Band 2, Heft 4

In this paper, we propose a model-selection approach to testing the expectations theory of the term structure of interest rates. Our method is based on the posterior information criterion (PIC) developed and analyzed by Phillips and Ploberger (1994, 1996) and extended to provide order estimation of cointegrating rank by Chao and Phillips (1997). This methodology has the advantage that issues of order selection—i.e., the determination of lag length and cointegrating rank in a vector autoregression—and hypothesis testing are treated within the same framework. Applying our procedure to interest-rate data from the International Financial Statistics, we find the expectations theory to be inconsistent with the data.

We examine the performance of artificial neural networks (ANNs) for technical trading rules for forecasting daily exchange rates. The main conclusion of our attempt is that ANNs perform well, and that they are often better than linear models. Furthermore, the precise number of hidden layer units in ANNs appears less important for forecasting performance than is the choice of explanatory variables.

In this paper, we examine the characteristics of market opening news and its impact on the estimated coefficients of the conditional volatility models of the GARCH class. We find that the differences between the opening price of one day and the closing price of the day before have different characteristics when considering various stock-market indices on which options are actively traded. The impact of a suitable positive-valued transformation of these differences has the effects of modifying the direct impact of daily innovations on volatility and reducing the estimated overall persistence of such innovations. The overall contribution of the variable is evaluated in an out-of-sample forecasting exercise, where we obtain significant improvements above the simple GARCH model.

We develop a class of ARCH models for series sampled at unequal time intervals set by trade orquote arrivals. Our approach combines insights from the temporal aggregation for GARCH models discussed byDrost and Nijman (1993) and Drost and Werker (1996), and the autoregressive conditional duration model ofEngle and Russell (1996) proposed to model the spacing between consecutive financial transactions.The class of models introduced here will be called ACD-GARCH. It can be described as a random coefficientGARCH, or doubly stochastic GARCH, where the durations between transactions determine the parameterdynamics. The ACD-GARCH model becomes genuinely bivariate when past asset-return volatilities are allowedto affect transaction durations, and vice versa. Otherwise, the spacings between trades are consideredexogenous to the volatility dynamics. This assumption is required in a two-step estimation procedure. Thebivariate setup enables us to test for Granger causality between volatility and intratrade durations. Undergeneral conditions, we propose several Generalized Method of Moments (GMM) estimation procedures, somehaving a Quasi Maximum Likelihood Estimation (QMLE) interpretation. As illustration, we present anempirical study of the IBM 1993 tick-by-tick data. We find some evidence that volatility of IBM stock pricesGranger-causes intratrade durations. We also find that the persistence in GARCH drops dramatically onceintratrade durations are taken into account.

Building upon Beaudry and Koop's (1993) analysis, we consider a "current depth of the recession" (CDR) variable in modeling the time-series behavior of the postwar quarterly U.S. unemployment rate. The CDR approach is consistent with the state-dependent behavior in the unemployment rate documented in the business-cycle asymmetry literature. We show that while the CDR effect is significant in-sample, no statistically significant out-of-sample forecast improvement is obtained relative to the linear alternative. Augmenting an AR(2) model by inclusion of the CDR term, however, does not significantly worsen the out-of-sample forecast performance.

The predictive accuracy of various econometric models, including random walks, vector-autoregressive and vector-error-correction models, are investigated using daily futures prices of four commodities (the S&P 500 index, treasury bonds, gold, and crude oil). All models are estimated using a rolling-window approach, and evaluated by both in-sample and out-of-sample performance measures. The criteria considered include system criteria, where we evaluate multiequation forecasting models, and univariate forecast-accuracy criteria. The five univariate criteria are root mean square error (RMSE), mean absolute deviation (MAD), mean absolute percentage error (MAPE), confusion matrix (CM), and confusion rate (CR). The five system criteria used include the trace of second-moment matrix of the forecast-errors matrix (TMSE), the trace of second-moment matrix of percentage-forecast errors (TMAPE), the generalized forecast-error second-moment matrix (GFESM), and a trading-rule profit criterion (TPC) based on a maximum-spread trading strategy. An in-sample criterion, the mean Schwarz information criteria (MSIC), is also computed. Our results suggest that error-correction models perform better in shorter forecast horizons, when models are compared based on quadratic loss measures and confusion matrices. However, the error-correction models which we consider perform better at all forecast horizons (one to five steps ahead) when models are compared based on a profit-maximization loss function. Further, our error-correction model, where the error-correction term is constructed according to a cost-of-carry equilibrium condition, outperforms our alternative error-correction model, which uses the price spreads as the error-correction term.