DSpace Collection: IOMS: Statistics Working Papers

no title

Sun, 25 May 2008 16:03:22 GMT

Were the 1996-2000 Yankees the Best Baseball Team Ever?

Sun, 29 Oct 2000 22:58:59 GMT

Title: Were the 1996-2000 Yankees the Best Baseball Team Ever?

Simon, Gary A.; Simonoff, Jeffrey S.

USING ORDER STATISTICS TO ESTIMATE PROBABILITIES OF PURCHASE FOR CONSUMER GOODS

Wed, 24 Jul 2002 22:58:59 GMT

Title: USING ORDER STATISTICS TO ESTIMATE PROBABILITIES OF PURCHASE FORCONSUMER GOODS

TASHJIAN, RICHARD H.; NEELANKAVIL, JAMES P.

Abstract: Much work has been done recently to develop models of individual brandchoice. This work has been especially fruitful in the area of“attribute investigation” (e.g., conjoint analysis) for thepurpose of uncovering an “ideal” set of product attributesfor a given product on a customer-by-customer basis. These techniqueshave been employed in several strategic areas of marketing, includingproduct specification, pricing, service prioritization, brandimage/equity, and satisfaction/loyalty. Similarly, some effort has beenmade with the issue of optimal advertising strategy. This paperconsiders the advertising effectiveness function within the context ofother interrelated variables such as consumer preference (brand choice)for a brand vis-à-vis its competitors. The model suggests, amongother things, that under certain reasonable conditions, the advertisingresponse function may not be “diminishing marginal returns”or ‘S-shaped’ as is usually assumed, but instead willincrease up to a point and then decline. The model also explicitlyconsiders the advertising expenditures of competing brands, as well asintrinsic "liking" for them. The consideration of competitiveactivity, in the present study may yield a more complete model ofadvertising response than is found elsewhere. Finally the model providesdirection for strategic purposes in its ability to illustrate in afairly straightforward and graphical sense how advertising and promotion“work” in much the same way that demand and supply curvesillustrate how the various economic inputs work.

Tree Induction vs Logistic Regression: A Learning Curve Analysis

Wed, 12 Dec 2001 22:58:59 GMT

Title: Tree Induction vs Logistic Regression: A Learning Curve Analysis

Perlich, Claudia; Provost, Foster; Simonoff, Jeffrey S.

Transformation-based density estimation for weighted distributions

Thu, 29 Oct 1998 22:58:59 GMT

Title: Transformation-based density estimation for weighted distributions

El Barmi, Hammou; Simonoff, Jeffrey S.

Abstract: In this paper we consider the estimation of a density f on the basis ofrandom sample from a weighted distribution G with density g given byg(x) = w(x)f(x)/ ÃÂµw, where w(u)>0 for all u andÃÂµw = Ã¢ÂÂ«w(u)f(u)du < Ã¢ÂÂ. A special case of thissituation is that of length-biased sampling, where w(x) = x. In thispaper we examine a simple transformation-based approach to estimatingthe density f. The approach is motivated by the form of thenonparametric estimator of f in the same context and under amonotonicity constraint. Since the method does not depend on thespecific density estimate used (only the transformation), it can be usedto construct both simple density estimates (histograms or frequencypolygons) and more complex methods with favorable properties (e.g.,local or penalized likelihood estimates). Monte Carlo simulationsindicate that transformation-based density estimation can outperform thekernel-based estimator of Jones (1991) depending on the weight functionw, and leads to much better estimation of monotone densities than thenonparametric maximum likelihood estimator.

Tracing the Source of Long Memory in Volatility

Wed, 12 Jan 2005 22:58:59 GMT

Title: Tracing the Source of Long Memory in Volatility

Deo, Rohit; Hsieh, Mengchen; Hurvich, Clifford M.

Abstract: We study the effects of trade duration properties on dependence incounts (number of trans-actions) and thus on dependence in volatility ofreturns. A return model is established to link counts and volatility. Wepresent theorems as well as a conjecture relating properties ofdurations to long memory in counts and thus in volatility. We then applyseveral parametric duration models to empirical trade durations anddiscuss our findings in the light of the theorems and conjecture.

Toward Enhancing the Quality and Quantity of Marketing Majors

Tue, 29 Oct 1996 22:58:59 GMT

Title: Toward Enhancing the Quality and Quantity of Marketing Majors

LaBarbera, Priscilla A.; Simonoff, Jeffrey S.

Abstract: This article reports the findings of a survey of undergraduate studentsdesigned to examine the key factors involved in selecting a marketingmajor. A discussion follows, dealing with the initiatives undertaken bymarketing departments at various universities in an attempt to enhancethe quality and quantity of marketing majors.

Tobit Model Estimation and Sliced Inverse Regression

Sat, 29 Oct 2005 22:58:59 GMT

Title: Tobit Model Estimation and Sliced Inverse Regression

Li, Lexin; Simonoff, Jeffrey S.; Tsai, Chih-Ling

Abstract: It is not unusual for the response variable in a regression model to besubject to censoring or truncation. Tobit regression models are aspecific example of such a situation, where for some observations theobserved response is not the actual response, but rather the censoringvalue (often zero), and an indicator that censoring (from below) hasoccurred. It is well-known that the maximum likelihood estimator forsuch a linear model (assuming Gaussian errors) is not consistent if theerror term is not homoscedastic and normally distributed. In this paperwe consider estimation in the Tobit regression context when thoseconditions do not hold, as well as when the true response is anunspecified nonlinear function of linear terms, using sliced inverseregression (SIR). The properties of SIR estimation for Tobit models areexplored both theoretically and based on Monte Carlo simulations. It isshown that the SIR estimator has good properties when the usual linearmodel assumptions hold, and can be much more effective than otherestimators when they do not. An example related to household charitabledonations demonstrates the usefulness of the estimator.

Three Sides of Smoothing: Categorical Data Smoothing, Nonparametric Regression, and Density Estimation

Tue, 29 Oct 1996 22:58:59 GMT

Title: Three Sides of Smoothing: Categorical Data Smoothing, NonparametricRegression, and Density Estimation

Simonoff, Jeffrey S.

Abstract: The past forty years have seen a great deal of research into theconstruction and properties of nonparametric estimates of smoothfunctions. This research has focused primarily on two sides of thesmoothing problem: nonparametric regression and density estimation.Theoretical results for these two situations are similar, andmultivariate density estimation was an early justification for theNadaraya-Watson kernel regression estimator. A third, lesswell-explored, strand of applications of smoothing is to the estimationof probabilities in categorical data. In this paper the position ofcategorical data smoothing as a bridge between nonparametric regressionand density estimation is explored. Nonparametric regression provides aparadigm for the construction of effective categorical smoothingestimates, and use of an appropriate likelihood function yields cellprobability estimates with many desirable properties. Such estimates canbe used to construct regression estimates when one or more of thecategorical variables are viewed as response variables. They also leadnaturally to the construction of well-behaved density estimates usinglocal or penalized likelihood estimation, which can then be used in aregression context. Several real data sets are used to illustrate these points.

The Restricted Likelihood Ratio Test at the Boundary in Autoregressive Series

Mon, 24 Aug 2009 19:51:18 GMT

Title: The Restricted Likelihood Ratio Test at the Boundary in Autoregressive Series

Chen, Willa W.; Deo, Rohit S.

Abstract: The restricted likelihood ratio test, RLRT, for the autoregressivecoefficient in autoregressive models has recently been shown to besecond order pivotal when the autoregressive coefficient is in theinterior of the parameter space and so is very well approximated by thechi-square distribution. In this paper, the non-standard asymptoticdistribution of the RLRT for the unit root boundary value is obtainedand is found to be almost identical to that of the chi-square in theright tail. Together, the above two results imply that the chi-squaredistribution approximates the RLRT distribution very well even for nearunit root series and transitions smoothly to the unit root distribution.

The Local Whittle Estimator of Long Memory Stochastic Volatility

Mon, 28 Apr 2003 22:58:59 GMT

Title: The Local Whittle Estimator of Long Memory Stochastic Volatility

Hurvich, Clifford M.; Ray, Bonnie K.

Abstract: We propose a new semiparametric estimator of the degree of persistencein volatility for long memory stochastic volatility (LMSV) models. Theestimator uses the periodogram of the log squared returns in a localWhittle criterion which explicitly accounts for the noise term in theLMSV model. Finite-sample and asymptotic standard errors for theestimator are provided. An extensive simulation study reveals that thelocal Whittle estimator is much less biased and that the finite-samplestandard errors yield more accurate confidence intervals than thewidely-used GPH estimator. The estimator is also found to be robustagainst possible leverage effects. In an empirical analysis of the dailyDeutsche Mark/US Dollar exchange rate, the new estimator indicatesstronger persistence in volatility than the GPH estimator, provided thata large number of frequencies is used.

The Local Whittle Estimator of Long Memory Stochastic Volatility

Thu, 29 Mar 2001 22:58:59 GMT

Title: The Local Whittle Estimator of Long Memory Stochastic Volatility

Hurvich, Clifford M.; Ray, Bonnie K.

Abstract: We propose a new semiparametric estimator of the degree of persistencein volatility for long memory stochastic volatility (LMSV) models. Theestimator uses the periodogram of the log squared returns in a localWhittle criterion which explicitly accounts for the noise term in theLMSV model. Finite-sample and asymptotic standard errors for theestimator are provided. An extensive simulation study reveals that thelocal Whittle estimator is much less biased and yields more accurateconfidence intervals than the widely-used GPH estimator. In an empiricalanalysis of the daily Deutschemark/Dollar exchange rate, the newestimator indicates stronger persistence in volatility than the GPHestimator, provided that a large number of frequencies is used.

The Emergence of Concentrated Ownership and the Rebalancing of Portfolios due to Shareholder Activism in a Financial Market Equilibrium

Wed, 29 Dec 1999 22:58:59 GMT

Title: The Emergence of Concentrated Ownership and the Rebalancing ofPortfolios due to Shareholder Activism in a Financial Market Equilibrium

Katz, Barbara G.; Owen, Joel

Abstract: Consider a financial market equilibrium with correlated firms and riskaverse investors holding diversified portfolios. When an activistinvestor has the ability to perform value-enhancing activities in asingle firm, and these activities increase with ownership, we show thatoptimizing behavior by all investors leads to a concentration of sharesin the hands of this activist. This concentration arises in thepresence of complete information and is a consequence of Walrasianequilibrium mechanisms that include all investors and give no specialpowers to any of them in the equilibrium process. By yielding moreownership to the activist, all investors alter the risk profiles oftheir holdings, ending with less balanced portfolios. This rebalancingeffect is accompanied by an increase in the price of the security thatthe activist can affect, as well as in total value of the market. Whenthe activist can affect more than one firm, rebalancing of allportfolios again occurs. Although the activist may not acquireincreased concentration in all the firms she might affect, prices changefor all those firms, and we give conditions under which at least oneprice must increase. We find that equilibrium results in a sharing ofthe costs and benefits of activism among all market participants,mitigating the free-rider problem. When we study multiple activists inmany firms, we show that concentration can occur for several activists,and rebalancing occurs for all investors. Predictions oninvestor-specific concentration are difficult and excessive portfoliochurning is present. The introduction of asymmetric informationconcerning activism again results in rebalancing and in concentration ofownership, but not necessarily in the hands of the activist.

The Conditional Breakdown Properties of Robust Local Polynomial Estimators

Thu, 13 Dec 2001 22:58:59 GMT

Title: The Conditional Breakdown Properties of Robust Local Polynomial Estimators

Giloni, Avi; Simonoff, Jeffrey S.

Abstract: Nonparametric regression techniques provide an e ective way ofidentifying and examining structure in regression data The standardapproaches to nonparametric regression such as local polynomial andsmoothing spline estimators are sensitive to unusual observations andalternatives designed to be resistant to such observations have beenproposed as a solution unfortunately there has been little examinationof the resistance properties of these proposed estimators In this paperwe examine the breakdown properties of several robust versions of localpolynomial estimation We show that for some estimators the breakdown atany evaluation point depends on the observed distribution ofobservations and the kernel weight function used Using synthetic andreal data we show how the breakdown point at an evaluation pointprovides a useful summary of the resistance of the regression estimatorto unusual obseravions

The Conditional Breakdown Properties of Least Absolute Value Local Polynomial Estimators

Sat, 06 Dec 2003 22:58:59 GMT

Title: The Conditional Breakdown Properties of Least Absolute Value LocalPolynomial Estimators

Giloni, Avi; Simonoff, Jeffrey S.

Abstract: Nonparametric regression techniques provide an effective way ofidentifying and examining structure in regression data. The standardapproaches to nonparametric regression, such as local polynomial andsmoothing spline estimators, are sensitive to unusual observations, andalternatives designed to be resistant to such observations have beenproposed as a solution. Unfortunately, there has been little examinationof the resistance properties of these proposed estimators. In this paperwe examine the breakdown properties of local polynomial estimation basedon least absolute values, rather than least squares. We show that thebreakdown at any evaluation point depends on the observed distributionof observations and the kernel weight function used, and makerecommendations regarding choice of kernel based on two differentbreakdown measures. The results suggest that the breakdown point at anevaluation point provides a useful summary of the resistance of theregression estimator to unusual observations.

The Averaged Periodogram Estimator for a Power Law in Coherency

Tue, 27 Sep 2011 17:30:31 GMT

Title: The Averaged Periodogram Estimator for a Power Law in Coherency

Sela, Rebecca J.; Hurvich, Clifford M.

Abstract: We prove the consistency of the averaged periodogram estimator (APE) intwo new cases. First, we prove that the APE is consistent for negativememory parameters, after suitable tapering. Second, we prove that theAPE is consistent for a power law in the cross-spectrum and thereforefor a power law in the coherency, provided that sufficiently manyfrequencies are used in estimation. Simulation evidence suggests thatthe lower bound on the number of frequencies is a necessary conditionfor consistency. For a Taylor series approximation to the estimator ofthe power law in the cross-spectrum, we consider the rate ofconvergence, and obtain a central limit theorem under suitableregularity conditions.

TESTING FOR LONG MEMORY IN VOLATILITY

Fri, 29 Oct 1999 22:58:59 GMT

Title: TESTING FOR LONG MEMORY IN VOLATILITY

Hurvich, Clifford M.; Soulier, Philippe

Abstract: We consider the asymptotic behavior of log-periodogram regressionestimators of the memory parameter in long-memory stochastic volatilitymodels, under the null hypothesis of short memory in volatility. We showthat in this situation, if the periodogram is computed from the logsquared returns, then the estimator is asymptotically normal, with thesame asymptotic mean and variance that would hold if the series wereGaussian. In particular, for the widely used GPH estimator dGPH underthe null hypothesis, the asymptotic mean ofmÃÂ½dGPH is zero and the asymptotic variance ispiÃÂ²/24 where m is the number of Fourier frequenciesused in the regression. This justifies an ordinary Wald test for longmemory in volatility based on the log periodogram of the log squared returns.

Structural Change Tests in Tail Behaviour and the Asian Crisis

Tue, 29 Dec 1998 22:58:59 GMT

Title: Structural Change Tests in Tail Behaviour and the Asian Crisis

Quintos, Carmela; Fan, Zhenhong; Phillips, Peter C. B.

Abstract: This paper explores tests of the hypothesis that the tail thickness of adistribution is constant over time. Using Hill's conditional maximumlikelihood estimator for the tail index of a distribution, tests of tailshape constancy are constructed that allow for an unknown breakpoint.The recursive test IS shown to be inconsistent in one direction, andonly a one-sided test is recommended. Specifically, the test can be usedwhen the alternative hypothesis is that the tail index decreases overtime. A rolling and sequential version of the test is consistent in bothdirections. The methods are illustrated on recent stock price data forThailand, Malaysia and Indonesia. The period covers the recent Asianfinancial crisis and enables us to assess whether breakpoints indomestic asset return distributions are related to known changes ininstitutional arrangements in the foreign currency markets of these countries.

Spectral tests of the martingale hypothesis under conditional heteroscedasticity

Tue, 29 Oct 1996 22:58:59 GMT

Title: Spectral tests of the martingale hypothesis under conditional heteroscedasticity

Deo, Rohit S.

Abstract: We study the asymptotic distribution of the sample standardized spectraldistribution function when the observed series is a conditionallyheteroscedastic martingale difference. We show that the asymptoticdistribution is no longer a Brownian bridge but another Gaussianprocess. Furthermore, this limiting process depends on the covariancestructure of the second moments of the series. We show that this causestest statistics based on the sample spectral distribution, such as theCramÃÂÃÂ©r von-Mises statistic, tohave heavily right skewed distributions, which will lead toover-rejection of the martingale hypothesis in favour of mean reversion.A non-parametric correction to the test statistics is proposed toaccount for the conditional heteroscedasticity. We demonstrate that thecorrected version of the CramÃÂÃÂ©rvon-Mises statistic has the usual limiting distribution which would beobtained in the absence of conditional heteroscedasticity. We alsopresent Monte Carlo results on the finite sample distributions ofuncorrected and corrected versions of theCramÃÂÃÂ©r von-Mises statistic. Oursimulation results show that this statistic can provide significantgains in power over the Box-Ljung-Pierce statistic against long-memoryalternatives. An empirical application to stock returns is also provided.

SPATIAL REGRESSION MODELS USING INTER-REGION DISTANCES IN A NON-RANDOM CONTEXT

Fri, 29 Oct 1999 22:58:59 GMT

Title: SPATIAL REGRESSION MODELS USING INTER-REGION DISTANCES IN A NON-RANDOM CONTEXT

Christou, Nicolas; Simon, Gary

Abstract: This paper considers spatial data z, z(s2), z(sn) collected at nlocations, with the objective of predicting z (s0) at another location.The usual method of analysis for this problem is kriging, but here weintroduce a new signal-plus-noise model whose essential feature is theidentification of hot spots. The signal decays in relation to distancefrom hot spots. We show that hot spots can be located with highaccuracy and that the decay parameter can be estimated accurately. Thisnew model compares well to kriging in simulations.

SOLVING A CLASS OF TRAVELING SALESMAN PROBLEMS ANALYTICALLY

Thu, 29 May 2003 22:58:59 GMT

Title: SOLVING A CLASS OF TRAVELING SALESMAN PROBLEMS ANALYTICALLY

TASHJIAN, RICHARD H.

Abstract: This paper addresses a class of Traveling Salesman Problems (TSP) inwhich a route must be made to a series of nodes and return to theoriginal location and attempts to solve it using analytical methods. Theproblem will be presented as a matrix of routes, much as might be seenin a national road map, excepting for there being in this case lessentries. This familiar arrangement of routes will be cast as a matrixproblem and solved using familiar formulations of quadratic forms. Thissolution, should it prove successful, can be contrasted with differingnumeric or even iterative methods, such as the well-known Gomory cutmethod of solving integer linear programs. The advantage, should itprove tenable, will be theoretic in that a familiar and accessible formof quadratic forms can be readily applied to the problem and to similar cases.

Should Governments Compete for Foreign Direct Investment?

Sat, 29 Mar 2003 22:58:59 GMT

Title: Should Governments Compete for Foreign Direct Investment?

Katz, Barbara G.; Owen, Joel

Abstract: We study two governments, each considering whether or not to compete toattract a foreign monopoly firm into its own domestic market. Thecompetition, should it occur, would involve offering incentives to thefirm. The incentives, which are costly for the governments to provide,lower the firm s marginal cost of production. Faced with the offers fromeach country, the firm must choose one of four options: to enter eitherof the markets, produce there and export to the other, to enter bothmarkets simultaneously with only local production, or to reject alloffers. We find conditions under which it would be optimal for one ofthe two countries not to compete with the other, preferring instead toimport the commodity from the country that attracted the firm, ratherthan incurring the additional costs that would have been necessary tomake its own economy more attractive to the foreign firm. We show thatwhen importing the good is a possibility, there are conditions underwhich, knowing that it will lose (win) the competition for the firm, thecountry nonetheless finds it optimal to (not) compete. Also, we derivethe market structure by establishing the relationship between the optionchosen by the firm and the characteristics of the two governments tryingto attract the firm.

Semiparametric Estimation of Multivariate Fractional

Thu, 13 Jun 2002 22:58:59 GMT

Title: Semiparametric Estimation of Multivariate Fractional

Chen, Willa W.; Hurvich, Califford M.

Abstract: We consider the semiparametric estimation of fractional cointegration ina multivariate process of cointegrating rank r > 0. We estimate thecointegrating relationships by the eigenvectors corresponding to the rsmallest eigenvalues of an averaged periodogram matrix of tapered,differenced observations. The number of frequencies m used in theperiodogram average is held fixed as the sample size grows. We firstshow that the averaged periodogram matrix converges in distribution to asingular matrix whose null eigenvectors span the space of cointegratingvectors. We then show that the angle between the estimated cointegratingvectors and the space of true cointegrating vectors is Op(nduôd)where d and du are the memory parameters of the observations andcointegrating errors, respectively. The proposed estimator is invariantto the labeling of the component series, and therefore does not requireone of the variables to be specified as a dependent variable. Wedetermine the rate of convergence of the r smallest eigenvalues of theperiodogram matrix, and present a criterion which allows for consistentestimation of r. Finally, we apply our methodology to the analysis offractional cointegration in interest rates.

Semiparametric Estimation of Fractional Cointegrating Subspaces

Thu, 18 Nov 2004 22:58:59 GMT

Title: Semiparametric Estimation of Fractional Cointegrating Subspaces

Chen, Willa W.; Hurvich, Clifford M.

Abstract: We consider a common components model for multivariate fractionalcointegration, in which the s ¸ 1 components have different memoryparameters. The cointegrating rank is allowed to exceed 1. The truecointegrating vectors can be decomposed into orthogonal fractionalcointegrating subspaces such that vectors from distinct subspaces yieldcointegrating errors with distinct memory parameters, denoted by dk, fork = 1; : : : ; s. We estimate each cointegrating subspace separatelyusing appropriate sets of eigenvectors of an averaged periodogram matrixof tapered, differenced observations. The averaging uses the first mFourier frequencies, with m fixed. We will show that any vector in thek’th estimated cointegrating subspace is, with high probability,close to the k’th true cointegrating subspace, in the sense thatthe angle between the estimated cointegrating vector and the truecointegrating subspace converges in probability to zero. This angle isOp(n¡®k ), where n is the sample size and ®k is theshortest distance between the memory parameters corresponding to thegiven and adjacent subspaces. We show that the cointegrating residualscorresponding to an estimated cointegrating vector can be used to obtaina consistent and asymptotically normal estimate of the memory parameterfor the given cointegrating subspace, using a univariate Gaussiansemiparametric estimator with a bandwidth that tends to 1 more slowlythan n. We also show how these memory parameter estimates can be used totest for fractional cointegration and to consistently identify thecointegrating subspaces.

Semiparametric and Additive Model Selection Using an Improved Akaike Information Criterion

Tue, 29 Oct 1996 22:58:59 GMT

Title: Semiparametric and Additive Model Selection Using an Improved AkaikeInformation Criterion

Simonoff, Jeffrey S.; Tsai, Chih-Ling

Abstract: An improved AIC-based criterion is derived for model selection ingeneral smoothing-based modeling, including semiparametric models andadditive models. Examples are provided of applications togoodness-of-fit, smoothing parameter and variable selection in anadditive model and semiparametric models, and variable selection in amodel with a nonlinear function of linear terms.

Score Tests for the Single Index Model

Fri, 29 Oct 1999 22:58:59 GMT

Title: Score Tests for the Single Index Model

Simonoff, Jeffrey S.; Tsai, Chih-Ling

Abstract: The single index model is a generalization of the linear regressionmodel with E(y|x) = g, where g is an unknown function. The modelprovides a flexible alternative to the linear regression model whileproviding more structure than a fully nonparametric approach. Althoughthe fitting of single index models does not require distributionalassumptions on the error term, the properties of the estimates depend onsuch assumptions, as does practical application of the model. In thisarticle score tests are derived for three potential misspecifications ofthe single index model: heteroscedasticity in the errors,autocorrelation in the errors, and the omission of an important variablein the linear index. These tests have a similar structure tocorresponding tests for nonlinear regression models. Monte Carlosimulations demonstrate that the first two tests hold their nominal sizewell and have good power properties in identifying model violations,often outperforming other tests. Testing for the need for additionalcovariates can be effective, but is more difficult. The score tests areapplied to three real datasets, demonstrating that the tests canidentify important model violations that affect inference, and thatapproaches that do not take model misspecifications into account canlead to very different results.

Robust Weighted LAD Regression

Thu, 24 Feb 2005 22:58:59 GMT

Title: Robust Weighted LAD Regression

Giloni, Avi; Simonoff, Jeffrey S.; Sengupta, Bhaskar

Abstract: The least squares linear regression estimator is well-known to be highlysensitive to unusual observations in the data, and as a result many morerobust estimators have been proposed as alternatives. One of theearliest proposals was least-sum of absolute deviations (LAD)regression, where the regression coefficients are estimated throughminimization of the sum of the absolute values of the residuals. LADregression has been largely ignored as a robust alternative to leastsquares, since it can be strongly affected by a single observation (thatis, it has a breakdown point of 1/n, where n is the sample size). Inthis paper we show that judicious choice of weights can result in aweighted LAD estimator with much higher breakdown point. We discuss theproperties of the weighted LAD estimator, and show via simulation thatits performance is competitive with that of high breakdown regressionestimators, particularly in the presence of outliers located at leveragepoints. We also apply the estimator to several real data sets.

Robust Analysis of Variance: Process Design and Quality Improvement

Thu, 28 Apr 2005 22:58:59 GMT

Title: Robust Analysis of Variance: Process Design and Quality Improvement

Giloni, Avi; Seshadri, Sridhar; Simonoff, Jeffrey S.

Abstract: We discuss the use of robust analysis of variance (ANOVA) techniques asapplied to quality engineering. ANOVA is the cornerstone for uncoveringthe effects of design factors on performance. Our goal is to utilizemethodologies that yield similar results to standard methods when theunderlying assumptions are satisfied, but also are relatively unaffectedby outliers (observations that are inconsistent with the general patternin the data). We do this by utilizing statistical software to implementrobust ANOVA methods, which are no more difficult to perform thanordinary ANOVA. We study several examples to illustrate how usingstandard techniques can lead to misleading inferences about the processbeing examined, which are avoided when using a robust analysis. Wefurther demonstrate that assessments of the importance of factors forquality design can be seriously compromised when utilizing standardmethods as opposed to robust methods.

REAL-TIME MULTIVARIATE DENSITY FORECAST EVALUATION AND CALIBRATION: MONITORlNG THE RISK OF HIGH-FREQUENCY RETURNS ON FOREIGN EXCHANGE

Sat, 28 Nov 1998 22:58:59 GMT

Title: REAL-TIME MULTIVARIATE DENSITY FORECAST EVALUATION AND CALIBRATION:MONITORlNG THE RISK OF HIGH-FREQUENCY RETURNS ON FOREIGN EXCHANGE

Diebold, Francis X.; Hahn, Jinyong; Tay, Anthony S.

Abstract: We provide a framework for evaluating and improving multivariate densityforecasts. Among other things, the multivariate framework lets usevaluate the adequacy of density forecasts involving cross-variableinteractions, such as time-varying conditional correlations. We alsoprovide conditions under which a technique of density forecast"calibration" can be used to improve deficient densityforecasts. Finally, motivated by recent advances in financial riskmanagement, we provide a detailed application to multivariatehigh-frequency exchange rate density forecasts.

RE-EM Trees: A New Data Mining Approach for Longitudinal Data

Wed, 10 Jun 2009 13:04:23 GMT

Title: RE-EM Trees: A New Data Mining Approach for Longitudinal Data

Sela, Rebecca J.; Simonoff, Jeffrey S.

Abstract: Longitudinal data refer to the situation where repeated observations areavailable for each sampled individual. Methodologies that take thisstructure into account allow for systematic differences betweenindividuals that are not related to covariates. A standard methodologyin the statistics literature for this type of data is the random effectsmodel, where these differences between individuals are represented byso-called “effects” that are estimated from the data. Thispaper presents a methodology that combines the flexibility of tree-basedestimation methods with the structure of random effects models forlongitudinal data. We apply the resulting estimation method, called theRE-EM tree, to pricing in online transactions, showing that the RE-EMtree is less sensitive to parametric assumptions and provides improvedpredictive power compared to linear models with random effects andregression trees without random effects. We also perform extensivesimulation experiments to show that the estimator improves predictiveperformance relative to regression trees without random effects and iscomparable or superior to using linear models with random effects inmore general situations.

Propagation of Memory Parameter from Durations to Counts

Mon, 07 Nov 2005 22:58:59 GMT

Title: Propagation of Memory Parameter from Durations to Counts

Deo, Rohit; Hurvich, Clifford M.; Soulier, Philippe; Wang, Yi

Abstract: We establish sufficient conditions on durations that are stationary withfinite variance and memory parameter d 2 [0; 1=2) to ensure that thecorresponding counting process N(t) satisfies VarN(t) » Ct2d+1 (C> 0) as t ! 1, with the same memory parameter d 2 [0; 1=2) that wasassumed for the durations. Thus, these conditions ensure that the memoryin durations propagates to the same memory parameter in counts andtherefore in realized volatility. We then show that any AutoregressiveConditional Duration ACD(1,1) model with a sufficient number of finitemoments yields short memory in counts, while any Long Memory StochasticDuration model with d > 0 and all finite moments yields long memoryin counts, with the same d. Finally, we present a result implying thatthe only way for a series of counts aggregated over a long time periodto have nontrivial autocorrelation is for the short-term counts to havelong memory. In other words, aggregation ultimately destroys allautocorrelation in counts, if and only if the counts have short memory.

Predictive Regressions: A Reduced-Bias Estimation Method

Mon, 03 May 2004 22:58:59 GMT

Title: Predictive Regressions: A Reduced-Bias Estimation Method

Amihud, Yakov; Hurvich, Clifford M.

Abstract: Standard predictive regressions produce biased coefficient estimates insmall samples when the regressors are Gaussian first-orderautoregressive with errors that are correlated with the error series ofthe dependent variable; see Stambaugh (1999) for the single-regressormodel. This paper proposes a direct and convenient method to obtainreduced-bias estimators for single and multiple regressor models byemploying an augmented regression, adding a proxy for the errors in theautoregressive model. We derive bias expressions for both the ordinaryleast squares and our reduced-bias estimated coefficients. For thestandard errors of the estimated predictive coefficients we develop aheuristic estimator which performs well in simulations, for both thesingle-predictor model and an important specification of themultiple-predictor model. The effectiveness of our method isdemonstrated by simulations and by empirical estimates of commonpredictive models in finance. Our empirical results show that some ofthe predictive variables that were significant under ordinary leastsquares become insignificant under our estimation procedure.

Predictive Regressions: A Reduced-Bias Estimation Method

Mon, 25 Nov 2002 22:58:59 GMT

Title: Predictive Regressions: A Reduced-Bias Estimation Method

Amihud, Yakov; Hurvich, Clifford M.

Abstract: We propose a direct and convenient reduced-bias estimator of predictiveregression coefficients, assuming that the regressors are Gaussianfirst-order autoregressive with errors that are correlated with theerror series of the dependent variable. For the single-regressor model,Stambaugh (1999) shows that the ordinary least squares estimator of thepredictive regression coefficient is biased in small samples. Ourestimation method employs an augmented regression which uses a proxy forthe errors in the autoregressive model. We also develop a heuristicestimator of the standard error of the estimated predictive coefficientwhich performs well in simulations, and show that the estimatedcoefficient of the errors and its squared standard error are unbiased.We analyze the case of multiple predictors that are first-orderautoregressive and derive bias expressions for both the ordinary leastsquares and our reduced-bias estimated coefficients. The effectivenessof our estimation method is demonstrated by simulations.

Predictive Regressions: A Reduced-Bias Estimation Method

Mon, 03 May 2004 22:58:59 GMT

Predictive Regression With Order-p Autoregressive Predictor

Fri, 11 Dec 2009 20:19:59 GMT

Title: Predictive Regression With Order-p Autoregressive Predictor

Amihud, Yakov; Hurvich, Clifford M.; Wang, Yi

Abstract: Studies of predictive regressions analyze the case where yt is predictedby xt-1 with xt being first-order autoregressive, AR(1). Under someconditions, the OLS- estimated predictive coefficient is known to bebiased. We analyze a predictive model where yt is predicted by xt-1,xt-2,... xt-p with xt being autoregressive of order p, AR(p) with p >1. We develop a generalized augmented regression method that produces areduced-bias point estimate of the predictive coefficients and derive anappropriate hypothesis testing procedure. We apply our method to theprediction of quarterly stock returns by dividend yield, which isapparently AR(2). Using our method results in the AR(2) predictor serieshaving insignificant effect, although under OLS, or the commonly assumedAR(1) structure, the predictive model is significant. We also generalizeour method to the case of multiple AR(p) predictors.

Predicting movie grosses: Winners and losers, blockbusters and sleepers

Thu, 29 Oct 1998 22:58:59 GMT

Title: Predicting movie grosses: Winners and losers, blockbusters and sleepers

Simonoff, Jeffrey S.; Sparrow, Ilana R.

Possible Sharing Arrangements in ARMA Supply Chains

Fri, 05 Oct 2012 15:50:20 GMT

Title: Possible Sharing Arrangements in ARMA Supply Chains

Kovtun, Vladimir; Giloni, Avi; Hurvich, Clifford

Abstract: We introduce a class of new sharing arrangements in a multi-stage supplychain in which the retailer observes stationary autoregressive movingaverage demand with Gaussian white noise (shocks). Similar to previousresearch, we assume each supply chain player constructs its best linearforecast of the leadtime demand and uses it to determine the orderquantity via a periodic review myopic order-up-to policy. We demonstratehow a typical supply chain player can create a sequence of partialinformation shocks (PIS) from its full information shocks FIS and sharethese with an adjacent upstream player. We go on to show how such asharing arrangement may be benecial to the upstream player bycharacterizing the player's FIS in such a case. Hence, we study how aplayer can determine its available information under PIS sharing, anduse this information to forecast leadtime demand. We characterize thevalue of FIS sharing for a typical supply chain player. Furthermore, weshow conditions under which a player is able to form and share valuablePIS without (i) revealing its historic demand sequence or (ii) revealingits FIS sequence. We also provide a way of comparing various PIS sharingarrangements with each other and with conventional sharing arrangementsinvolving demand sharing or FIS sharing. We show that demand propagatesthrough a supply chain where any player may share nothing or a sequenceof PIS shocks with an adjacent upstream player as quasi-ARMA in -quasi-ARMA out.

PLUG-IN SELECTION OF THE NUMBER OF FREQUENCIES IN REGRESSION ESTIMATES OF THE MEMORY PARAMETER OF A LONG-MEMORY TIME SERIES

Sun, 28 Jun 1998 22:58:59 GMT

Title: PLUG-IN SELECTION OF THE NUMBER OF FREQUENCIES IN REGRESSION ESTIMATESOF THE MEMORY PARAMETER OF A LONG-MEMORY TIME SERIES

Hurvich, Clifford M.; Deo, Rohit S.

Abstract: We consider the problem of selecting the number of frequencies, m, in alog-periodogram regression estimator of the memory parameter d of aGaussian long-memory time series. It is known that under certainconditions the optimal m, minimizing the mean squared error of thecorresponding estimator of d, is given by m(opt) = Cn4/5, where n is thesample size and C is a constant. In practice, C would be unknown sinceit depends on the properties of the spectral density near zerofrequency. In this paper, we propose an estimator of C based again on alog-periodogram regression and derive its consistency. We also derivean asymptotically valid confidence interval for d when the number offrequencies used in the regression is deterministic and proportional ton4/5. In this case, squared bias cannot be neglected since it is of thesame order as the variance. In a Monte Carlo study, we examine theperformance of the plug-in estimator of d, in which m is obtained byusing the estimator of C in the formula for m(opt) above. We also studythe performance of a bias-corrected version of the plug-in estimator ofd. Comparisons with the choice m = nÃÂ½frequencies, as originally suggested by Geweke and Porter-Hudak.

PARAMETRIC ESTIMATION OF HAZARD FUNCTIONS WITH STOCHASTIC COVARIATE PROCESSES

Tue, 29 Oct 1996 22:58:59 GMT

Title: PARAMETRIC ESTIMATION OF HAZARD FUNCTIONS WITH STOCHASTIC COVARIATE PROCESSES

Berman, Simeon M.; Frydman, Halina

Abstract: Let X(t), t Ã¢ÂÂ¥ 0, be a real or vectorvalued stochastic process and T a random killing-time of the processwhich generally depends on the sample function. In the context ofsurvival analysis, T represents the time to a prescribed event (e.g.system failure, time of disease symptom, etc.) and X(t) is a stochasticcovariate process, observed up to time T. The conditional distributionof T, given X(t), t Ã¢ÂÂ¥ 0, is assumed tobe of a known functional form with an unknown vector parameterÃÂ¸; however, the distributions ofX(Ã¢ÂÂ¢) are not specified. For anarbitrary fixed ÃÂ± > 0 the observable data froma single realization of T and X(Ã¢ÂÂ¢) ismin(T, ÃÂ±), X(t), 0Ã¢ÂÂ¤ tÃ¢ÂÂ¤ min(T, ÃÂ±).For n Ã¢ÂÂ¥ 1 the maximum likelihoodestimator of ÃÂ¸ is based on n independent copies ofthe observable data. It is shown that solutions of the likelihoodequation are consistent and asymptotically normal and efficient underspecified regularity conditions on the hazard function associated withthe conditional distribution of T. The Fisher information matrix isrepresented in terms of the hazard function. The form of the hazardfunction is very general, and is not restricted to the commonlyconsidered cases where it depends onX(Ã¢ÂÂ¢) only through the present pointX(t). Furthermore, the process X(Ã¢ÂÂ¢) isa general, not necessarily Markovian process.

ON THE LOG PERIODOGRAM REGRESSION ESTIMATOR OF THE MEMORY PARAMETER IN LONG MEMORY STOCHASTIC VOLATILITY MODELS

Wed, 29 Oct 1997 22:58:59 GMT

Title: ON THE LOG PERIODOGRAM REGRESSION ESTIMATOR OF THE MEMORY PARAMETER INLONG MEMORY STOCHASTIC VOLATILITY MODELS

Deo, Rohit S.; Hurvich, Clifford M.

Abstract: We consider semiparametric estimation of the memory parameter in a longmemory stochastic volatility model. We study the estimator based on alog periodogram regression as originally proposed by Geweke andPorter-Hudak (1983, Journal of Time Series Analysis 4,221ÃÂ¢ÃÂÃÂ238).Expressions for the asymptotic bias and variance of the estimator areobtained, and the asymptotic distribution is shown to be the same asthat obtained in recent literature for a Gaussian long memory series.The theoretical result does not require omission of a block offrequencies near the origin. We show that this ability to use thelowest frequencies is particularly desirable in the context of the longmemory stochastic volatility model.

ON THE LOG PERIODOGRAM REGRESSION ESTIMATOR OF THE MEMORY PARAMETER IN LONG MEMORY STOCHASTIC VOLATILITY MODELS

Fri, 29 Oct 1999 22:58:59 GMT

Title: ON THE LOG PERIODOGRAM REGRESSION ESTIMATOR OF THE MEMORY PARAMETER INLONG MEMORY STOCHASTIC VOLATILITY MODELS

Deo, Rohit S.; Hurvich, Clifford M.

Abstract: We consider semiparametric estimation of the memory parameter in a longmemory stochastic volatility model. We study the estimator based on alog periodogram regression as originally proposed by Geweke andPorter-Hudak (1983, Journal of Time Series Analysis 4, 221 238).Expressions for the asymptotic bias and variance of the estimator areobtained, and the asymptotic distribution is shown to be the same asthat obtained in recent literature for a Gaussian long memory series.The theoretical result does not require omission of a block offrequencies near the origin. We show that this ability to use the lowestfrequencies is particularly desirable in the context of the long memorystochastic volatility model.

On the Correlation Matrix of the Discrete Fourier Transform and the Fast Solution of Large Toeplitz Systems For Long-Memory Time Series

Tue, 27 Jul 2004 22:58:59 GMT

Title: On the Correlation Matrix of the Discrete Fourier Transform and the FastSolution of Large Toeplitz Systems For Long-Memory Time Series

Chen, Willa; Hurvich, Clifford M.; Lu, Yi

Abstract: For long-memory time series, we show that the Toeplitz system§n(f)x = b can be solved in O(n log5=2 n) operations using awell-known version of the preconditioned conjugate gradient method,where §n(f) is the n£n covariance matrix, f is the spectraldensity and b is a known vector. Solutions of such systems are neededfor optimal linear prediction and interpolation. We establishconnections between this preconditioning method and the frequency domainanalysis of time series. Indeed, the running time of the algorithm isdetermined by rate of increase of the condition number of thecorrelation matrix of the discrete Fourier transform vector, as thesample size tends to 1. We derive an upper bound for this conditionnumber. The bound is of interest in its own right, as it sheds somelight on the widely-used but heuristic approximation that thestandardized DFT coefficients are uncorrelated with equal variances. Wepresent applications of the preconditioning methodology to theforecasting and smoothing of volatility in a long memory stochasticvolatility model, and to the evaluation of the Gaussian likelihoodfunction of a long-memory model.

On testing the adequacy of stable processes under conditional heteroscedasticity

Tue, 16 Oct 2001 22:58:59 GMT

Title: On testing the adequacy of stable processes under conditional heteroscedasticity

Deo, Rohit S.

Abstract: We consider a recently proposed method of estimating the tail index andtesting the goodness-of-fit of dependent stable processes. Through MonteCarlo simulations, we evaluate the ability of the procedure todistinguish between stable and non-stable processes in the presence ofnon-linear dependence and to estimate the tail index of thedistribution. We then apply the test to black market East Europeanexchange rates, whose distributional and tail behaviour has beenanalysed previously in the literature. After adjusting for seasonality,we conclude, unlike the earlier analysis, that a stable process cannotbe rejected as a model for some of the currencies. Estimates of the tailindex for these currencies are also obtained.

Multistep forecasting of long memory series using fractional exponential models

Fri, 29 Oct 1999 22:58:59 GMT

Title: Multistep forecasting of long memory series using fractional exponential models

Hurvich, Clifford M.

Abstract: We develop forecasting methodology for the fractional exponential (FEXP)model. First, we devise algorithms for fast exact computation of thecoefficients in the infinite order autoregressive and moving averagerepresentations of a FEXP process. We also describe an algorithm toaccurately approximate the autocovariances and to simulate realizationsof the process. Next, we present a fast frequency-domain crossvalidation method for selecting the order of the model. This modelselection method is designed to yield the model which provides the bestmultistep forecast for the given lead time, without assuming that theprocess actually obeys a FEXP model. Finally, we use the infinite orderautoregressive coefficients of a fitted FEXP model to constructmultistep forecasts of inflation in the United Kingdom. These forecastsare substantially different than those from a fitted ARFIMA model.

Moving Toward the Rule of Law in the Face of Corruption: Re-examining the Big-Bang

Fri, 29 Oct 2004 22:58:59 GMT

Title: Moving Toward the Rule of Law in the Face of Corruption: Re-examiningthe Big-Bang

Katz, Barbara G.; Owen, Joel

Abstract: We investigate the claim that the establishment of property rights in aneconomy in transition would create its own demand for the enforcement oflaws to protect those rights. Our model contains a government seekingactivities to accomplish certain objectives that depend on publicsupport for the enforcement of the rule of law. It also contains agentswho interpret the activities of the government as signals as to theintent of the government to enforce the rule of law. The agents use thesignals in their choice of whether to support the objectives of thegovernment. With both the government and the agents playing an activerole, we establish conditions under which the activities chosen by thegovernment will maximize its benefits and, at the same time, maximizethe constituency in support of enforcement. These conditions provide abasis for the argument for the implementation of the big-bang policy ineconomies in transition. However, when these conditions do not hold, weshow that in pursuing its own goals, the government reduces support forthe enforcement of the rule of law, which, in our model, leads to anincrease in corruption. Two characteristics play an important role inthese conditions: the initial level of corruption in the economy and thetypes of activities the government chooses to undertake. We present fourexamples to determine the relative importance to our conclusions of eachof these characteristics.

Model Selection in Regression Based on Presmoothing

Sat, 29 Oct 2005 22:58:59 GMT

Title: Model Selection in Regression Based on Presmoothing

Aerts, Marc; Hens, Niel; Simonoff, Jeffrey S.

Abstract: In this paper we investigate the effect of presmoothing on modelselection. Christobal Christobal et al. (1987) showed the beneficialeffect of presmoothing for estimating the parameters in a linearregression model. Here, in a regression setting, we show that smoothingthe response data prior to model selection by Akaike's InformationCriterion can lead to an improved selection procedure. The bootstrap isused to control the magnitude of the random error structure in thesmoothed data. The effect of presmoothing on model selection is shown insimulations. The method is illustrated in a variety of settings,including the selection of the best fractional polynomial in ageneralized linear model.

MODEL SELECTION FOR BROADBAND SEMIPARAMETRIC ESTIMATION OF LONG MEMORY IN TIME SERIES

Fri, 26 Feb 1999 22:58:59 GMT

Title: MODEL SELECTION FOR BROADBAND SEMIPARAMETRIC ESTIMATION OF LONG MEMORYIN TIME SERIES

Hurvich, Clifford M.

Abstract: We study the properties of MallowsÃ¢ÂÂ CLcriterion for selecting a fractional exponential (FEXP) model for aGaussian long-memory time series. The aim is to minimize the meansquared error of a corresponding regression estimator dFEXP of thememory parameter, d. Under conditions which do not require that thedata were actually generated by a FEXP model, it is known that the meansquared error MSE = E[dFEXP Ã¢ÂÂd]ÃÂ² can converge to zero as fast as (log n)/n,where n is the sample size, assuming that the number of parameters growsslowly with n in a deterministic fashion. Here, we suppose that thenumber of parameters in the FEXP model is chosen so as to minimize alocal version of CL, restricted to frequencies in a neighborhood ofzero. We show that, under appropriate conditions, the expected value ofthe local CL is asymptotically equivalent to MSE. A combination oftheoretical and simulation results give guidance as to the choice of thedegree of locality in CL.

Minimax and the Value of Information

Wed, 18 Apr 2012 14:38:02 GMT

Title: Minimax and the Value of Information

Sadler, Evan

Abstract: In his discussion of minimax decision rules, Savage (1954, p. 170)presents an example purporting to show that minimax applied to negativeexpected utility (referred to by Savage as 'negative income') is aninadequate decision criterion for statistics; he suggests theapplication of a minimax regret rule instead. The crux of Savage'sobjection is the possibility that a decision maker would choose toignore even 'extensive' information. More recently, Parmigiani (1992)has suggested that minimax regret suffers from the same flaw. Hedemonstrates the existence of 'relevant' experiments that a minimaxregret agent would never pay a positive cost to observe. On closerinspection, I find that minimax regret is more resilient to thiscritique than would first appear. In particular, there are cases whereno experiment has any value to an agent employing the minimax negativeincome rule, while we may always devise a hypothetical experiment that aminimax regret agent would pay for. The force of Parmigiani's critiqueis further blunted by the observation that 'relevant' experiments existfor which a Bayesian agent would never pay. I conclude by discussing thenotion of pessimism in the context of minimax decision rules.

MAXIMUM LIKELIHOOD ESTIMATION OF HIDDEN MARKOV PROCESSES

Sun, 29 Oct 2000 22:58:59 GMT

Title: MAXIMUM LIKELIHOOD ESTIMATION OF HIDDEN MARKOV PROCESSES

Frydman, Halina; Lakner, Peter

Abstract: We consider the process dYt = ut dt + dWt , where u is a process notnecessarily adapted to F Y (the filtration generated by the process Y)and W is a Brownian motion. We obtain a general representation for thelikelihood ratio of the law of the Y process relative to Brownianmeasure. This representation involves only one basic filter (expectationof u conditional on observed process Y). This generalizes the result ofKailath and Zakai [Ann.Math. Statist. 42 (1971) 130â140] whereit is assumed that the process u is adapted to F Y . In particular, weconsider the model in which u is a functional of Y and of a randomelement X which is independent of the Brownian motion W. For example, Xcould be a diffusion or a Markov chain. This result can be applied tothe estimation of an unknown multidimensional parameter Î¸appearing in the dynamics of the process u based on continuousobservation of Y on the time interval [0,T ]. For a specific hiddendiffusion financial model in which u is an unobserved mean-revertingdiffusion, we give an explicit form for the likelihood function ofÎ¸. For this model we also develop a computationallyexplicit EâM algorithm for the estimation of Î¸. Incontrast to the likelihood ratio, the algorithm involves evaluation of anumber of filtered integrals in addition to the basic filter.

Machine learning for targeted display advertising: Transfer learning in action

Tue, 19 Feb 2013 15:41:19 GMT

Title: Machine learning for targeted display advertising: Transfer learning in action

Perlich, C; Dalessandro, B; Stitelman, O; Raeder, T; Provost, F

Abstract: This paper presents a detailed discussion of problem formulation anddata representation issues in the design, deployment, and operation of amassive-scale machine learning system for targeted display advertising.Notably, the machine learning system itself is deployed and has been incontinual use for years, for thousands of advertising campaigns (incontrast to simply having the models from the system be deployed). Inthis application, acquiring sufficient data for training from the idealsampling distribution is prohibitively expensive. Instead, data aredrawn from surrogate domains and learning tasks, and then transferredto the target task. We present the design of this multistage transferlearning system, highlighting the problem formulation aspects. We thenpresent a detailed experimental evaluation, showing that the differenttransfer stages indeed each add value. We next present productionresults across a variety of advertising clients from a variety ofindustries, illustrating the performance of the system in use. We closethe paper with a collection of lessons learned from the work over half adecade on this complex, deployed, and broadly used machine learning system.

Limit Laws in Trasnaction-Level Asset Price Models

Thu, 19 Jul 2012 13:30:03 GMT

Title: Limit Laws in Trasnaction-Level Asset Price Models

Aue, Alexander; Horvath, Lajos; Hurvich, Clifford; Soulier, Philippe

Abstract: We consider pure-jump transaction-level models for asset prices incontinuous time, driven by point processes. In a bivariate model thatadmits cointegration, we allow for time deformations to account for sucheffects as intraday seasonal patterns in volatility, and non-tradingperiods that may be different for the two assets. We also allow forasymmetries (leverage effects). We obtain the asymptotic distribution ofthe log-price process. For the weak fractional cointegration case, weobtain the asymptotic distribution of the ordinary least-squaresestimator of the cointegrating parameter based on data sampled from anequally-spaced discretization of calendar time, and justify a feasiblemethod of hypothesis testing for the cointegrating parameter based onthe corresponding t-statistic. In the strong fractional cointegrationcase, we obtain the limiting distribution of a continuously-averagedtapered estimator as well as other estimators of the cointegratingparameter, and find that the rate of convergence can be affected byproperties of intertrade durations. In particular, the persistence ofdurations (hence of volatility) can affect the degree of cointegration.We also obtain the rate of convergence of several estimators of thecointegrating parameter in the standard cointegration case. Finally, weconsider the properties of the ordinary least squares estimator of theregression parameter in a spurious regression, i.e., in the absence of cointegration.

Limit Laws in Transaction-Level Asset Price Models

Wed, 27 May 2009 14:30:41 GMT

Title: Limit Laws in Transaction-Level Asset Price Models

Aue, Alexander; Horvath, Lajos; Hurvich, Clifford

Abstract: We consider pure-jump transaction-level models for asset prices incontinuous time, driven by point processes. In a bivariate model thatadmits cointegration, we allow for time deformations to account for sucheffects as intraday seasonal patterns in volatility, and non-tradingperiods that may be different for the two assets. Most assumptions arestated directly on the point process, though we provide sufficientconditions on the corresponding inter-trade durations for theseassumptions to hold. We obtain the asymptotic distribution of thelog-price process. We also obtain the asymptotic distribution of theordinary least-squares estimator of the cointegrat- ing parameter basedon data sampled from an equally-spaced discretization of calendar time,in the case of weak fractional cointegration. Finally, we obtain thelimiting distribution of the ordinary least-squares estimator of theautoregressive parameter in a simplified transaction-level univariatemodel with a unit root.

Forecasting and Information Sharing in Supply Chains Under Quasi-ARMA Demand

Mon, 13 Jul 2009 17:15:07 GMT

Title: Forecasting and Information Sharing in Supply Chains Under Quasi-ARMA Demand

Giloni, Avi; Hurvich, Clifford; Seshadri, Sridhar

Abstract: In this paper, we revisit the problem of demand propagation in amulti-stage supply chain in which the retailer observes ARMA demand. Incontrast to previous work, we show how each player constructs the orderbased upon its best linear forecast of leadtime demand given itsavailable information. In order to characterize how demand propagatesthrough the supply chain we construct a new process which we callquasi-ARMA or QUARMA. QUARMA is a generalization of the ARMA model. Weshow that the typical player observes QUARMA demand and places ordersthat are also QUARMA. Thus, the demand propagation model isQUARMA-in-QUARMA-out. We study the value of information sharing betweenadjacent players in the supply chain. We demonstrate that under certainconditions information sharing can have unbounded bene¯ts. Ouranalysis hence reverses and sharpens several previous results in theliterature involving information sharing and also opens up manyquestions for future research.

Financial Buyers in Takeovers: Focus on Cost Efficiency

Tue, 31 Dec 2002 22:58:59 GMT

Title: Financial Buyers in Takeovers: Focus on Cost Efficiency

Frydman, Halina; Frydman, Roman; Trimbath, Susanne

Abstract: This paper examines whether financial buyers are more likely to initiatetakeovers of inefficient firms. We show that they indeed are and thusconclude that takeovers by financial buyers play a potentiallybeneficial role in the allocation of corporate assets in the U.S.economy. Our analysis of determinants of takeovers initiated byfinancial buyers uses an application of the methodology developed inTrimbath, Frydman and Frydman (2001). In order to illustrate efficiencyenhancements introduced by financial buyers, we select Forstmann andLittle’s acquisition of General Instrument for a brief case study.We show that their aggressive programs of cost management substantiallyimproved the efficiency of General Instrument. Moreover, it allowedGeneral Instrument to expand research and development to become theglobal leader in high definition television.

Evaluating Density Forecasts with Applications to Financial Risk Management

Thu, 29 Oct 1998 22:58:59 GMT

Title: Evaluating Density Forecasts with Applications to Financial Risk Management

Diebold, Francis X.; Gunther, Todd A.; Tay, Anthony S.

Abstract: Density forecasting is increasingly more important and commonplace, forexample in financial risk management, yet little attention has beengiven to the evaluation of density forecasts. We develop a simple andoperational framework for density forecast evaluation. We illustrate theframework with a detailed application to density forecasting of assetreturns in environments with time-varying volatility. Finally, wediscuss several extensions.

Estimation of Long Memory in Volatility

Fri, 29 Oct 1999 22:58:59 GMT

Title: Estimation of Long Memory in Volatility

Deo, Rohit S.; Hurvich, C. M.

Abstract: We discuss some of the issues pertaining to modelling and estimatinglong memory in volatility. The main focus is on semi parametricestimation of the memory parameter in the long memory stochasticvolatility model. We present the asymptotic properties of the logperiodogram regression estimator of the memory parameter in this model.A modest simulation study of the estimator is also presented to studyits behaviour when the volatility possesses only short memory. Weconclude with a discussion of the appropriate choice of transformationof returns to measure persistence in volatility.

Estimation of long memory in the presence of a smooth nonparametric trend

Wed, 24 Jul 2002 22:58:59 GMT

Title: Estimation of long memory in the presence of a smooth nonparametric trend

Hurvich, Clifford; Lang, Gabriel; Soulier, Philippe

Abstract: We consider semi parametric estimation of the long-memory parameter of astationary process in the presence of an additive nonparametric meanfunction. We use a semi parametric Whittle type estimator, applied tothe tapered, differenced series. Since the mean function is notnecessarily a polynomial of finite order, no amount of differencing willcompletely remove the mean. We establish a central limit theorem for theestimator of the memory parameter, assuming that a slowly increasingnumber of low frequencies are trimmed from the estimator's objectivefunction. We find in simulations that tapering and trimming areessential for the good performance of the estimator in practice.

Estimation in the Mixture of Markov Chains

Wed, 08 Jan 2003 22:58:59 GMT

Title: Estimation in the Mixture of Markov Chains

Frydman, Halina

Abstract: This paper considers a new mixture of time homogeneous finite Markovchains where the mixing is on the rate of movement and develops the EMalgorithm for the maximum likelihood estimation of the parameters of themixture. A continuous and discrete time versions of the mixture aredefined and their estimation is considered separately. The simulationstudy is carried out for the continuous time mixture. To simplify theexposition the results are derived for a mixture of two Markov chains,but can be easily extended to a mixture of any finite number of Markovchains. The class of mixture models proposed in this paper provides aframework for modeling population heterogeneity with respect to the rateof movement. The proposed mixture generalizes the mover-stayer model,which has been widely employed in applications.

Estimation in the continuous time mover-stayer model with an application to bond ratings migration

Wed, 18 Dec 2002 22:58:59 GMT

Title: Estimation in the continuous time mover-stayer model with an applicationto bond ratings migration

Frydman, Halina; Kadam, Ashay

Abstract: The usual tool for modeling bond ratings migration is a discrete,timehomogeneuous Markov chain. Such model assumes that all bonds arehomogeneous with respect to their movement behavior among ratingcategories and that the movement behavior does not change over time.However, among recognized sources of heterogeneity in ratings migrationis age of a bond (time elapsed since issuance). It has been observedthat young bonds have a lower propensity to change ratings, and thus todefault, than more seasoned bonds. The aimof this paper is to introducea continuous, time-nonhomogeneuous model for bond ratings migration,which also incorporates a simple form of population heterogeneity. Thespecific form of heterogeneity postulated by the proposed model appearsto be suitable for modeling the effect of age of a bond on itspropensity to change ratings. This model, called a mover-stayer model,is an extension of a time-nonhomogeneuous Markov chain. This paperderives the maximum likelihood estimators for the parameters of acontinuous time mover-stayer model based on a sample of independentcontinuously monitored histories of the process, and develops thelikelihood ratio test for discriminating between the Markov chain andthe mover-stayer model. The methods are illustrated using a sample ofrating histories of young corporate issuers. For this sample, thelikelihood ratio test rejects a Markov chain in favor of a mover-stayermodel. For young bonds with lowest rating the default probabilitiespredicted by the mover-stayer model are substantially lower than thosepredicted by the Markov chain.

Estimating long memory in volatility

Tue, 29 Jan 2002 22:58:59 GMT

Title: Estimating long memory in volatility

Hurvich, Clifford M.; Moulines, Eric; Soulier, Philippe

Abstract: We consider semiparametric estimation of the memory parameter in a modelwhich includes as special cases both the long-memory stochasticvolatility (LMSV) and fractionally integrated exponential GARCH(FIEGARCH) models. Under our general model the logarithms of the squaredreturns can be decomposed into the sum of a long-memory signal and awhite noise. We consider periodogram-based estimators which explicitlyaccount for the noise term in a local Whittle criterion function. Weallow the optional inclusion of an additional term to allow for acorrelation between the signal and noise processes, as would occur inthe FIEGARCH model. We also allow for potential nonstationarity involatility, by allowing the signal process to have a memory parameter d1=2. We show that the local Whittle estimator is consistent for d 2 (0;1). We also show that a modi ed version of the local Whittle estimatoris asymptotically normal for d 2 (0; 3=4), and essentially recovers theoptimal semiparametric rate of convergence for this problem. Inparticular if the spectral density of the short memory component of thesignal is suficiently smooth, a convergence rate of n2=5-δ for d 2(0; 3=4) can be attained, where n is the sample size and δ > 0is arbitrarily small. This represents a strong improvement over theperformance of existing semiparametric estimators of persistence involatility. We also prove that the standard Gaussian semiparametricestimator is asymptotically normal if d = 0. This yields a test forlong memory in volatility.

Estimating fractional cointegration in the presence of polynomial trends

Wed, 05 Feb 2003 22:58:59 GMT

Title: Estimating fractional cointegration in the presence of polynomial trends

Chen, Willa W.; Hurvich, Clifford M.

Abstract: We propose and derive the asymptotic distribution of a taperednarrow-band least squares estimator (NBLSE) of the cointegrationparameter ÃÂÃÂ² in the framework offractional cointegration. This tapered estimator is invariant todeterministic polynomial trends. In particular, we allow for arbitrarylinear time trends that often occur in practice. Our simulations showthat, in the case of no deterministic trends, the estimator is superiorto ordinary least squares (OLS) and the nontapered NBLSE proposed byP.M. Robinson when the levels have a unit root and the cointegratingrelationship between the series is weak. In terms of rate ofconvergence, our estimator converges faster under certain circumstances,and never slower, than either OLS or the nontapered NBLSE. In a dataanalysis of interest rates, we find stronger evidence of cointegrationif the tapered NBLSE is used for the cointegration parameter than if OLSis used.

Estimating Fractional Cointegration in the Presence of Polynomial Trends

Mon, 07 Oct 2002 22:58:59 GMT

Title: Estimating Fractional Cointegration in the Presence of Polynomial Trends

Chen, Willa W.; Hurvich, Clifford M.

Abstract: We propose and derive the asymptotic distribution of a taperednarrow-band least squares estimator (NBLSE) of the cointegrationparameter Î² in the framework of fractional cointegration.This tapered estimator is invariant to deterministic polynomial trends.In particular, we allow for arbitrary linear time trends that oftenoccur in practice. Our simulations show that, in the case of nodeterministic trends, the estimator is superior to ordinary leastsquares (OLS) and the nontapered NBLSE proposed by P.M. Robinson whenthe levels have a unit root and the cointegrating relationship betweenthe series is weak. In terms of rate of convergence, our estimatorconverges faster under certain circumstances, and never slower, thaneither OLS or the nontapered NBLSE. In a data analysis of interestrates, we find stronger evidence of cointegration if the tapered NBLSEis used for the cointegration parameter than if OLS is used.

Efficiency and Consistency for Regularization Parameter Selection in Penalized Regression: Asymptotics and Finite-Sample Corrections

Thu, 17 Nov 2011 16:59:41 GMT

Title: Efficiency and Consistency for Regularization Parameter Selection inPenalized Regression: Asymptotics and Finite-Sample Corrections

Flynn, Cheryl J.; Hurvich, Clifford M.; Simonoff, Jeffrey S.

Abstract: This paper studies the asymptotic and nite-sample performance ofpenalized regression methods when different selectors of theregularization parameter are used under the assumption that the truemodel is, or is not, included among the candidate model. In the lattersetting, we relax assumptions in the existing theory to show thatseveral classical information criteria are asymptotically efficientselectors of the regularization parameter. In both settings, we assessthe nite-sample performance of these as well as other common selectorsand demonstrate that their performance can suffer due to sensitivity tothe number of variables that are included in the full model. Asalternatives, we propose two corrected information criteria which areshown to outperform the existing procedures while still maintaining thedesired asymptotic properties. In the non-true model world, we relaxthe assumption made in the literature that the true error variance isknown or that a consistent estimator is available to prove that Akaike'sinformation criterion (AIC), Cp and Generalized cross-validation (GCV)themselves are asymptotically efficient selectors of the regularizationparameter and we study their performance in nite samples. In classicalregression, AIC tends to select overly complex models when the dimensionof the maximum candidate model is large relative to the sample size.Simulation studies suggest that AIC suffers from the same shortcomingswhen used in penalized regression. We therefore propose the use of theclassical AICc as an alternative. In the true model world, a similarinvestigation into the nite sample properties of BIC reveals analogousoverfitting tendencies and leads us to further propose the use of acorrected BIC (BICc). In their respective settings (whether the truemodel is, or is not, among the candidate models), BICc and AICc have thedesired asymptotic properties and we use simulations to assess theirperformance, as well as that of other selectors, in nite samples forpenalized regressions fit using the Smoothly clipped absolute deviation(SCAD) and Least absolute shrinkage and selection operator (Lasso)penalty functions. We nd that AICc and 10-fold cross-validationoutperform the other selectors in terms of squared error loss, and BICcavoids the tendency of BIC to select overly complex models when thedimension of the maximum candidate model is large relative to the sample size.

Economic Trade: a Solution to the Production Frontier of Two Economies in Trade

Mon, 28 Nov 2005 22:58:59 GMT

Title: Economic Trade: a Solution to the Production Frontier of Two Economiesin Trade

Tashjian, Richard

Abstract: It is often possible to intuit the function describing the productionfrontier of two economies in trade. For example, when the two individualfrontiers are linear, the aggregate frontier can be rendered by anintuitive sense of the two derivatives, as will be shown below. But whenthe two functions are complex, no amount of judgment will suffice tosketch the aggregate production frontier. To solve this conundrum, wewill employ polar coordinates and Lagrange Multipliers to yield ananalytical function describing the aggregate production frontier.

Drift in Transcation-Level Asset Price Models

Tue, 20 Nov 2012 18:02:03 GMT

Title: Drift in Transcation-Level Asset Price Models

Cao, Wen; Hurvich, Clifford; Soulier, Philippe

Abstract: We study the effect of drift in pure-jump transaction-level models forasset prices in continuous time, driven by point processes. The drift isassumed to arise from a nonzero mean in the efficient shock series. Itfollows that the drift is proportional to the driving point processitself, i.e. the cumulative number of transactions. This link reveals amechanism by which properties of intertrade durations (such as heavytails and long memory) can have a strong impact on properties of averagereturns, thereby potentially making it extremely difficult to determinegrowth rates. We focus on a basic univariate model for log price,coupled with general assumptions on durations that are satisfied byseveral existing flexible models, allowing for both long memory andheavy tails in durations. Under our pure-jump model, we obtain thelimiting distribution for the suitably normalized log price. Thislimiting distribution need not be Gaussian, and may have either finitevariance or infinite variance. We show that the drift can affect notonly the limiting distribution for the normalized log price, but alsothe rate in the corresponding normalization. Therefore, the drift (orequivalently, the properties of durations) affects the rate ofconvergence of estimators of the growth rate, and can invalidatestandard hypothesis tests for that growth rate. Our analysis also shedssome new light on two longstanding debates as to whether stock returnshave long memory or infinite variance.

Discrete Quantile Estimation

Sun, 28 Jan 2007 22:58:59 GMT

Title: Discrete Quantile Estimation

Frydman, Halina; Simon, Gary

Abstract: We consider estimation of a quantile from a discrete distribution. Thisgives rise to three new ideas, the confidence set for such a quantile,the notion that the associated confidence level can be increased afterthe data are collected, and that it is legitimate to strive to obtain asingleton confidence set. We develop properties of the sample quantilenoting that the behavior for discrete populations is very different fromthe behavior for continuous populations. We illustrate the results withsimulations and examples.

Cost Inefficiency, Size of Firms and Takeovers

Sun, 29 Oct 2000 22:58:59 GMT

Title: Cost Inefficiency, Size of Firms and Takeovers

Trimbath, Susanne; Frydman, Halina; Frydman, Roman

Abstract: This study, using the Cox proportional hazards model, finds that therisk of takeover rises with cost inefficiency. It also finds that a firmfaces a significantly higher risk of takeover if its cost performancelags behind its industry benchmark. Moreover, these findings appear tobe remarkably stable over the nearly two decades spanned by the sample.The effect of the variables used to measure the risk-size relationship,however, indicates temporal changes. Lastly, the study presents evidencefrom fixed-effects models of ex post cost efficiency improvements thatsupport the hypothesis that takeover targets are selected based on thepotential for improvement.

Conditions for the Propagation of Memory Parameter from Durations to Counts and Realized Volatility

Tue, 15 May 2007 22:58:59 GMT

Title: Conditions for the Propagation of Memory Parameter from Durations toCounts and Realized Volatility

Deo, Rohit; Soulier, Philippe; Hurvich, Clifford M.; Soulier, Philippe; Wang, Yi

Abstract: We establish sufficient conditions on durations that are stationary withfinite variance and memory parameter d 2 [0; 1=2) to ensure that thecorresponding counting process N(t) satisfies VarN(t) » Ct2d+1 (C> 0) as t ! 1, with the same memory parameter d 2 [0; 1=2) that wasassumed for the durations. Thus, these conditions ensure that the memoryparameter in durations propagates to the same memory parameter in thecounts. We then show that any Autoregressive Conditional DurationACD(1,1) model with a sufficient number of finite moments yields shortmemory in counts, while any Long Memory Stochastic Duration model withand all finite moments yields long memory in counts, with the same d.Next, we present a result implying that the only way for a series ofcounts aggregated over a long time period to have nontrivialautocorrelation is for the counts to have long memory. In other words,aggregation ultimately destroys all autocorrelation in counts, if andonly if the counts have short memory. Finally, under assumptions on thepure-jump price process, we show that the memory parameter in durationspropagates all the way to the realized volatility, under bothcalendar-time sampling and transaction-time sampling.

Computationally Efficient Gaussian Maximum Likelihood Methods for Vector ARFIMA Models

Mon, 13 Oct 2008 21:28:35 GMT

Title: Computationally Efficient Gaussian Maximum Likelihood Methods for VectorARFIMA Models

Sela, Rebecca J.; Hurvich, Clifford M.

Abstract: In this paper, we discuss two distinct multivariate time series modelsthat extend the univariate ARFIMA model. We describe algorithms forcomputing the covariances of each model, for computing the quadraticform and approximating the determinant for maximum likelihoodestimation, and for simulating from each model. We compare the speed andaccuracy of each algorithm to existing methods and measure theperformance of the maximum likelihood estimator compared to existingmethods. We also fit models to data on unemployment and inflation in theUnited States, to data on goods and services inflation in the UnitedStates, and to data about precipitation in the Great Lakes.

Collaborating on Multi-party Information Systems Development Projects: A Collective Reflection-in-Action View

Wed, 29 Oct 2003 22:58:59 GMT

Title: Collaborating on Multi-party Information Systems Development Projects: ACollective Reflection-in-Action View

Levina, Natalia

Abstract: Growth of business-to-consumer (B2C) applications such as electronicstorefronts, catalogues, and customer support websites has drawn a greatnumber of diverse stakeholders into the IS Development (ISD) practice.Marketing, strategy, and graphic design specialists have joined avariety of technical professionals and business stakeholders indeveloping B2C applications. Oftentimes, these professionals work fordifferent organizations with different histories, cultures, and rewardstructures. A longitudinal qualitative field study of a B2C applicationdevelopment project was undertaken in order to build an in-depthunderstanding of the collaborative practices of diverse professionals inISD projects. The paper proposes that the multi-party collaborativepractice can be understood as a âcollectivereflection-in-actionâ cycle through which an IS design emergesas a result of agents producing, sharing, and reflecting upon materialobjects. Agents from diverse backgrounds exert different influences overemergent designs depending on their organization, profession, andproject involvement-based power relations. These power relations shapewhether collaborators âadd toâ âignore,âor âchallengeâ the work produced by others. In turn,agentsâ actions either reinforce or transform existing powerrelations depending on who gets to claim credit for and ownership of theemergent design. Implications for the study of boundary objects, teamdiversity, organizational learning, and contemporary ISD are drawn.

Cointegration and Long-Horizon Forecasting

Mon, 06 Oct 1997 22:58:59 GMT

Title: Cointegration and Long-Horizon Forecasting

Christoffersen, Peter F.; Diebold, Francis X.

BROADBAND SEMIPARAMETRIC ESTIMATION OF THE MEMORY PARAMETER OF A LONG-MEMORY TIME SERIES USING FRACTIONAL EXPONENTIAL MODELS

Mon, 28 Sep 1998 22:58:59 GMT

Title: BROADBAND SEMIPARAMETRIC ESTIMATION OF THE MEMORY PARAMETER OF ALONG-MEMORY TIME SERIES USING FRACTIONAL EXPONENTIAL MODELS

Hurvich, Clifford M.; Brodsky, Julia

Abstract: We consider a fractional exponential, or FEXP estimator of the memoryparameter of a stationary Gaussian long-memory time series. Theestimator is constructed by fitting a FEXP model of slowly increasingdimension to the log periodogram at all Fourier frequencies by ordinaryleast squares, and retaining the corresponding estimated memoryparameter. We do not assume that the data were necessarily generated bya FEXP model, or by any other finite-parameter model. We do, however,impose a global differentiability assumption on the spectral densityexcept at the origin. Because of this, and its use of all Fourierfrequencies, we refer to the FEXP estimator as a broadbandsemiparametric estimator. We demonstrate the consistency of the FEXPestimator, and obtain expressions for its asymptotic bias and variance.It the true spectral density is sufficiently smooth, the FEXP estimatorcan strongly outperform existing semiparametric estimators, such as theGeweke-Porter-Hudak (GPH) and Gaussian semiparametric estimators (GSE),attaining an asymptotic mean squared error proportional to (log n)/n,where n is the sample size. In a simulation study, we demonstrate themerits of using a finite-sample correction to the asymptotic variance,and we also explore the possibility of automatically selecting thedimension of the exponential model usingMallowsÃ¢ÂÂ CL criterion.

Bootstrapping Multivariate Spectra

Fri, 22 Aug 1997 22:58:59 GMT

Title: Bootstrapping Multivariate Spectra

Berkowitz, Jeremy; Diebold, Francis X.

Abstract: We generalize the Franke-HÃÂ¤rdle (1992) spectraldensity bootstrap to the multivariate case. The extension is non-trivialand facilitates use of the Franke-HÃÂ¤rdlebootstrap in frequency-domain econometric work, which often centers oncross-variable dynamic interactions. We document thebootstrapÃ¢ÂÂs good finite-sample performancein a small Monte Carlo experiment, and we conclude by highlighting keydirections for future research.

Bias Reduction and Likelihood Based Almost-Exactly Sized Hypothesis Testing in Predestricted Likelihoodictive Regressions using the R

Mon, 24 Aug 2009 19:54:56 GMT

Title: Bias Reduction and Likelihood Based Almost-Exactly Sized HypothesisTesting in Predestricted Likelihoodictive Regressions using the R

Chen, Willa W.; Deo, Rohit S.

Abstract: Difficulties with inference in predictive regressions are generallyattributed to strong persistence in the predictor series. We show thatthe major source of the problem is actually the nuisance interceptparameter and propose basing inference on the RestrictedLikelihood,which is free of such nuisance location parameters and alsopossesses small curvature, making it suitable for inference. The bias ofthe Restricted Maximum Likelihood (REML) estimates is shown to beapproximately 50% less than that of the OLS estimates near the unitroot, without loss of efficiency. The error in the chi-squareapproximation to the distribution of the REML based Likelihood RatioTest (RLRT) for no predictability is shown to be 3/4 − ρ2n−1 (G3 (·) − G1 (·)) + O n−2 ,where |ρ| < 1 is the correlation of the innovation series and Gs(·) is the c.d.f. of a χ2s random variable. This very smallerror, free of the AR parameter, suggests that the RLRT forpredictability has very good size properties even when the regressor hasstrong persistence. The Bartlett corrected RLRT achieves an On−2 error. Power under local alternatives is obtained andextensions to more general univariate regressors and vector AR(1)regressors, where OLS may no longer be asymptotically efficient, areprovided. In simulations the RLRT maintains size well, is robust tonon-normal errors and has uniformly higher power than theJansson-Moreira test with gains that can be substantial. The Campbell-Yogo Bonferroni Q test is found to have size distortions and can besignificantly oversized.

Asymptotics for Duration-Driven Long Range Dependent Processes

Thu, 28 Aug 2003 22:58:59 GMT

Title: Asymptotics for Duration-Driven Long Range Dependent Processes

Hsieh, Mengchen; Hurvich, Clifford M.; Souliery, Philippe

Abstract: We consider processes with second order long range dependence resultingfrom heavy tailed durations. We refer to this phenomenon asduration-driven long range dependence (DDLRD), as opposed to the morewidely studied linear long range dependence based on fractional dierencing of an iid process. We consider in detail two speci c processeshav- ing DDLRD, originally presented in Taqqu and Levy (1986), and Parke(1999). For these processes, we obtain the limiting distribution ofsuitably standardized discrete Fourier trans forms (DFTs) and sampleautocovariances. At low frequencies, the standardized DFTs converge to astable law, as do the standardized autocovariances at xed lags. Finitecollections of standardized autocovariances at a xed set of lagsconverge to a degenerate distribution. The standardized DFTs at highfrequencies converge to a Gaussian law. Our asymptotic results arestrikingly similar for the two DDLRD processes studied. We calibrate ourasymptotic results with a simulation study which also investigates theproperties of the semiparametric log periodogram regression estimator ofthe memory parameter.

Assessing the Difference Between Shock Sharing and Demand Sharing in Supply Chains

Wed, 03 Oct 2012 13:53:19 GMT

Title: Assessing the Difference Between Shock Sharing and Demand Sharing inSupply Chains

Kovtun, Vladimir; Giloni, Avi; Hurvich, Clifford

Abstract: We consider the problem of assessing value of demand sharing in amulti-stage supply chain in which the retailer observes stationaryautoregressive moving average demand with Gaussian white noise (shocks).Similar to previous research, we assume each supply chain playerconstructs its best linear forecast of the leadtime demand and uses itto determine the order quantity via a periodic review myopic order-up-topolicy. We demonstrate how a typical supply chain player can determinethe extent of its available information under demand sharing by studyingthe properties of the moving average polynomials of adjacent supplychain players. Hence, we study how a player can determine its availableinformation under demand sharing, and use this information to forecastleadtime demand. We characterize the value of demand sharing for atypical supply chain player. Furthermore, we show conditions under which(i) it is equivalent to no sharing, (ii) it is equivalent to fullinformation shock sharing, and (iii) it is intermediate in value to thetwo previously described arrangements. We then show that demandpropagates through a supply chain where any player may share nothing,its demand, or its full-information shocks with an adjacent upstreamplayer as quasi-ARMA in - quasi-ARMA out. We also provide a convenientform for the propagation of demand in a supply chain that will lenditself to future research applications.

APPROXIMATING SEPARABLE NONLINEAR FUNCTIONS VIA MIXED ZERO-ONE PROGRAMS

Mon, 28 Sep 1998 22:58:59 GMT

Title: APPROXIMATING SEPARABLE NONLINEAR FUNCTIONS VIA MIXED ZERO-ONE PROGRAMS

Padberg, M.

Abstract: We discuss two models from the literature that have been developed toformulate piecewise linear approximation of separable nonlinearfunctions by way of mixed-integer programs. We show that the mostcommonly proposed method is computationally inferior to a lesser knowntechnique by comparing analytically the linear programming relaxationsof the two formulations. A third way of formulating the problem, thatshares the advantages of the better of the two known methods, is also proposed.

An Investigation of Missing Data Methods for Classification Trees

Mon, 13 Oct 2008 21:33:40 GMT

Title: An Investigation of Missing Data Methods for Classification Trees

Ding, Yufeng; Simonoff, Jeffrey S.

Abstract: There are many different missing data methods used by classificationtree algorithms, but few studies have been done comparing theirappropriateness and performance. This paper provides both analytic andMonte Carlo evidence regarding the effectiveness of six popular missingdata methods for classification trees. We show that in the context ofclassification trees, the relationship between the missingness and thedependent variable, rather than the standard missingness classificationapproach of Little and Rubin (2002) (missing completely at random(MCAR), missing at random (MAR) and not missing at random (NMAR)), isthe most helpful criterion to distinguish different missing datamethods. We make recommendations as to the best method to use in varioussituations. The paper concludes with discussion of a real data setrelated to predicting bankruptcy of a firm.

An Investigation of Missing Data Methods for Classiffcation Trees

Sat, 02 Dec 2006 22:58:59 GMT

Title: An Investigation of Missing Data Methods for Classiffcation Trees

Ding, Yufeng; Simonoff, Jeffrey S.

Abstract: There are many different missing data methods used by classificationtree algorithms, but few studies have been done comparing theirappropriateness and performance. This paper provides both analytic andMonte Carlo evidence regarding the effectiveness of six popular missingdata methods for classification trees. We show that in the context ofclassification trees, the relationship between the missingness and thedependent variable, rather than the standard missingness classificationapproach of Little and Rubin (2002) (missing completely at random(MCAR), missing at random (MAR) and not missing at random (NMAR)), isthe most helpful criterion to distinguish different missing datamethods. We make recommendations as to the best method to use in varioussituations. The paper concludes with discussion of a real data setrelated to predicting bankruptcy of a firm.

An Empirical Study of Factors Relating to the Success of Broadway Shows

Fri, 29 Oct 1999 22:58:59 GMT

Title: An Empirical Study of Factors Relating to the Success of Broadway Shows

Simonoff, Jeffrey S.; Ma, Lan

Abstract: This article uses the Cox proportional hazards model to analyze recentBroadway show data to investigate the factors that relate to thelongevity of shows. The type of show, whether a show is a revival, andfirst-week attendance for the show are predictive for longevity.Favorable critic reviews in the New York Daily News are related togreater success, but reviews in the New York Times are not. Winningmajor Tony Awards is associated with a longer run for a show, but beingnominated for Tonys and then losing is associated with a shorterpostaward run.

AN EFFICIENT TAPER FOR POTENTIALLY OVERDIFFERENCED LONG-MEMORY TIME SERIES

Mon, 29 Dec 1997 22:58:59 GMT

Title: AN EFFICIENT TAPER FOR POTENTIALLY OVERDIFFERENCED LONG-MEMORY TIME SERIES

Hurvich, Clifford M.; Chen, Willa W.

Abstract: We propose a new complex-valued taper and derive the properties of atapered Gaussian semiparametric estimator of the long-memory parameter dÃÂÃÂ (-0.5, 1.5). The estimator and itsaccompanying theory can be applied to generalized unit root testing. Inthe proposed method, the data are differenced once before the taper isapplied. This guarantees that the tapered estimator is invariant withrespect to deterministic linear trends in the original series. Anydetrimental leakage effects due to the potential noninvertibility of thedifferenced series are strongly mitigated by the taper. The proposedestimator is shown to be more efficient than existing invariant taperedestimators. Invariance to kth order polynomial trends can be attainedby differencing the data k times and then applying a stronger taper,which is given by the kth power of the proposed taper. We show thatthis new family of tapers enjoys strong efficiency gains over comparableexisting tapers. Analysis of both simulated and actual data highlightspotential advantages of the tapered estimator of d compared with thenontapered estimator.

Alternative Methods of Linear Regression

Sat, 28 Apr 2001 22:58:59 GMT

Title: Alternative Methods of Linear Regression

Giloni, A.; Padberg, M.

Abstract: This paper is a survey on traditional linear regression techniques usingthe lÃÂ±-, l2-, andlÃ¢ÂÂ-norm. We derive the characterization ofthe respective regression estimates (including optimality and uniquenesscriteria), as well as discuss some of their statistical properties.

ALMOST PERFECT MATRICES AND GRAPHS

Mon, 28 Sep 1998 22:58:59 GMT

Title: ALMOST PERFECT MATRICES AND GRAPHS

Padberg, M.

Abstract: We introduce the notions of w-projection and k-projection that mapalmost integral polytopes associated with almost perfect graphs G with nnodes from Rn into Rn-w where w is the maximum clique size in G. We showthat C. Berge's strong perfect graph conjecture is correct if and onlyif the projection (of either kind) of such polytopes is again almostintegral in Rn-w. Several important properties of w-projections andk-projections are established. We prove that the strong perfect graphconjecture is wrong if an w-projection and a related k-projection of analmost integral polytope with 2 Ã¢ÂÂ¤ wÃ¢ÂÂ¤ (n - 1)/2 produce differentpolytopes in Rn-w.

Activism's Impact on Diversified Investors and the Market

Mon, 28 Jan 2013 15:21:28 GMT

Title: Activism's Impact on Diversified Investors and the Market

Katz, Barbara; Owen, Joel

Abstract: We model activism as it affects the future distribution of prices in aportfolio context with risk-averse expected utility of end-of-periodwealth maximizing investors. We characterize activism as affecting themean, the variance, and/or the covariance of the target firm’sprice with the prices of the other firms. This characterization allowsus to investigate conditions under which the activist would choose tobecome an activist and, subsequently, to derive the sequence ofequilibria that begins with the surreptitious acquisition of shares bythe activist and ends at the moment of the activist’s divestitureof these shares. We investigate the impact of activism not only on thetarget firm’s price over time and the activist’s profit, butalso on the redistribution of portfolio holdings of all marketparticipants that this activism induces. We propose a method to evaluateactivism and show that, while activism may augment the share price ofthe target firm, there exist conditions under which activism would notnecessarily increase the value of the market. Furthermore, we show thatthe pro.t of the activist is at the expense of the group of otherinvestors. We compare our results to recent empirical findings.

A Small Sample Study of Goodness-of-fit Tests for Time Series Models

Fri, 29 Oct 1999 22:58:59 GMT

Title: A Small Sample Study of Goodness-of-fit Tests for Time Series Models

Chen, Willa W.; Deo, Rohit S.

Abstract: We study the small sample behaviour of two goodness-of-fit tests fortime series models which have been proposed recently in the literature.Both tests are generalizations of the popular Box- Ljung-Pierceportmanteau test, one in the time domain and the other in the frequencydomain. The tests are found to be oversized under the null of whitenoise but undersized under other null hypotheses. The cause for thiseffect is investigated and a finite sample correction proposed whichameliorates this effect. It is found that the corrected versions of thetests have markedly better size properties. The correction is also foundto result in an overall increase in power which can be significant incertain alternatives. Furthermore, the corrected tests also haveuniformly better power than the Box-Ljung-Pierce portmanteau test,unlike the uncorrected versions.

A Pure-Jump Transaction-Level Price Model Yielding Cointegration, Leverage, and Nonsynchronous Trading Effects

Sun, 03 Dec 2006 22:58:59 GMT

Title: A Pure-Jump Transaction-Level Price Model Yielding Cointegration,Leverage, and Nonsynchronous Trading Effects

Hurvich, Clifford M.; Wang, Yi

Abstract: We propose a new transaction-level bivariate log-price model, whichyields fractional or standard cointegration. Most existing models forcointegration require the choice of a fixed sampling interval ¢t.By contrast, our proposed model is constructed at the transaction level,thus determining the properties of returns at all sampling frequencies.The two ingredients of our model are a Long Memory Stochastic Durationprocess for the waiting times f¿kg between trades, and a pair ofstationary noise processes (fekg and f´kg) which determine thejump sizes in the pure-jump log-price process. The fekg, assumed to bei:i:d: Gaussian, produce a Martingale component in log prices. We assumethat the microstructure noise f´kg obeys a certain model withmemory parameter d´ 2 (¡1=2; 0) (fractional cointegrationcase) or d´ = ¡1 (standard cointegration case). Ourlog-price model includes feedback between the disturbances of the twolog-price series. This feedback yields cointegration, in that thereexists a linear combination of the two series that reduces the memoryparameter from 1 to 1 + d´ 2 (0:5; 1) [ f0g. Returns at samplinginterval ¢t are asymptotically uncorrelated at any fixed lag as¢t increases. We prove that the cointegrating parameter can beconsistently estimated by the ordinary least-squares estimator, andobtain a lower bound on the rate of convergence. We proposetransaction-level method-of-moments estimators of several of the otherparameters in our model. We present a data analysis, which providesevidence of fractional cointegration. We then consider special cases andgeneralizations of our model, mostly in simulation studies, to arguethat the suitably-modified model is able to capture a variety ofadditional properties and stylized facts, including leverage, portfolioreturn autocorrelation due to nonsynchronous trading, Granger causality,and volatility feedback. The ability of the model to capture theseeffects stems in most cases from the fact that the model treats the(stochastic) intertrade durations in a fully endogenous way.

A Pure-Jump Transaction-Level Price Model Yielding Cointegration, Leverage, and Nonsynchronous Trading Effects

Wed, 07 Jan 2009 19:25:54 GMT

Title: A Pure-Jump Transaction-Level Price Model Yielding Cointegration,Leverage, and Nonsynchronous Trading Effects

Hurvich, Clifford; Wang, Yi

Abstract: We propose a new transaction-level bivariate log-price model, whichyields fractional or standard cointegration. The model provides a linkbetween market microstructure and lower-frequency observations. The twoingredients of our model are a Long Memory Stochastic Duration processfor the waiting times between trades, and a pair of stationary noiseprocesses which determine the jump sizes in the pure-jump log-priceprocess. Our model includes feedback between the disturbances of the twolog-price series at the transaction level, which induces standard orfractional cointegration for any fixed sampling interval. We prove thatthe cointegrating parameter can be consistently estimated by theordinary least-squares estimator, and obtain a lower bound on the rateof convergence. We propose transaction-level method-of-momentsestimators of the other parameters in our model and discuss theconsistency of these estimators. We then use simulations to argue thatsuitably-modified versions of our model are able to capture a variety ofadditional properties and stylized facts, including leverage, andportfolio return autocorrelation due to nonsynchronous trading. Theability of the model to capture these effects stems in most cases fromthe fact that the model treats the (stochastic) intertrade durations ina fully endogenous way.

A Mathematical Programming Approach for Improving the Robustness of LAD Regression

Thu, 22 Jul 2004 22:58:59 GMT

Title: A Mathematical Programming Approach for Improving the Robustness of LAD Regression

Giloni, Avi; Sengupta, Bhaskar; Simonoff, Jeffrey

Abstract: This paper discusses a novel application of mathematical programmingtechniques to a regression problem. While least squares regressiontechniques have been used for a long time, it is known that theirrobustness properties are not desirable. Specifically, the estimatorsare known to be too sensitive to data contamination. In this paper weexamine regressions based on Least-sum of Absolute Deviations (LAD) andshow that the robustness of the estimator can be improved significantlythrough a judicious choice of weights. The problem of finding optimumweights is formulated as a nonlinear mixed integer program, which is toodifficult to solve exactly in general. We demonstrate that our problemis equivalent to one similar to the knapsack problem and then solve itfor a special case. We then generalize this solution to generalregression designs. Furthermore, we provide an efficient algorithm tosolve the general non-linear, mixed integer programming problem when thenumber of predictors is small. We show the efficacy of the weighted LADestimator using numerical examples.

A Generalized Portmanteau Goodness-of-fit Test for Time Series Models

Fri, 29 Oct 1999 22:58:59 GMT

Title: A Generalized Portmanteau Goodness-of-fit Test for Time Series Models

Chen, Willa W.; Deo, Rohit S.

Abstract: We present a goodness of fit test for time series models based on thediscrete spectral average estimator. Unlike current tests of goodness offit, the asymptotic distribution of our test statistic allows the nullhypothesis to be either a short or long range dependence model. Our testis in the frequency domain, is easy to compute and does not require thecalculation of residuals from the fitted model. This is especiallyadvantageous when the fitted model is not a finite order autoregressivemodel. The test statistic is a frequency domain analogue of the test byHong (1996) which is a generalization of the Box-Pierce (1970) teststatistic. A simulation study shows that our test has power comparableto that of Hong's test and superior to that of another frequency domaintest by Milhoj (1981).

“Last licks”: Do they really help?

Fri, 29 Oct 2004 22:58:59 GMT

Title: “Last licks”: Do they really help?

Simon, Gary A.; Simonoff, Jeffrey S.

Abstract: Much has been written about the home field advantage in sports. Baseballand softball are unusual games, in that the rules are explicitlydifferent for home versus visiting teams, since by rule home teams batsecond in each inning (they have “last licks”). This isgenerally considered to be an advantage, which seems to be contradictedby the apparent weakness of the home field advantage in baseballcompared to that in other sports. In this paper we examine the effect of“last licks” on baseball and softball team success usingneutral site college baseball and softball playoff games. We find littleevidence of an effect in baseball, but much greater evidence insoftball, related to whether a game is close late in the game. Insoftball games that are tied at the end of an inning, batting last seemsto be disadvantageous later in the game, apparently related to thechances of the team scoring first to break the tie. By also examininggames where one team was playing on its home field, we are able to saysomething about benefits from playing at home that are not related to“last licks.”