the best specifications to be used for this forecasting horizon = 1 is with • mls(x, 4:12, 4, nealmon) with parameters x = c(2, 10, 1, -0.1) (the first one representing an impact parameter and the last three being the parameters of the normalized weighting function), and • mls(z, 12:20, 12, nealmon) with parameters z = c(-1, 2, -0.1), i.e., with one parameter less in the weighting function.
Given already preselected and evaluated models, a user can use the function average_forecast to evaluate the forecasting performance. To use this function at first it is necessary to fit the model and then pass it to function average_forecast specifying the in-sample and outof-sample data, accuracy measures and weighting scheme in a similar manner to function select_and_forecast R> mod1 <-midas_r(y~trend + mls(x, 4:14, 4, nealmon) + mls(z, + 12:22, 12, nealmon), start = list(x = c(10, 1, -0.1), z = c(2, -0.1))) R> avgf <-average_forecast(list(mod1), data = list(y = y, x = x, z = z, + trend = trend), insample = 1:200, outsample = 201:250, type = "fixed", + measures = c("MSE", "MAPE", "MASE"), + fweights = c("EW", "BICW", "MSFE", "DMSFE")) It should also be pointed out that the forecast combinations in select_and_forecast are obtained only from the forecasts linked to different restriction functions on parameters. The forecasts related to different lag specifications are not combined, but the best lag order is chosen in terms of a given information criterion. If there is a need to get forecast combinations for a group of models which the user selected using other criteria, the function average_forecast should be used in a manner outlined in the previous example.

Forecasting GDP growth
We replicate the example provided in Ghysels (2013). In particular we run MIDAS regression to forecast quarterly GDP growth with monthly non-farms payroll employment growth. The forecasting equation is the following where y t is the log difference of quarterly seasonally adjusted real US GDP and x 3t is the log difference of monthly total employment non-farms payroll. The data is taken from the St. Louis FRED website.
First we load the data and perform necessary transformations.
To specify the model for the midas_r function we rewrite it in the following equivalent form: As in Ghysels (2013) we restrict the estimation sample from the first quarter of 1985 to the first quarter of 2009. We evaluate the models with the Beta polynomial, Beta with non-zero and U-MIDAS weight specifications.

Forecasting realized volatility
As another demonstration we use the package midasr to forecast the daily realized volatility. A simple model for forecasting the daily realized volatility was proposed by Corsi (2009). The heterogeneous autoregressive model of realized volatility (HAR-RV) is defined as where RV t is the daily realized volatility and RV (w) t and RV (m) t are weekly and monthly averages: where we assume that a week has 5 days, and a month has 4 weeks. This model is a special case of a MIDAS regression: For empirical demonstration we use the realized variance data on stock indices provided by Heber, Lunde, Shephard, and Sheppard (2009). We estimate the model for the annualized realized volatility of the S&P500 index, which is based on 5-minute returns data. R> data("rvsp500", package = "midasr") R> spx2_rvol <-100 * sqrt(252 * as.numeric(rvsp500[, "SPX2.rv"])) R> mh <-midas_r(rv~mls(rv, 1:20, 1, harstep), data = list(rv = spx2_rvol), + start = list(rv = c(1, 1, 1))) R> summary ( For comparison we also estimate the model with the normalized exponential Almon weights R> mr <-midas_r(rv~mls(rv, 1:20, 1, nealmon), data = list(rv = spx2_rvol), + start = list(rv = c(0, 0, 0)), weight_gradients = list()) R> summary(mr) We can see that the null hypothesis pertaining to the HAR-RV-implied constraints in the MIDAS regression model is rejected at the 0.05 significance level, whereas the null hypothesis that the exponential Almon lag restriction is adequate cannot be rejected.  For the exponential Almon lag specification we can choose the number of lags via AIC or BIC.
We can look into the forecast performance of both models, using a rolling forecast with 1000 observations window. For comparison we also calculate the forecasts of an unrestricted AR(20) model.

Final remarks
Only a part of the available functionality of the discussed functions of the package midasr was presented. As it is usual in R, much more information on the resulting objects and all the information on the package-specific functions can be retrieved by using the generic functions objects and ?, respectively. Furthermore, in order to save space, the coding examples provided were almost always presented with minimal accompanying output obtained after running the code. The package page contains all the codes and complete output together with some additional illustration of the functionality of the package. Other information with a list of the functions and a number of demonstration codes is accessible using the usual ??midasr.