RATIO ANALYSIS USING RANK TRANSFORMATION
The kurtosis of the sample ratio distributions in this study also demonstrate severe
departures from normality. In all cases, the untransformed ratios exhibit leptokurtic properties,
with kurtosis ranging from 675.6361 to 7669.9680. Ranked transformations result
in a kurtosis of 21.20, slightly deviant from normal and platykurtic. This improvement in
kurtosis is unmatched by log or square root transformation, with the exception of the
square root of leverage (3.2825) which becomes almost normal. Other transformations,
however, do in some cases achieve substantive reductions in kurtosis. These include the
logarithmic transformation for Quick Ratio, Growth and Payout ratios.
Table 2, Panel A, presents the simple regression results for each of the ratios we tested,
using the full data set, and without truncation of any distribution. For both return measures,
ROA and ROE, the regression using the ranked ratio has substantively lower mean
square error, the highest F-value, and the highest R2 of all the transformations we tested.
For the other ratios, however, the improvement from the use of rank transformation, while
detectable, appears quite marginal. The quick ratio is only significantly associated with
stock returns when rank transformations are used. The sign of the coefficient, however, is
opposite that expected, i.e., negative because lower liquidity suggests higher risk and
return. In addition, the degree of explanatory power yielded by this regression, as measured
by r-squared, is almost zero. The payout ratio, leverage, and growth rate, while
significant in all cases and with coefficient signs of the conforming to our priors, yield
similarly weak associations with stock returns. Rank transformations, however, do yield
marginally superior goodness-of-fit statistics in all cases except for the leverage ratio.
As discussed earlier, a common procedure used to control for the problems with ratios
is truncation. Although rank transformation is still arguably a superior technique, since it
requires no subjective choices as to cutoff and/or outlier deletion, we wished to determine
whether truncation has any impact on the general ability of rank transformations to
outperform their untransformed, logarithmic, and square-root ratio relatives.We truncated
our sample on both ends, deleting a total of 5% of all sample observations. Results are
presented in Table 2, Panel B. Information in Table 2 reveals similar results to those
obtained with the full dataset. The models using the ranked ratios are superior for ROA,
ROE, Lev, and Growth, with higher F-values, lower mean square errors, and higher R2s.
The raw ratio for Payout is best and, for QR, the logarithmic model shows the best
goodness-of-fit statistics.
The multiple regression results are presented in Table 3. Panel A presents diagnostic
and goodness-of-fit regression results for each of the thirteen years in our sample. In all
cases, for the goodness-of-fit measures that we examined (mean square error and
r-squared) rank transformations showed substantive improvement over models that used
untransformed ratios. In comparison with models that used log-transformed, and square
roots of the accounting ratios, the multiple regression models using the ranked ratios
generally exhibited the best goodness-of-fit statistics and were most significant, with the
highest F-value, lowest mean square error, and highest adjusted R2. The exceptions were:
1980, 1981, and 1985, when log-transforms performed slightly better; and 1985 when the
model that used square root transforms also performed slightly better than rank transforms,
showing marginally improved r-squared and mean square error statistics.