Testing for structural stability of factor augmented forecasting models

Mild factor loading instability, particularly if sufficiently independent across the different constituent variables, does not affect the estimation of the number of factors, nor subsequent estimation of the factors themselves (see e.g.  Stock and Watson (2009)). This result does not hold in the presence of large common breaks in the factor loadings, however. In this case, information criteria overestimate the number of breaks. Additionally, estimated factors are no longer consistent estimators of “true” factors. Hence, various recent research papers in the diffusion index literature focus on testing the constancy of factor loadings. However, forecast failure of factor augmented models can be due to either factor loading instability, regression coefficient instability, or both. To address this issue, we develop a test for the joint hypothesis of structural stability of both factor loadings and factor augmented forecasting model regression coefficients. Our proposed test statistic has a chi-squared limiting distribution, and we are able to establish the first order validity of (block) bootstrap critical values. Empirical evidence is also presented for 11 US macroeconomic indicators.     

Unstable Diffusion Indexes: With an Application to Bond Risk Premia

This paper studies the empirically relevant problem of estimation and inference in diffusion index forecasting models with structural instability. Factor model and factor augmented regression both experience a structural change with different unknown break dates. In the factor model, we estimate factors and loadings by principal components. We consider least squares estimation of the factor augmented regression and propose a break test. The empirical application uncovers instabilities in the linkages between bond risk premia and macroeconomic factors.     

Inference on factor structures in heterogeneous panels

This paper develops an estimation and testing framework for a stationary large panel model with observable regressors and unobservable common factors. We allow for slope heterogeneity and for correlation between the common factors and the regressors. We propose a two stage estimation procedure for the unobservable common factors and their loadings, based on Common Correlated Effects estimator and the Principal Component estimator. We also develop two tests for the null of no factor structure: one for the null that loadings are cross sectionally homogeneous, and one for the null that common factors are homogeneous over time. Our tests are based on using extremes of the estimated loadings and common factors. The test statistics have an asymptotic Gumbel distribution under the null, and have power versus alternatives where only one loading or common factor differs from the others. Monte Carlo evidence shows that the tests have the correct size and good power.     

Identifying relevant and irrelevant variables in sparse factor models

This paper considers factor estimation from heterogeneous data, where some of the variables—the relevant ones—are informative for estimating the factors, and others—the irrelevant ones—are not. We estimate the factor model within a Bayesian framework, specifying a sparse prior distribution for the factor loadings. Based on identified posterior factor loading estimates, we provide alternative methods to identify relevant and irrelevant variables. Simulations show that both types of variables are identified quite accurately. Empirical estimates for a large multi‐country GDP dataset and a disaggregated inflation dataset for the USA show that a considerable share of variables is irrelevant for factor estimation.     

Testing for structural breaks in dynamic factor models

In this paper we investigate the consequences of structural breaks in the factor loadings for the specification and estimation of factor models based on principal components and suggest procedures for testing for structural breaks. It is shown that structural breaks severely inflate the number of factors identified by the usual information criteria. The hypothesis of a structural break is tested by using LR, LM and Wald statistics. The LM test (which performs best in our Monte Carlo simulations) is generalized to test for structural breaks in factor models where the break date is unknown and the common factors and idiosyncratic components are serially correlated. The proposed test procedures are applied to datasets from the US and the euro area.     

Detecting big structural breaks in large factor models

Time invariance of factor loadings is a standard assumption in the analysis of large factor models. Yet, this assumption may be restrictive unless parameter shifts are mild (i.e., local to zero). In this paper we develop a new testing procedure to detect big breaks in these loadings at either known or unknown dates. It relies upon testing for parameter breaks in a regression of one of the factors estimated by Principal Components analysis on the remaining estimated factors, where the number of factors is chosen according to Bai and Ng’s (2002) information criteria. The test fares well in terms of power relative to other recently proposed tests on this issue, and can be easily implemented to avoid forecasting failures in standard factor-augmented (FAR, FAVAR) models where the number of factors is a priori imposed on the basis of theoretical considerations.     

Bayesian exploratory factor analysis

This paper develops and applies a Bayesian approach to Exploratory Factor Analysis that improves on ad hoc classical approaches. Our framework relies on dedicated factor models and simultaneously determines the number of factors, the allocation of each measurement to a unique factor, and the corresponding factor loadings. Classical identification criteria are applied and integrated into our Bayesian procedure to generate models that are stable and clearly interpretable. A Monte Carlo study confirms the validity of the approach. The method is used to produce interpretable low dimensional aggregates from a high dimensional set of psychological measurements.     

Fitting dynamic factor models to non-stationary time series

Factor modelling of a large time series panel has widely proven useful to reduce its cross-sectional dimensionality. This is done by explaining common co-movements in the panel through the existence of a small number of common components, up to some idiosyncratic behaviour of each individual series. To capture serial correlation in the common components, a dynamic structure is used as in traditional (uni- or multivariate) time series analysis of second order structure, i.e. allowing for infinite-length filtering of the factors via dynamic loadings. In this paper, motivated from economic data observed over long time periods which show smooth transitions over time in their covariance structure, we allow the dynamic structure of the factor model to be non-stationary over time by proposing a deterministic time variation of its loadings. In this respect we generalize the existing recent work on static factor models with time-varying loadings as well as the classical, i.e. stationary, dynamic approximate factor model. Motivated from the stationary case, we estimate the common components of our dynamic factor model by the eigenvectors of a consistent estimator of the now time-varying spectral density matrix of the underlying data-generating process. This can be seen as a time-varying principal components approach in the frequency domain. We derive consistency of this estimator in a “double-asymptotic” framework of both cross-section and time dimension tending to infinity. The performance of the estimators is illustrated by a simulation study and an application to a macroeconomic data set.     

Simultaneous Parameter Estimation in Exploratory Factor Analysis: An Expository Review

The classical exploratory factor analysis (EFA) finds estimates for the factor loadings matrix and the matrix of unique factor variances which give the best fit to the sample correlation matrix with respect to some goodness-of-fit criterion. Common factor scores can be obtained as a function of these estimates and the data. Alternatively to the classical EFA, the EFA model can be fitted directly to the data which yields factor loadings and common factor scores simultaneously. Recently, new algorithms were introduced for the simultaneous least squares estimation of all EFA model unknowns. The new methods are based on the numerical procedure for singular value decomposition of matrices and work equally well when the number of variables exceeds the number of observations. This paper provides an account that is intended as an expository review of methods for simultaneous parameter estimation in EFA. The methods are illustrated on Harman's five socio-economic variables data and a high-dimensional data set from genome research.     

Macroeconomic forecasting using approximate factor models with outliers

In this paper we consider estimating an approximate factor model in which candidate predictors are subject to sharp spikes such as outliers or jumps. Given that these sharp spikes are assumed to be rare, we formulate the estimation problem as a penalized least squares problem by imposing a norm penalty function on those sharp spikes. Such a formulation allows us to disentangle the sharp spikes from the common factors and estimate them simultaneously. Numerical values of the estimates can be obtained by solving a principal component analysis (PCA) problem and a one-dimensional shrinkage estimation problem iteratively. In addition, it is easy to incorporate methods for selecting the number of common factors in the iterations. We compare our method with PCA by conducting simulation experiments in order to examine their finite-sample performances. We also apply our method to the prediction of important macroeconomic indicators in the U.S., and find that it can deliver performances that are comparable to those of the PCA method.     

Asymptotics of the principal components estimator of large factor models with weakly influential factors

This paper introduces a drifting-parameter asymptotic framework to derive accurate approximations to the finite sample distribution of the principal components (PC) estimator in situations when the factors’ explanatory power does not strongly dominate the explanatory power of the cross-sectionally and temporally correlated idiosyncratic terms. Under our asymptotics, the PC estimator is inconsistent. We find explicit formulae for the amount of the inconsistency, and propose an estimator of the number of factors for which the PC estimator works reasonably well. For the special case when the idiosyncratic terms are cross-sectionally but not temporally correlated (or vice versa), we show that the coefficients in the OLS regressions of the PC estimates of factors (loadings) on the true factors (true loadings) are asymptotically normal, and find explicit formulae for the corresponding asymptotic covariance matrix. We explain how to estimate the parameters of the derived asymptotic distributions. Our Monte Carlo analysis suggests that our asymptotic formulae and estimators work well even for relatively small n$n$ and T$T$. We apply our theoretical results to test a hypothesis about the factor content of the US stock return data.     

Are more data always better for factor analysis?

Factors estimated from large macroeconomic panels are being used in an increasing number of applications. However, little is known about how the size and the composition of the data affect the factor estimates. In this paper, we question whether it is possible to use more series to extract the factors, and yet the resulting factors are less useful for forecasting, and the answer is yes. Such a problem tends to arise when the idiosyncratic errors are cross-correlated. It can also arise if forecasting power is provided by a factor that is dominant in a small dataset but is a dominated factor in a larger dataset. In a real time forecasting exercise, we find that factors extracted from as few as 40 pre-screened series often yield satisfactory or even better results than using all 147 series. Weighting the data by their properties when constructing the factors also lead to improved forecasts. Our simulation analysis is unique in that special attention is paid to cross-correlated idiosyncratic errors, and we also allow the factors to have stronger loadings on some groups of series than others. It thus allows us to better understand the properties of the principal components estimator in empirical applications.     

CCE estimation of factor‐augmented regression models with more factors than observables

This paper considers estimation of factor‐augmented panel data regression models. One of the most popular approaches towards this end is the common correlated effects (CCE) estimator of Pesaran (Estimation and inference in large heterogeneous panels with a multifactor error structure. Econometrica, 2006, 74, 967–1012, 2006). For the pooled version of this estimator to be consistent, either the number of observables must be larger than the number of unobserved common factors, or the factor loadings must be distributed independently of each other. This is a problem in the typical application involving only a small number of regressors and/or correlated loadings. The current paper proposes a simple extension to the CCE procedure by which both requirements can be relaxed. The CCE approach is based on taking the cross‐section average of the observables as an estimator of the common factors. The idea put forth in the current paper is to consider not only the average but also other cross‐section combinations. Asymptotic properties of the resulting combination‐augmented CCE (C³E) estimator are provided and tested in small samples using both simulated and real data.     

Principal components estimation and identification of static factors

It is known that the principal component estimates of the factors and the loadings are rotations of the underlying latent factors and loadings. We study conditions under which the latent factors can be estimated asymptotically without rotation. We derive the limiting distributions for the estimated factors and factor loadings when N$N$ and T$T$ are large and make precise how identification of the factors affects inference based on factor augmented regressions. We also consider factor models with additive individual and time effects. The asymptotic analysis can be modified to analyze identification schemes not considered in this analysis.     

Consistent noisy independent component analysis

We study linear factor models under the assumptions that factors are mutually independent and independent of errors, and errors can be correlated to some extent. Under the factor non-Gaussianity, second-to-fourth-order moments are shown to yield full identification of the matrix of factor loadings. We develop a simple algorithm to estimate the matrix of factor loadings from these moments. We run Monte Carlo simulations and apply our methodology to data on cognitive test scores, and financial data on stock returns.     

Estimating a common deterministic time trend break in large panels with cross sectional dependence

This paper develops an estimation procedure for a common deterministic time trend break in large panels. The dependent variable in each equation consists of a deterministic trend and an error term. The deterministic trend is subject to a change in the intercept, slope or both, and the break date is common for all equations. The estimation method is simply minimizing the sum of squared residuals for all possible break dates. Both serial and cross sectional correlations are important factors that decide the rate of convergence and the limiting distribution of the break date estimate. The rate of convergence is faster when the errors are stationary than when they have a unit root. When there is no cross sectional dependence among the errors, the rate of convergence depends on the number of equations and thus is faster than the univariate case. When the errors have a common factor structure with factor loadings correlated with the intercept and slope change parameters, the rate of convergence does not depend on the number of equations and thus reduces to the univariate case. The limiting distribution of the break date estimate is also provided. Some Monte Carlo experiments are performed to assess the adequacy of the asymptotic results. A brief empirical example using the US GDP price index is offered.     

Generalized M-fluctuation tests for parameter instability

A general class of fluctuation tests for parameter instability in an M-estimation framework is suggested. Tests from this framework can be constructed by first choosing an appropriate estimation technique, deriving a partial sum process of the estimation scores that captures instabilities over time, and aggregating this process to a test statistic by using a suitable scalar functional. Inference for these tests is based on functional central limit theorems, which are derived under the null hypothesis of parameter stability and local alternatives. For (generalized) linear regression models, concrete tests are derived, which cover several known tests for (approximately) normal data but also allow for testing for parameter instability in regressions with binary or count data. The usefulness of the test procedures—complemented by powerful visualizations derived from these—is illustrated using Dow Jones industrial average stock returns, youth homicides in Boston, USA, and illegitimate births in Grossarl, Austria.     

Bayesian forecasting and portfolio decisions using dynamic dependent sparse factor models

We extend the recently introduced latent threshold dynamic models to include dependencies among the dynamic latent factors which underlie multivariate volatility. With an ability to induce time-varying sparsity in factor loadings, these models now also allow time-varying correlations among factors, which may be exploited in order to improve volatility forecasts. We couple multi-period, out-of-sample forecasting with portfolio analysis using standard and novel benchmark neutral portfolios. Detailed studies of stock index and FX time series include: multi-period, out-of-sample forecasting, statistical model comparisons, and portfolio performance testing using raw returns, risk-adjusted returns and portfolio volatility. We find uniform improvements on all measures relative to standard dynamic factor models. This is due to the parsimony of latent threshold models and their ability to exploit between-factor correlations so as to improve the characterization and prediction of volatility. These advances will be of interest to financial analysts, investors and practitioners, as well as to modeling researchers.     

An endogenously clustered factor approach to international business cycles

Factor models have become useful tools for studying international business cycles. Block factor models can be especially useful as the zero restrictions on the loadings of some factors may provide some economic interpretation of the factors. These models, however, require the econometrician to predefine the blocks, leading to potential misspecification. In Monte Carlo experiments, we show that even a small misspecification can lead to substantial declines in fit. We propose an alternative model in which the blocks are chosen endogenously. The model is estimated in a Bayesian framework using a hierarchical prior, which allows us to incorporate series‐level covariates that may influence and explain how the series are grouped. Using international business cycle data, we find our country clusters differ in important ways from those identified by geography alone. In particular, we find that similarities in institutions (e.g., legal systems, language diversity) may be just as important as physical proximity for analyzing business cycle comovements.     

The robustness of maximum likelihood and distribution-free estimators to non-normality in confirmatory factor analysis

The present Monte Carlo compares the estimates produced by maximum likelihood (ML) and asymptotically distribution-free (ADF) methods. The study extends prior research by investigating the combined effects of sample size, magnitude of correlation among observed indicators, number of indicators, magnitude of skewness and kurtosis, and proportion of indicators with non-normal distributions. Results indicate that both ML and ADF showed little bias in estimates of factor loadings under all conditions studied. As the number of indicators in the model increased, ADF produced greater negative bias in estimates of uniquenesses than ML. In addition, the bias in standard errors for both ML and ADF estimation increased in models with more indicators, and this effect was more pronounced for ADF than ML. Increases in skewness and kurtosis resulted in greater underestimating of standard errors; ML standard errors showed greater bias than ADF under conditions of non-normality, and ML chi-square statistics were also inflated. However, when only half the indicators departed from normality, the inflation in ML chi-square decreased.