RDP 2004-02: The Impact of Rating Changes in Australian Financial Markets 3. Methodology

We examine the impact of the announcement of rating changes on debt and equity prices using relatively standard event study techniques.^[4] Given the depth of information available about companies from a variety of sources, our null hypothesis is that announcements by rating agencies should not be associated with impacts on market prices. We depart from a number of earlier studies by excluding rating changes that coincide with other major corporate news such as announcements concerning earnings, mergers and divestments. The reason for excluding these events, which Hand et al (1992) refer to as ‘contaminated’ events, is to ensure that we are capturing the impact (if any) of rating changes per se on market prices and not simply the effect of other value-relevant news.^[5] In our case, this rules out more than half of all rating changes, and the announcement effects that we document in Section 5 would be substantially larger had we kept all the ‘contaminated’ events in the sample.

We define the date of the announcement (the ‘event’) as t = 0, and a window of 20 days on either side of the event as the ‘event window’ (Figure 1). We attempt to isolate the movement of financial prices in the event window that is not due to factors influencing the overall market. In the case of equities, this is the ‘abnormal return’, and in the case of bonds, this is the change in spreads, which we define below. We assess the statistical significance of these abnormal movements based on the movements in financial prices in the ‘estimation window’, the 100-day period prior to the event window.

In the case of equities, we estimate a standard market model using daily returns in the estimation window. For each event (i), the daily (log-differenced) equity price return for the relevant company (R_it) is regressed upon the corresponding broad market return (the All Ordinaries Index, R_mt) using ordinary least squares:

Abnormal returns for the event windows (AR_it) are then defined as the difference between actual returns and the returns predicted by the market model using the parameters from the estimation window:^[6]

The daily average abnormal return (AAR_t) for any n events (either upgrades or downgrades) is then calculated by summing (in event-time) across the n events. The cumulative average abnormal return (CAAR(τ₁,τ₂)) between any two days τ₁ and τ₂ within the event window is defined as the sum of the average abnormal returns over that period. The statistical significance of average abnormal returns in the event window can be assessed using the estimate of the standard deviation of average abnormal returns in the estimation window which is denoted s(AAR_t). Under the assumption of i.i.d. normally distributed abnormal returns, the ratio of AAR_t to s(AAR_t) is distributed as a Student's t with n degrees of freedom. In addition, under these assumptions the standard deviation of any cumulative average abnormal return is given by s(AAR_t), multiplied by the square root of the number of days in the period.

In the case of debt securities, we analyse the impact on bond yields, or more specifically the impact on bond spreads, relative to the Commonwealth Government bond of comparable maturity. The movement in each bond's yield spread (measured in basis points, or hundredths of one percentage point) provides a ready-made proxy for its abnormal performance relative to the overall market.^[7]

The reason for considering spreads rather than bond returns is that the bonds in our sample have a range of maturities, so we would expect that a given impact on required bond yields or spreads would have different effects on the prices of different bonds, depending on the maturity of the bond. Hence, it may make little sense to examine abnormal returns of bond prices.

For our statistical tests, we calculate the average basis point changes for each day of the event and estimation windows, for upgrades and downgrades separately. These average changes can be summed over time to compute cumulative average basis point changes. The statistical significance of the average and cumulative spread changes in the event window can then be determined by comparing them to the standard deviation of spread changes in the estimation window. The expectation is that if announcements of upgrades (downgrades) convey information to market participants, spreads on average will fall (rise) immediately following news of the rating change.

In addition to the statistical tests for average abnormal returns or average spread changes, we also present tests based on the proportion of positive or negative changes in market prices. This type of test may be useful if abnormal returns or spread changes are not normally distributed. For this test, we compare the actual proportion of events that are positive and compare this with the theoretical distribution (a binomial test) under the null hypothesis that this proportion is equal to 0.5. Here the expectation is that if rating announcements convey information, the proportion of abnormal returns that are positive will be greater than (less than) 0.5 following upgrades (downgrades), and the proportion of spread changes that are positive will be less than (greater than) 0.5 for upgrades (downgrades).

Footnotes

See MacKinlay (1997) for a general discussion of event studies. [4]

Of course, reports about major companies occur quite frequently. In cases where there were simultaneous news reports unrelated to the rating announcement, we omitted events if this other news appeared substantially more important than the rating announcement. [5]

An alternative approach would be to recognise – based on earlier US studies cited in Section 2 (and our data) – that decisions to change ratings may not be completely exogenous with respect to recent return outcomes. In particular, estimates of α_i in the estimation window are unlikely to be reliable – downgrades are likely to be associated with negative estimates of α_i and upgrades with positive estimates. On the other hand, estimates of β_i should not be problematic. This would suggest estimating the market model including a constant term in the estimation window, but then defining the abnormal return in the event window to be given by R_it – β_iR_mt; i.e., removing the impact of the problematic constant term in the abnormal return. The result of this change is to make the pre-announcement and announcement effects in Figure 4 and Table 2 somewhat larger, but overall the results are little changed. [6]

Because corporate spreads tend to vary over time (i.e., corporate yields do not always move one-for-one with government yields), in principle one could measure spreads relative to other similarly rated bonds. However, this is difficult given the relatively small number of bonds within any single rating and maturity category in existing Australian indices. [7]