3.12. To give an impression of the revisions process through which balance of payments statistics pass, Appendix 1 shows graphically the way in which estimates of the principal aggregates evolve over time. Graphs are included for monthly, quarterly and annual statistics which compare the latest (i.e. at December 1994) estimate with the initial estimate and successive revisions to the initial estimate. The comparison expresses estimates as a ratio of the latest estimate (i.e. a ratio of 0.50 means the value of the estimate is half the value of the latest estimate). The graph plots this ratio for the initial estimate (the first point on the horizontal axis) and a number of revised estimates 24 revisions for monthly data, 12 revisions for quarterly data and 6 revisions for annual data. Medians of the ratios calculated at each estimation point have been used. For example, the ratio plotted at, say, revision number 3 is the median of the ratios for all observation periods, where the ratio for each period is calculated by dividing the estimate after the third revision by the corresponding latest estimate. Also shown on the graphs, as dashed lines, are the upper boundaries of the first and third quartiles of calculated ratios. The area between these bounds therefore represents the behaviour of the middle 50% of observations or, alternatively, the 25% of ratio observations that lie below the median and the 25% of ratio observations above it.
3.13. Examination of these graphs gives an indication of the bias present in preliminary estimates (and therefore in revisions to those estimates), the spread of ratios around the median ratio at different stages in the estimation cycles, and how quickly the preliminary estimates approach the latest value.
Bias in monthly statistics
3.14. Looking first at the current account, it can be seen that, merchandise exports and imports (Graphs A.1 and A.2) display negligible bias and initial estimates are revised to the latest estimates very quickly. Similarly the spread of ratios calculated for the initial estimates of the middle 50% of observations is small - within about 1% of the median latest value for exports, and even closer for imports.
3.15. On the other hand, the non-merchandise components of the current account are less reliable. The most striking feature of the non-merchandise credits and debits aggregates (Graphs A.3 and A.4) is that the median values are all estimated to be less than their latest values, and even after two years of revision, the upper boundary of the third quartile of estimate ratios remains below the final estimate. On average, the preliminary estimates can therefore be said to display a clear negative bias. Both the credits and debits median measures do improve over the first 24 months of revision, and the ratio spreads narrow.
3.16. Of the non-merchandise components, unrequited transfers credits (Graph A.9)) displays the least bias and the spread of ratios for the middle 50% of observations is initially within about 3% of the median, and at the median after one annual revision cycle. On the other hand, the initial estimates of services credits (Graph A.7), with a median ratio of about 0.88, are the most biased. While services credits estimates improve in accuracy over two years of revisions, the median estimate at the end of the period remains at less than 94% of the final estimate. A number of methodological changes introduced in the June quarter 1994 contribute to this. First, improved estimates of expenditure in Australia by non-resident travellers were introduced reflecting, principally, the inclusion of expenditures on pre-paid package tours. Second, introduction of a new collection methodology for the transportation items resulted in improved coverage of port expenditures in Australia by foreign carriers and freight on exports earned by resident carriers. Third, some relatively minor upward revisions were made as a result of changes to the collection methodology for insurance services. Together these changes accounted for an increase of about 10% in total services credits estimates for 1992-93 and for progressively smaller proportional increases in earlier years.
3.17. Services debits (Graph A.8), with a median ratio for initial estimates of about 0.94, also displays significant bias throughout the two year revision period. The ratios accounted for by the middle 50% of observations are relatively evenly spread around the median. While the median ratio improves over the revision period shown, it does not exceed 98%.
3.18. Income credits (Graph A.5) has a much broader spread for the middle 50% of observations than other components exhibit, with the ratios of initial estimates to the latest estimates ranging from about 0.78 to 1.13 indicating a marked variability in the reliability of estimates of this aggregate. However, over two years of revisions, median estimates of income credits approach the latest value and the spread of ratios narrows markedly until about the sixteenth revision. An important reason for the variability of monthly income credit estimates is the method used to estimate reinvested earnings on Australian direct investment abroad, an item that has been particularly variable and therefore difficult to forecast. At present, initial estimates are based on annual forecasts evenly distributed over the year. Once the results of the annual Survey of Foreign Investment become available, which may be up to 18 months after the initial estimate has been made, there is often a need for substantial revision (both increases and decreases).
3.19. Introduction, in the January 1994 issue of 5301.0, of an improved methodology for estimating income earned by residents of Australia while travelling outside Australia, also resulted in revisions to past periods' estimates. Revised estimates were made using the results of a new ABS survey (Survey of Returned Australian Travellers) designed to give benchmark estimates of residents' earnings and expenditure while abroad. Although these revisions were not as substantial or variable as those made to estimates of reinvested earnings, they have resulted in increases in total income credits estimates for 199192 and 1992-93 respectively of about 5% and 10%.
3.20. Income debits (Graph A.6) behave quite differently. While both initial and first revision median estimates have ratios less than 0.96, from the second revision the median estimates rapidly approach the final estimate which is closely approximated from the eighth revision onwards with a narrowing band of ratios around it. While revisions to estimates of reinvested earnings debits are significant in absolute terms, they are a smaller component of the very much larger income debits total than is the case for credits.
3.21. Median initial estimates of transfers debits (Graph A.10) are a little more accurate than most other components (with ratios close to 97%) and with ratio spreads that are reasonably tightly clustered around the median. However, median estimates after two years of revision are only marginally closer to final estimation levels, and the spreads only a little tighter.
3.22. The net result of the estimation process for the components of the current account is reflected in the behaviour of the balance on current account (Graph A.11). Analysis of such net results needs to be undertaken with care because these series will reflect all methodological changes impacting on component series. The initial median estimate of the balance is very close to the final estimate, although the spread around this estimate is relatively large. Revisions over the two year period appear to increase the gap between the median (revised) estimates and the final estimates, with the median estimate at the end of the period about 3% higher than the final estimate. The median ratios shown in the graph display significant positive bias developing after the seventh revision. This somewhat paradoxical pattern reflects the significant methodological improvements recently made to services credits estimates, the persistency of the negative bias that exists in initial income credits estimates, and the improving accuracy of the debits estimates over the first two years of revision. The methodological improvements to services credits estimates have a paradoxical effect because previously revisions to the various current account items tended to cancel each other out, but this is happening to a lesser extent with the improved methodology for services credits. It should be noted that a positive bias in estimates of this balance refers to an overstatement of the current account deficit.
3.23. In monthly statistics, capital account data are available only for net transactions of the official sector (the Reserve Bank and the general government sector) and certain non-official public sector transactions. Other capital account transactions (mainly those of the non-official private sector), together with the balancing item, are derived as a residual after the balance on current account has been deducted from the identified public sector capital account transactions. Graphs A.12, A.13 and Graph A.14 in Appendix 1 show the patterns of behaviour of the official sector.
3.24. Estimates of net capital transactions of the Reserve Bank (Graph A.12) are very reliable. Net capital transactions of the general government sector (Graph A.13) are less reliable and display negative bias (an understatement of net inflows or overstatement of net outflows) in initial estimates (a median ratio of about 0.91) which is progressively reduced until, from about the sixth monthly revision, the estimate remains close to the latest estimate. The middle 50% of observations initially shows a wide distribution of ratios around the median (from 0.66 to 1.14). However, within about one year, this area has narrowed to span a band of ratios from 0.85 to 1.14.
Bias in quarterly statistics
3.25. The graphs in Appendix 1 relating to quarterly statistics, in general, display a similar picture to that for monthly statistics. Within the current account, merchandise exports and imports statistics are very reliable and unbiased (Graphs A.15 and A.16). As with monthly statistics, the non-merchandise items show clear negative bias for both credit and debit items. On average, ratios of initial estimates to latest estimates for total non-merchandise credits (Graph A.17) have a value of about 0.92. The equivalent average ratio for non-merchandise debits (Graph A.18) is a little better at 0.94. The quarterly ratios improve to about 97% (credits) and 99% (debits) over the three years of revision history incorporated in the graphs.
3.26. As with the non-merchandise aggregate series, quarterly component series behave very similarly to the monthly series. Unrequited transfers credits (Graph A.23) displays the least median bias, with medians for initial estimates and for all subsequent revisions at or very close to the final estimates. By the fourth revision, all estimates in the second and third quartiles are also at the final estimates.
3.27. The initial estimates that appear least reliable are income credits (Graph A.19) and services credits (Graph A.21) where the median ratios are about 0.92 and 0.90, respectively. The income graph is similar to the equivalent monthly graph in that it shows the wide spread of ratios displayed by the middle 50% of initial estimates and early revisions. The main reasons for this variability have been noted earlier.
3.28. The quarterly graph for services credits (Graph A.21) is notable for the relatively large proportion of revisions that remain to be applied after the twelfth quarterly revision. The median ratio value of less than 0.95 after three years of revision is little better than achieved after the two years of revision presented in the monthly graph. It indicates that, on average, the estimates will be increased by more than 5% to reach latest values. It should also be borne in mind that 50% of quarterly estimates of that aggregate are even less reliable at that stage. The reasons for this outcome are discussed above in noting the performance of the monthly series.
3.29. The quarterly graph of the balance on current account (Graph A.25) records the net outcome of all current account transactions. As with the monthly series, the quarterly median series shows that the initial estimate is very close to the final estimate but the second and subsequent estimates show some positive bias (i.e. an overstatement of the current account deficit) which remains for the three years of revision under analysis and primarily reflects the impact of the methodological improvements noted for estimating services credits. The spread around the median initial estimate is relatively large but it narrows progressively over the three year period.
3.30. Quarterly capital account statistics are generally less reliable than quarterly current account statistics, with most of the capital account items included in Appendix 1 showing negative bias. The exceptions are the estimates of the Reserve Bank transactions, which are extremely reliable (Graphs A.28 and A.29), and investment abroad by general government (Graph A.27). In the case of the general government investment abroad series, all median estimates are at final levels, but there is a wide spread of estimates for each revision period (Graph A.27). It needs to be borne in mind that the net investment transactions abroad by the general government sector vary significantly from period to period and in some periods may have values close to zero. When values close to zero are revised, even by small absolute amounts, large proportional changes may occur, accounting for the wide dispersion of ratios.
3.31. Of the remaining aggregates analysed, foreign investment in the general government sector (Graph A.26) shows the least bias in initial estimates, with a median ratio at about 0.79. This bias in initial estimates is due in part to revisions to estimates of foreign borrowing domiciled in Australia for 1989-90 and earlier years resulting from improved coverage of Australian securities held by nominees on behalf of non-residents. Median estimates are close to final levels from the sixth revision. The middle 50% of ratio observations are widely dispersed for initial estimates and early revisions but narrow rapidly until the sixth quarterly revision.
3.32. The negative bias present in estimates of transactions of the non-official sector (Graphs A.30 and A.31) is most pronounced for Australian investment abroad (AIA) with the median ratio of initial estimates to latest estimates being about 0.66. On average, therefore, initial estimates are revised upward by about 50% before reaching their latest value. The middle 50% of observations classified to AIA display a spread that is very broad, from as low as 0.37 to 0.91. Reasons for the degree of variability found in quarterly AIA estimates are associated with the difficulty of measuring the reinvestment of earnings component of Australian investment abroad (see earlier comments in the context of estimates of the reinvested earnings component of income credits) and achieving satisfactory quarterly coverage of other investment flows.
3.33. Estimates of transactions of the non-official sector (Graph A.30) relating to foreign investment in Australia (FIA) are also negatively biased on average, with the median ratio for the initial estimate at about 0.69. However, the median is revised upward relatively quickly - after three revisions it is within about 10% of the latest value. After then, the revisions only slowly raise the median towards the latest values. The middle 50% of initial observations classified to FIA are clustered across a slightly tighter spread than was displayed for AIA, and this spread tightens more quickly around the median with the result that it is much narrower by the sixth revision.
3.34. Also included in Appendix 1 are graphs of the balance on capital account (Graph A.32) and the balancing item (Graph A.33). Because they refer to residual balances which may vary considerably in size from period to period, and in the case of the balancing item frequently change sign, care is needed in their interpretation. The quarterly graph of the balance on capital account records the net outcome of all capital account transactions. The quarterly median series shows a marked negative bias (i.e. an understatement of the capital account surplus), with the ratio for the initial estimate around 0.78 and for the estimate after three years around 0.93. The spread around the median estimates is large and volatile up to the sixth quarterly revision when it narrows; and from then to the end of the three year revision period the spread remains fairly constant.
Bias in annual statistics
3.35. Appendix 1 Graphs A.34 to A.52 relate to annual statistics. The graphs display generally similar patterns of bias to monthly and quarterly statistics but because only annual revisions cycles are captured, and because the period of analysis extends from 1993-94 back to 1981-82 for some series, care is needed in both interpreting the data presented and comparing the annual series with either the monthly or quarterly series which cover a shorter time span (from January 1986 to June 1994). Three particular points should be noted. First, the ongoing improvements in methodology implemented over the last decade have given rise to often substantial revisions to item estimates for the period 1982 to 1985, which are captured in some of the annual analyses that cover this earlier period. However, this early revisions experience is excluded from the analyses for the monthly and quarterly series which commence in 1986. Some of the annual series may therefore appear to be both more biased and less reliable than the estimates with greater periodicity. Second, in the monthly graphs, the 24 months of revisions experience that is shown captures only two annual revisions cycles, while for quarterly series, 3 annual revision cycles are displayed. In both cases quite significant revisions are often shown as remaining to be applied between the last revision cycle presented and the final estimate produced in the December 1994 (quarterly) or January 1995 (monthly) balance of payments publications. For the annual series, the graphs capture explicitly most of the extra revisions experience so that many of the series are shown with medians rising to, or very close to, the final estimate. Third, any comparison of revisions to monthly, quarterly and annual capital account series is complicated by the net basis of recording capital account entries and the frequent turnarounds from inflows to outflows and vice versa in some series. This can mean, for example, that the annual estimate for some items in some periods is lower than some or all of the corresponding quarterly estimates, without regard to sign.
3.36. As with monthly and quarterly data, the annual merchandise trade estimates are very reliable and unbiased while most non-merchandise components of the current account show clear negative bias for both credit and debit items. For the non-merchandise credit and debit aggregates (Graphs A.36 and A.37) the median ratios of initial annual estimates to latest estimates (about 0.88 for credits and 0.94 for debits) are a little less reliable (i.e., lower) than the equivalent ratios for monthly and quarterly statistics. As mentioned in the previous paragraph, this reflects the revisions experience of the early 1980s.
3.37. The least reliable annual current account aggregate, in terms of the median ratio of initial estimates to latest estimates, is income credits which has a ratio of about 0.86 (Graph A.38). Like the quarterly statistics, the middle 50% of ratios for income credits is initially spread across a wide band (of about 30 ratio points) but after two revisions this narrows substantially to a band of two ratio points. Reasons for this variability have been suggested in the discussion on bias in monthly statistics.
3.38. Overall, service credits is the least reliable annual series. While the median initial estimate has a ratio of about 0.91 (Graph A.40) which is better than for income credits, the series is notable for the relatively large proportional revision that remains to be applied after the last revision cycle shown on the graph. Graph A.40 in Appendix 1 shows that the annual estimates that have passed through six revision cycles remain to be revised upward by, on average, about 6%. This is a reflection of the revisions introduced in June quarter 1994 that have been described in the context of quarterly statistics.
3.39. Annual series for other components of the current account behave similarly to their quarterly equivalents. However, the different pattern for the second quartile of ratios in unrequited transfer credits (Graph A.42) is a reflection of revisions to estimates of transfers by immigrants introduced in December quarter 1985, and which are therefore not reflected in either the monthly or quarterly analyses.
3.40. Graph A.44 in Appendix 1 shows the reliability pattern displayed by the balance on current account. While this balance is an important key aggregate in its own right, it reflects the net sum of revisions to all the components of the account, including methodological revisions that shift the series across time, and therefore the usefulness of the reliability measures presented in the graph is limited. The annual median series shows relatively small bias in the 6 years of revision that are under examination. On average, the initial estimates of the current account deficit understates the final estimate by about 2%. Subsequent revisions result in the estimate fluctuating at or slightly above the latest figure. On average, after 6 revisions the estimate overstates the latest figure by 2%. Half of the initial annual estimates of the deficit fall between an understatement of 6% or an overstatement of 3%. This spread narrows progressively over the 6 year period and results in an overstatement of between 1 and 3% by the sixth year of revisions.
3.41. Graphs for annual capital account aggregates (Appendix 1 Graphs A.45 to A.52) behave similarly to their quarterly equivalents with aggregates for the Reserve Bank being completely free from bias and the median ratios for other items showing negative bias. The median ratio for transactions of the general government sector classified to FIA is initially a little more biased than its quarterly equivalent, reflecting revisions experience in the early 1980s. The band of ratios representing the middle 50% of ratios is more tightly clustered, as is expected in an annual analysis where initial estimates already incorporate several revisions emerging from monthly and quarterly source data. Like the quarterly series, the spread of ratios for general government sector transactions classified to AIA (Graph A.46) is erratic but it must be remembered that these ratios reflect estimates that are relatively small and very variable in size and direction.
3.42. Reliability of annual non-official sector statistics for AIA is a little better than the equivalent quarterly statistics (Graph A.50), reflecting the substantial upward revisions experienced over the first three quarterly revision periods. The middle 50% of ratios of initial estimates to latest estimates for AIA are as widely spread around the median ratio as the quarterly series, narrowing after two annual revisions to about the same spread as is displayed after eight quarterly revisions. For annual non-official sector statistics relating to FIA (Graph A.49), the initial estimates show a substantial negative bias on average (at a median ratio of 0.73) as do their quarterly counterparts (at a median ratio of 0.69). The bias in the median series reduces sharply after one annual revision cycle and continues to fall until the final estimate is reached after five years.
3.43. As with the quarterly series, graphs for the net series balance on capital account (Graph A.51) and the balancing item (Graph A.52) should be interpreted with care. The annual graph of the balance on capital account records the net outcome of all capital account transactions. The annual median series shows a strong negative bias initially which reduces progressively until the final estimate is reached after five years. On average, the initial estimate of the capital account surplus understates the final estimate by around 20%. Half of the initial annual estimates of the surplus involve understatements of between 6% and 27%. This spread narrows to the point where, after five years, the understatement is between zero and 2%. The fact that estimates of the balance on capital account become final more quickly than estimates of the balance on current account largely reflects differing revisions practices with the different data sources used. In general, revisions to capital account series are not identified for periods earlier than the latest five years, whereas revisions to current account series can often be backcast for much longer periods.