SBR-9422575 Robert Engle As computers increase in power and memory it becomes feasible to collect and analyze data at higher and higher frequencies. Data sets that record every transaction-- the highest frequency possible-- now exist for many financial data sets as well as microeconomic transactions such as telephone calls and credit card purchases that are recorded by computers. The analysis of such data sets poses new and interesting econometric challenges, one of them being the choice of the proper interval of time within which to aggregate the data so as to generate a data set with observations spaced evenly apart. The problem with fixed interval analysis is that it can leave the investigator with many uninformative data points or disguise the periods of most interest. The PI proposes an alternative to fixed interval analysis which he calls autoregressive conditional duration. Instead of selection a fixed interval for analyzing the data, it is proposed to let the interval between transactions be the random variable to be analyzed. Thus the data set becomes a list of durations and characteristics of each transaction. This procedure models the time intervals directly without using auxiliary data or imposing assumptions on the causes of the time flow. The ACD model is used to analyze the price, volume and duration process for IBM stock transactions. This research can help institutions forecast market liquidity and volatility on a high frequency basis. It can also have implications for government interventions in financial markets either through transactions or through circuit breakers. The research also includes a number of other projects. One project examines non-linear business cycles. The nature of the non-linearities is always difficult to extract because there are relatively few cycles in recent history. However, if sectors and final goods categories regions of the country all participate in the business cycle, then the non-linearities ought to be common and easier to detect. An economic model is developed to motive these ideas. Another project involves using ARCH models of volatility to produce a term structure of volatility forecasts. These forecasts can be used to construct a portfolio whose value is unaffected by portfolio shocks. Such portfolios provide a new dimension in which to examine the accuracy of volatility forecasts: Is the term structure of volatility adequate to reduce the variance of such multiple maturity portfolios?
