INTRODUCTION . Deciding when a series is a good candidate for seasonal adjustment can be difficult. There ar2 situations where a series may show evidence of seasonality, gut because of a dominating irregular component, for exampl2, or a volatile seasonal component., many of its seasonal factors cannot be estimated reliably. In these circumstances, the estimates of a given month's seasonal factor can change substantially when more data are added to the series and earlier data deleted. Some

more »

... al adjustment programs, such as X-11 and X-ll-ARIMA, provide diagnostics which can be used to help the analyst make this decision. We have found, however, that the diagnostics provided by X-11 and X-ll-ARIMA are sometimes inadequate. In this article, we will discuss two new sets of measures which help to determine when a series can be seasonally adjusted reliably by a proposed seasonal adjustment methodology. The first set, described in section 3, compares seasonal (and trading day) adjustments performed on sliding spans of data. These enable the analyst to se2 how stable the estimates are of seasonal factors and of month-to-month changes in the seasonally adjusted data. If too many months have unstable estimates, it is an indication that the adjustment nethod used cannot reliably adjust the ssries being examined. 2 The second set, describ2d in sectiorl 5, uses the rgvisions history of a series to provide measures of (a) how much the initial seasonal adjustments get revised in later years and (b) how rapidly these adjustments converge to their final value. We will us2 these measures to determine whether the seasonal adjustments of the series being analyzed are subject to excessive amounts of revision and to help ascertain if the final adjustments are merely artifacts of the finite lengths of the adjustment filt2rs used. In either of these situations, the 'szasonal adjustments are likely to be unreliable. In the remaining sections, we us2 these methods in conjunction with others to * analyze a number of Census Rureau series. In sections 4 and 6, a Census Bureau series called XU3 (exports of mineral fuels, lubricants and related materials) serves as an example for a detailed illustration of the us2 of our new techniques to determine if a series is a candidate for seasonal adjustment using X-11 (or X-ll-ARIMA without the ARIMA forecasts). Then, in section 7, thirty regional foreign trade series are analyzed with these new techniques and with some conventional diagnostics. The reader who is chiefly interested in our conclusions regarding the adjustment of these series can proceed directly to sections 3, 7.1 and 7.4. Although we use X-11 and X-ll-ARIMA in this studyl, the techniques presented here can be adapted for use with other seasonal adjustment methods. CONVENTIONAL ANALYSIS OF XU3 The graph of the series X113, given in Figure 2.1, does not reveal any obvious persistent seasonal pattern, apart from a trough each December. Also, it suggests that the series undergoes a significant change around 1974. The analyst should carefully consider the question of what data span to use. For illustrative purposes, we will begin with an analysis of the 3 f?rll series (January 1956 to 92cemher 19q3) and later give a summary of an analysis performed on the shortened series (January 1974 to December 1983). (Insert Figure 2.1 about here) X-ll-ARIMA was used, without forecasting, to seasonally adjust the full series u-sing 3x9 seasonal filters. Some of the diagnostics from X-ll-ARIMA support seasonally adjusting the series, but others are cautionary. A summary *of conventional diagnostics is given in Table 2.1. The F-tests used to detect stab12 seasonality tentatively suggest that there is significant seasonal variation in this series. The F-test for (linearly) moving seasonality indicates that there is no linear movement in the pattern of seasonality which would prevent its reliable estimation. (For more information on these F-tests, see r21, r31, C41.) (Table 2.1 goes near here) However, there are also indications -that this series may not be a good s candidate for seasonal adjustment. There are several signs that the series is highly irregular. The proportion of the sum of squared percent changes attributed to the irregular component is high for this adjustment (59.0 percent at lag one, 27.6 percent at lag 3 according to table F 2.B of the X-ll-ARIMA analysis for XU32). A graph of the SI ratios 3 for the last six years of data (given in Figure 2 .2) shows how this irregularity is reflected in the spread of the values of the SI ratios for the individual calendar months. This kind of spread can lead to degraded estimates of seasonal. factors obtained as weighted averages of these SI ratios.