j. david cummin and alwys n powell€¦ · j. david cummin and alwys n powell 1. introduction when...
TRANSCRIPT
Astin Bulletin 11 (1980) 91-106
THE PERFORMANCE OF ALTERNATIVE MODELS FORFORECASTING AUTOMOBILE INSURANCE PAID CLAIM COSTS
J. DAVID CUMMINS and ALWYN POWELL
1. INTRODUCTION
When setting or modifying insurance rates on the basis of observed data,adjustments should be made for inflation between the average claim date ofthe observation period and the average claim date covered by policies to whichthe new rates will apply. In automobile insurance ratemaking in the UnitedStates, the adjustment is made as follows 1:
(1) La = U x (i + p)&
where La = trended losses and loss adjustment expenses incurred,Lo — observed losses and loss adjustment expenses incurred during the
experience period,(5 = the predicted rate of inflation per quarter, andk = the number of quarters for which losses are being trended.
The inflation coefficient, (3, is an exponential growth factor that is estimatedusing the following equation 2:
(2) lnj^ = x
where y% = average paid claim costs for the year ending in quarter t,t = time, in quarters, t = 1, 2, . . . , 12, and
et = a random disturbance term.
The estimation period for the equation is the twelve quarters immediatelypreceding the first forecast period, and the estimation is conducted utilizingordinary least squares. Annual averages of claim costs are used as the dependentvariable to smooth random and seasonal fluctuations.
The use of equation (2) can be expected to lead to two major problems:l. Since the equation merely extrapolates past inflation rates into the future,
the forecasts are likely to be seriously inaccurate when the inflation rate inthe forecast period is substantially different from that in the preceding twelvequarters.
1 In some cases, an adjustment also is made for frequency. The formula in such casesis: La = L0(i + J3s)
fc . (i + (3/)*, where (3, = the predicted rate of change in severityand (3/ = the predicted rate of change in frequency. This article is limited to an analysisof severity, which is the predominant source of claims cost inflation.
2 Trend factors are discussed in COOK (1970). When frequency trend factors are em-ployed, they also are obtained through exponential trending.
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
0,2 J. DAVID CUMMINS AND ALWYN POWELL
2. Forecasts based on four quarter moving averages are likely to respondslowly to changing inflation rates. Thus, premium rates will be unresponsive,contributing to the underwriting cycle.
An alternative approach to forecasting average paid claim costs is to usean econometric model which predicts claim costs as a function of relevanteconomic variables. This approach has become feasible in recent years asprogress has been made in developing forecasting models for major economicaggregates in most industrialized nations 3. The purpose of this article is todevelop and test the accuracy of econometric models for predicting automobileliability insurance paid claim costs. The estimation and testing are conductedusing claim cost data for the United States.
2. EQUATION SPECIFICATION AND METHODOLOGY
Among the advantages of the exponential trend model (equation (2)) are itsfamiliarity and simplicity. These are important qualities because rate increasesin most states must receive regulatory approval. Since regulators are likelyto be suspicious of highly technical models, any methodology proposed as areplacement for exponential trending should be as simple as possible. Hence,the approach taken in this article is to begin by utilizing ordinary least squares(OLS) to estimate equations with the simplest possible specification. Teststhen are conducted to determine whether the resulting equations depart fromthe assumptions underlying the OLS technique. Where departures are found,the equations are reestimated with the appropriate adjustments.
The three basic specifications used in the research are the following:
(3) yt = ct. + $xt + zt
(4) In yt = a + (3 In xt + st
(5) yt = cL
where yt = average paid claim costs in quarter t,xt = a price or wage index, andS{ = the random disturbance term.
Linear and log-linear equations in which two economic indices appear asexplanatory variables also were estimated, and dummy variables for seasonalitywere tested with each type of equation. Since the same estimation methodsapply to (3) and (4), the examples pertaining to these equations in the followingdiscussion are based only on equation (3).
3 Models of the U.S. economy are reviewed in KLEIN and BURMEISTER (1976). Formodels applicable to other countries, see WAELBROECK (1976).
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
AUTOMOBILE INSURANCE CLAIM COSTS 93
The classical ordinary least squares model incorporates the following as-sumptions about £(4:
(6) E(et) = o
(7) E[(xt-Xt)et] = o
(8) £(s«) = a*
(g) E{tt . zt~i) = o, for i = l , 2, . . . t.
If the assumptions are satisfied, ordinary least squares yields parameterestimates for equations (3) and (4) that are best linear unbiased, i.e., that havethe smallest variances among all linear unbiased estimators. It is not necessaryto assume normality in order to achieve this result, although the usual tests ofsignificance for the coefficients will be correct only asymptotically if normalityis not present. Even if all of the assumptions are satisfied, the ordinary leastsquares estimators for equation (5) are biased although they are consistentand asymptotically efficient.
As assumptions (6) and (7) rarely cause problems, attention is focused onassumptions (8) and (9), i.e., homoskedasticity and the absence of serialcorrelation, respectively. If either assumption is violated, ordinary leastsquares estimators of the parameters in equations (3) and (4) are unbiased andconsistent but are inefficient both in finite samples and asymptotically. Forequation (5), violation of assumption (8) implies that ordinary least squaresestimators are consistent but asymptotically inefficient, while violation ofassumption (9) renders the OLS estimators both inconsistent and asymptotical-ly inefficient.
Because of the potential importance of assumptions (8) and (9) for forecastaccuracy, tests are conducted to determine whether they are satisfied in theestimated relationships. For equations (3) and (4), the Durbin-Watson dstatistic is used to test for serial correlation, while the hypothesis of homo-skedasticity is tested using a procedure suggested by GOLDFELD and QUANDT
(1965) 5. Because the Durbin-Watson d statistic is inappropriate for equationswith lagged dependent variables among the explanatory variables, an alter-native procedure suggested by DURBIN (1970) is used to test for serial correla-tion in equation (5) 6. The Goldfeld-Quandt heteroskedasticity test also isapplied to equation (5).
4 The discussion of estimation theory in this section draws heavily on KMENTA (1971).5 The Goldfeld-Quandt test has been shown to be at least as powerful as other major
heteroskedasticity tests. See HARVEY and PHILLIPS (1974) and HARRISON and MCCABE
(1979)-6 DURBIN proposed two tests—one based on the statistic h and the other on regressions
involving the OLS residuals of the equation being tested. The latter test is used herebecause the h statistic has been shown to be biased in small samples. See SPENCER (1975).
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
94 J- DAVID CUMMINS AND ALWYN POWELL
If the tests indicate that autocorrelation is present in equation (3) or (4),the approach is to hypothesize that e< = pst-i + Ut, where ut satisfies (6)through (9). The equation then is reestimated by solving the following problem:
T
(10) Minimize 2 [yt - pyt-i - a(l - p) - $(xt - pxt-1)]2
(a .P .p) ( - «
Minimizing (10) is asymptotically equivalent to maximum likelihood estimationif Ut is normally distributed 7.
A method suggested by GLEJSER (1969) is used to adjust for heteroskedasti-city. This technique is based on the hypothesis that st — P{z)wt, where P(z)is a function of a mathematical variable z, and wt satisfies (6) through (9).The variable z is taken to be (one of) the independent variable (s) in the re-gression, in this case xt. The estimation is conducted as follows:
(1) Estimate equation (3) or (4) using OLS and obtain the residuals, sj =
yt-x- $xt.(2) Regress [ s* | on simple functions of the independent variable such as
To + Yixt> To + Yixl< To + Ti^2» e^c- ar*d choose as the functional form P(xt) theequation with the highest explanatory power (R2).
(3) Transform the regression equation by dividing each term by P(xt).E.g., if the best regression for | e* | is yo+yi*2, the transformed variables are
equal to y't = ytl(y0 + r>«), x't = *(/(YO + Y>?)> a n d c't = ^/(Yo + Yi**).
where c't is the transformed constant term.(4) Estimate specification (3) or (4) using the transformed variables and
OLS. For example, equation (3) becomes: y't = a.c't+ $x't + wt.If the hypothesis is correct, the error term of the transformed equation
satisfies (6) through (9). However, the correction will be only approximate forfinite sample sizes.
When the error terms in equation (3) or (4) are characterized by both auto-correlation and heteroskedasticity, the estimation problem is more difficult.The ideal solution would be to use a generalized nonlinear estimation procedurewhich would correct simultaneously for both problems. Since computerprograms to perform this type of estimation were not readily available to theauthors, two alternative approaches were adopted, based on extensions ofGlejser's method.
Method 1. The following error structure is hypothesized: sj = pst-i + Ut,where ut = P{xt)wt) E{utzt-i) = E(utut-i) = E(wtwt-i) = o, for i = 1, 2,. . . , t; and E(w2
t) = a2. Under this hypothesis, the first step in the estimationis to solve the minimization problem in expression (10). The absolute valuesof the estimated residuals from this equation then are regressed on various
' Minimization was conducted using the Time Series Processor (TSP) program. SeeHALL, HALL, and BECKETTI (1977).
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
rAUTOMOBILE INSURANCE CLAIM COSTS 95
functions of xt to determine P(xt). Finally, OLS is used to estimate the fol-lowing equation:
(U) y t h t - i = « ( 1 P ) + $(xtpxt-i) + ut
P(xt) P(xt) P(xt) P(xt)
where p = the estimate of p obtained in the first stage by solving the mini-mization problem in (io), and
P(xt) = the functional form of xt which provides the highest explanatorypower in regressions of the absolute residuals from (io).
Method 2. In this case, the error structure is assumed to be: zt =P{xt)\put-i + wi\, w h e r e E(ut-iWt) = E(wtwt-i) = o , i = 1 , 2 , ...,t; a n d
E(w2t) = a2. The estimation procedure is to obtain OLS estimates of a and [3,
define %t = yt — u— fi%t, and regress | et | on various functions of xt. Denotingthe best of these functions as P{xt), the final step is to minimize the followingexpression with respect to a, (3, and p:
( yP(xt) P(xi) P{xt)
Equations estimated through both methods were used in the forecastingexperiments with the choice between the two based on forecast accuracy.
Efficient estimation of equations with lagged dependent variables amongthe regressors when the residuals are autocorrelated or heteroskedastic is adifficult econometric problem requiring specialized computer programs notavailable to the authors. Accordingly, no adjustments were made whenequation (5) was characterized by these problems, and the OLS versions wereused in the forecasting experiments.
3. THE DATA AND THE FORECASTING EXPERIMENTS
The data used in estimating the forecasting equations are automobile bodilyinjury and property damage liability insurance average paid claim costsreported by the Insurance Services Office (ISO). The ISO is the leading ratingbureau in the United States, maintaining automobile insurance data for morethan one hundred contributing insurance companies. Pooled data from allcontributing companies are used by ISO to develop trend factors that affectthe rates of a subset of contributors (others price independently). A credibilityweighted average of state and national data is used in the trending processfor rates applicable to a particular state. The forecasts in this article are basedon the national averages.
The data period for automobile property damage liability claim costs isthe first quarter of 1954 (1954.1) to the second quarter of 1978 (1978.2). For
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
96 J. DAVID CUMMINS AND ALWYN POWELL
bodily injury liability, the period is 1964.1 to 1978.2. The property damagefigures are for total limits coverage, while the bodily injury data are for$ 10,000 basic limits coverage (i.e., claims in excess of $ 10,000 appear in theaverages as $ 10,000). Each series is obtained by dividing the total dollaramount of claim payments in a given quarter by the total number of claims inthe quarter. The claims cost series, converted to index form, are presentedin the Appendix 8.
Several price and wage indices were tested as independent variables. Amongthose with the highest explanatory power are the implicit price deflator forgross national product, the implicit price deflator for autos and parts, andwage indices for the service sector and for the private sector of the economy.Actual values of the deflators and indices are available in published U.S.government sources 9, while predicted values were obtained from WhartonEconometric Forecasting Associates (WEFA), one of the leading U.S. fore-casting firms 10.
The initial screening of predictor variables was based on summary regressionstatistics such as the coefficient of determination (i?2) u . After selecting thetwo or three best variables, subsequent evaluations of variables and specifica-tions were based on predictive accuracy. The approach taken in the experimentswas to forecast average paid claim costs for periods of eight quarters beginningat one-year intervals from 1971 through 1977 12. The third quarter of eachyear was arbitrarily selected as the initial forecast period. The goal in theexperiments was to reproduce as accurately as possible the results that wouldhave been obtained if the forecasting equations actually had been employedin the periods under consideration. Accordingly, the econometric models wereestimated using actual values of the dependent and independent variables forthe period ending in the quarter prior to the initial forecast period, and theforecasts were generated using predicted values of the independent variables.
8 Transforming data into indices affects the coefficients of any regressions based onthe transformed data but not the values of the essential test statistics. KMENTA (1971),PP- 377-379-
9 The full title of the autos and parts deflator is the implicit deflator, personal consump-tion expenditures, durables, autos and parts. It and the GNP deflator are available in theSurvey of Current Business (Washington, D.C.: U.S. Government Printing Office, monthly).The wage indices are compensation per man hour, private sector and compensation perman hour, service sector. They are constructed from series available in the Survey ofCurrent Business and in Employment and Earnings (Washington, D.C.: U.S. GovernmentPrinting Office, monthly).
10 WEFA forecasts are released quarterly. A summary version of the forecast is presentedeach quarter in The Wharton Magazine, which is published at the University of Penn-sylvania, Philadelphia, Pa.
11 i?2 is not a good indicator of predictive accuracy. However, it was considered ac-ceptable as an initial screening statistic.
12 Data availability limited the last forecast period to four quarters.
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
AUTOMOBILE INSURANCE CLAIM COSTS 97
The accuracy of the forecasts is determined by reference to the followingstatistics:
(13) TPCE = (ys~ yB
\x 100
MAPE = V l ^ ^ J x 100^-—i Vt
RMSPE = IOO x 1/ y {p.—
where TPCE = total predicted change error,MAPE = mean absolute percentage error,
RMSPE = root mean square percentage error,yt = the predicted value of average paid claim costs in quarter t
of the forecast period,yt = actual average paid claim costs in quarter t of the forecast
period, andyo = the actual value of average paid claim costs in the quarter
immediately preceding the forecast period.
The TPCE is the most important measure of forecasting accuracy forratemaking purposes because it measures the percentage by which predictedclaim costs exceed (fall short of) actual claim costs for the period to which thenew rates would apply.
Tests with the ordinary least squares versions of equations (3), (4), and (5)revealed little difference among alternative predictor variables in terms ofgoodness-of-fit or forecast accuracy. Since the private sector wage rate per-formed at least as well as the other variables during most forecast periods,it is used exclusively in the reported regressions. Equations with more thanone economic explanatory variable were not much better than the univariateequations in terms of goodness-of-fit and, in fact, often yielded unrealisticsigns for the coefficients because of multicollinearity. Accordingly, the resultspresented below are based on equations with one economic regressor. The bestforecasts were obtained with equations estimated over the maximum availableobservation period.
4. RESULTS: AUTOMOBILE PROPERTY DAMAGE LIABILITY PAID CLAIM COSTS
For automobile property damage liability paid claim costs, the OLS versionof equation (5) yielded forecasts which generally were more accurate than thosefrom the OLS versions of equations (3) and (4). Tests revealed that all threespecifications are subject to autocorrelation and heteroskedasticity in nearly
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
98 J. DAVID CUMMINS AND ALWYN POWELL
every estimation period 13. Since it was not feasible to adjust equation (5) forthese problems due to program limitations, the next-best OLS model, equation(3), was chosen for adjustment. The exponential trend equations and the OLSversion of equation (5) are presented in Table 1. Equation (3), adjusted forautocorrelation, is shown in the top panel of Table 2, while the bottom panelshows this specification adjusted for both autocorrelation and heteroskedasti-city according to Method 1 14. Dummy variables for seasonality are included inthe equations presented in Table 2 but not in the OLS version of equation (5)because preliminary results indicated that the inclusion of seasonal dummiesin the OLS equations leads to less accurate forecasts.
The forecast results for property damage liability paid claim costs areshown in Table 3. The table gives the TPCE, the MAPE, and the RMSPE forthe exponential trend model (ET), the OLS model with lagged endogenousregressor (LE), equation (3) adjusted for autocorrelation (AR), and the samespecification adjusted for both autocorrelation and heteroskedasticity (ARH).To provide a benchmark for analyzing the forecast errors, a shift index tomeasure unanticipated inflation has been included in the table. The index isthe ratio of the (quarterly) geometric mean rate of claims cost inflation duringeach forecast period to the geometric mean inflation rate during the precedingtwelve quarters.
In terms of TPCE, all three econometric models outperform the exponentialtrend model in five of the seven forecast periods. As expected, the performanceof the trend model is worst when unanticipated inflation is highest. Amongthe econometric models, those adjusted for both autocorrelation and heteroske-dasticity do best during periods characterized by high unanticipated inflation(e.g., the forecast periods beginning in 1973.4, 1974.4, and 1975.4). The OLSmodel also does quite well during these periods, suggesting its use as a rela-tively simple replacement for the exponential trend model.
5. RESULTS: AUTOMOBILE BODILY INJURY LIABILITY PAID CLAIM COSTS
The exponential time trend and best OLS equations for automobile bodilyinjury liability claim costs are presented in Table 4. In this case, equation (3)is the best OLS model. The hypothesis of homoskedasticity is rejected for thismodel during the last five estimation periods. The Durbin-Watson test in-dicates the presence of autocorrelation during the last four periods and givesinconclusive results for the period ending in 1973.2. Since the behavior ofthe test statistics when autocorrelation and heteroskedasticity are presentsimultaneously is not clear, the equations for all periods were reestimated
13 All significance test results reported in this article are at the 5 percent level.14 For the first two sample periods, P{xt) = y0 + yj%t '• while for the last five periods
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
TA
BL
E 1
: E
XPO
NE
NT
IAL
T
RE
ND
A
ND
LA
GG
ED
EN
DO
GE
NO
US
PRE
DIC
TIO
N
EQ
UA
TIO
NS
FO
R A
UT
OM
OB
ILE
PR
OPE
RT
Y
DA
MA
GE
LIA
BIL
ITY
AV
ER
AG
E P
AID
CLA
IM C
OST
S
Equ
atio
n fo
rFo
reca
stB
egin
ning
in
1971
-3
1972
-3
1973
-3
1974
-3
1975
-3
1976
-3
1977
-3
lny t
a
5-39
35(1
822.4
4)
5-50
68(1
065.1
2)
5-6
113
(84
6.1
48
)
5-67
56(1
323-
17)
5-69
97(8
28.7
77)
5-74
O9
(721.4
99)
5.8
09
9
(170
5-41
)
- &+u
0.0
236
(58.
740)
0.0
194
(27-
675)
0.0
132
(14.
656)
0.0
116
(19-
905)
0.0
16
3(1
7.42
6)
0.0
220
(20.
314)
0.0
25
4
•997
.987
•956
•975
.968
.976
•997
Vt
a
-11
.89
94
(-2
.50
6)
-11
.13
40
(-2
.19
8)
-10
.92
35
(-2
-34
9)
-5-8
70
5(-
1.3
66
)
-9.4
80
7(-
2.8
98)
-8.5
85
5(-
2.9
82)
-9-9
91
3(-
3-5
33
)
=
a +
$xw
t
k11
.506
2(2
.136
)
12.5
708
(2.2
20)
12.5
643
(2.3
26)
8.0
022
(1-5
29)
12.3
019
(2.9
04)
11-5
338
(2.9
24)
13-2
348
(3.2
80)
.885
015
(12.
793)
.860
157
(n.9
92
)
.858
908
(12.3
35)
•908
715
(13-
119)
•854
314
(14.
685)
.862
572
(15.
620)
.84
07
51
(14-
754)
R*
•995
•995
.996
.996
•996
•997
•997
-6-7
5
-3-9
6
-4.2
1
-4.6
7
-3-3
1
-3-5
O
-4.0
4
GQ
*
2-54
(29.
29)
4-52
(31.
31)
4.98
(33.
33)
6.86
(35.
35)
7.2
6
(37.
37)
6.63
(39.
39)
6.97
(41.4
1)
c,TOMO BILE INSURi\NCE CLAIM
Key
: yt
=
aver
age
pai
d c
laim
cost
s in
qu
art
er
t; y
t =
av
erag
e pai
d c
laim
cost
s fo
r th
e y
ear
endin
g i
n q
uart
er
t; t
=
ti
me,
in q
uar-
ters
; an
d w
t =
p
riv
ate
sec
tor
wag
es p
er
man
-hour
in q
uart
er
t.N
ote
: T
he
est
imat
ion f
or
each
tim
e tr
en
d e
quat
ion i
s th
e tw
elve
qu
art
ers
pre
cedin
g t
he
fir
st f
ore
cast
per
iod.
Fo
r ea
ch g
eom
etri
c la
geq
uat
ion
, th
e e
stim
atio
n p
erio
d b
egin
s in
195
4. l a
nd
ends
in th
e q
uart
er
pre
cedin
g t
he
firs
t fo
reca
st p
erio
d.
Th
e n
um
ber
s in
par
enth
eses
bel
ow
th
e e
stim
ated
coef
fici
ents
are
-s
tati
stic
s.
a T
his
colu
mn
pre
sen
ts r
esu
lts
of
a t
est
su
gges
ted b
y D
UR
BIN
(1
970)
for
auto
corr
elat
ion in
model
s w
ith l
agged
en
dogen
ous
ex-
pla
nato
ry v
aria
ble
s. D
urb
in s
ugges
ts t
wo
test
s, o
ne
bas
ed o
n a
sta
tist
ic h
, an
d t
he
oth
er b
ased
on
a r
egre
ssio
n i
nvolv
ing t
he
res
idual
sof
th
e O
LS
equat
ion
bei
ng
test
ed
. T
he
latt
er
test
is
use
d h
ere.
Sin
ce i
t is
a £
-tes
t w
ith m
ore
th
an 3
0 d
egre
es o
f fr
eedom
, th
e s
tan
dard
no
rmal
tab
le c
an
be
use
d t
o t
est
sig
nif
ican
ce.
At
the
5 p
erce
nt
level
, th
e hypoth
esis
of
no a
uto
corr
elat
ion i
s re
ject
ed f
or
ever
y e
quat
ion
show
n i
n t
his
sec
tion o
f th
e t
ab
le.
t> T
his
colu
mn
pre
sen
ts t
he
GO
LD
FEL
D-Q
UA
ND
T
(196
5) t
est
st
ati
stic
s fo
r h
eter
osk
edas
tici
ty.
Th
e h
ypoth
esis
of
hom
osk
edas
tici
tyis
rej
ecte
d i
f th
e st
ati
stic
valu
e ex
ceed
s F
a,V
tlV
2.
Th
e deg
rees
of
free
dom
para
mete
rs,
v x a
nd
v2 a
pp
ear
in p
aren
thes
es b
elow
th
e st
ati
stic
.A
t th
e 5
per
cen
t le
vel
, th
e h
yp
oth
esis
is
reje
cted
for
ever
y e
quat
ion s
how
n i
n t
his
sec
tion o
f th
e tab
le.
o o w o
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
100 J. DAVID CUMMINS AND ALWYN POWELL
TABLE 2 : AUTOREGRESSIVE EQUATIONS FOR AUTOMOBILE PROPERTY DAMAGE LIABILITY
AVERAGE PAID CLAIM COSTS:
EQUATIONS IN LOWER PANEL ADJUSTED FOR HETEROSKEDASTICITY
Equation forForecastBeginning in
1971-3
1972.3
1973-3
1974-3
1975-3
1976.3
1977-3
1971-3
1972-3
1973-3
1974-3
1975-3
1976.3
1977-3
yt
85.6807(13.780)
77.2711(21.109)
76.0740(23-796)
72.2098
(29-452)
72.7631
(31.229)
72.4086
(35-250)
73-4537(35-232)
89.6291
(37-599)
77.9910(33-828)
75-5438(38.015)
71.0324
(43-322)
72.0044
(45-473)
72.2156
(49-337)
72.7016
(51-657)
= p*-1 + *(.
Yi
— 9.3200(-11.828)
— 9.8140(-10.518)
—10.2579
(-10.725)
— 10.6073
(-9-39O)
— 10.8620( — 10.101)
— 10.7907(-10.588)
-11-5857(-11.108)
-14.1917(-17.047)
-10.7988
(-10.874)
— 9.4062(-11.164)
-8.3723(-10.736)
-9-3375(—10.980)
-10.1239
(-H-175)
-10.3194
(-11-635)
-P) + P K -
Ya
-2.3364(—2.960)
-2-7595(-2.940)
-2.7396
(-2.835)
-3.0991(-2.688)
-35125(-3-196)
-3-511O(-3-359)
-3.6506(-3-396)
— 0.7009(-0.879)
-2.6878
(-2.778)
-2.8493(-3-498)
-3.1814(-4.244)
-3-3O19(-4.012)
-33716(-3-826)
-3.4609(-4.016)
M + fA-
Ya
-5-9115(-7.464)
-5-8231(-6.211)
-5-5684
(-5784)
-5.6290
(-4-938)
-57196(-5-238)
-5.6970
(-5.489)
-6.1675
(-5-77O)
-6.4039
(-7.4297)
-5-7983(-5-938)
-5-7251(-6.979)
-5-9O34(-7.786)
-5-9526
(-7-179)
-58504(-6.614)
-6.1045(-7.050)
HYaSa + Y3S3
P
•950495(64-735)
.924875
(50.785)
.918491(49.380)
.897217
(38.963)
.895484
(39-829)
•893974
(41-125)
.892990
(40.288)
•95O495
•924875
.918491
.897217
.895484
•893974
.892990
R*
•9983
.9980
•9982
•9977
.9981
.9985
.9987
•9751
.8983
•8030
.5889
•7483
•8724
.8612
Key: yt = average paid claim costs in quarter t; wt — private sector wages per man-hour in quarter t; and St = 1.0 in quarter i of each year, o otherwise.Note: The estimation period for each equation begins in 1954.1 and ends in the quarterpreceding the first forecast period. The number in parentheses below each estimatedcoefficient is the ratio of the coefficient to its estimated standard deviation. The constantterm a was estimated in an earlier version of the equations but was found to be insig-nificant.
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
TA
BL
E 3
: R
ESU
LT
S O
F E
IGH
T Q
UA
RT
ER
FO
RE
CA
STS
OF
AU
TO
MO
BIL
E P
RO
PER
TY
DA
MA
GE
LIA
BIL
ITY
AV
ER
AG
E P
AID
CLA
IM C
OST
S
Fir
st P
erio
dof
For
ecas
t
1971
-3
1972
-3
1973
-3
1974
-3
1975
-3
1976
.3
1977
-3 °
Abs
olut
e A
vera
ges
ET
To
tal
Pre
dict
edC
hang
e E
rro
rs a
LE
AR
10.8
%
5.4
%
3.7
%
5-4
-10
.3
-15
-6
-8-7
-6.9
-1.6 8-
5
7-3
1.8
-2.4 1.6
-4-3
-3-O
%
3-7°
/
6.0 o-3
0.6 °-5
-5°
-3-4
D
2.8
%
AR
H
4-3%
6-3
0 0 0
-5°
-3-6 2-
7%
Mea
n A
bsol
ute
Per
cent
. E
rro
rs a
ET
6.0%
5-7
1.1
5-6
5-6
2.9
1.2 4-o%
LE
3-8%
2.4 3°
2-9 i-7
1-9
1-9
2-5/
0
AR
2-5%
1.9
2.1 1.2
1.8
2.0
2.0 1-9
/o
AR
H
2-4%
1-9
2.0 1-4
1-5
2.1
2.2 1-9%
Roo
t M
ean
Squa
rePe
rcen
t. E
rror
s a
ET
6-9%
5-9
i-7
6.4
5-9
3-2 1.2 4-5%
LE
4-4%
3-2
3-7
3-1 1.9
2.4
2.2 3-o%
AR
2-7%
2.4
2.9 1.6
2.0
2-5
2-3 2.3%
AR
H
2-7%
2-5
2.9 1.8
1.8
2-5 2-5 2-4%
Shi
ftIn
dex
t>
51
%
61
12
6
20
1
15
0
13
8
11
5
C O y 0 0 w
Key
: E
T
=
expo
nent
ial
tren
d m
odel
, L
E
=
econ
omet
ric
mod
el w
ith
lagg
ed
endo
geno
us r
egre
ssor
, A
R
=
auto
regr
essi
ve m
odel
wit
ho
ut
hete
rosk
edas
tici
ty a
djus
tmen
t, A
RH
=
au
tore
gres
sive
mod
el w
ith
hete
rosk
edas
tici
ty
adju
stm
ent.
No
te:
The
eco
nom
etri
c fo
reca
sts
(col
umns
lab
elle
d L
E,
AR
, an
d A
RH
) ar
e ba
sed
on p
redi
cted
val
ues
of t
he
pri
vat
e se
ctor
wag
e ra
tepr
omul
gate
d b
y W
har
ton
Eco
nom
etri
c F
orec
asti
ng A
ssoc
iate
s, I
nc.
The
equ
atio
ns w
ere
esti
mat
ed u
sing
rev
ised
val
ues
of t
he
wag
era
te s
erie
s, w
hich
dif
fer
slig
htly
in
mos
t ca
ses
from
th
e pu
blis
hed
valu
es a
vail
able
at
the
tim
e th
e W
har
ton
fore
cast
s w
ere
mad
e. T
hus,
the
pred
icte
d va
lue
of t
he
wag
e ra
te,
wt,
for
use
in t
he
clai
m c
ost
fore
cast
s is
wt
= w
oc .
u>
twl^
ov>
whe
re w
oc
= t
he
curr
ent
(rev
ised
)va
lue
of w
for
the
peri
od p
rece
ding
th
e fi
rst
fore
cast
per
iod,
wop
=
th
e (p
reli
min
ary)
val
ue o
f w
for
the
peri
od p
rece
ding
th
e fi
rst
fore
-ca
st p
erio
d at
th
e ti
me
the
Wh
arto
n fo
reca
st w
as m
ade,
and
wt w
=
th
e W
har
ton
fore
cast
of
w f
or p
erio
d t.
a T
ota
l pr
edic
ted
chan
ge e
rror
(T
PC
E),
mea
n ab
solu
te p
erce
ntag
e er
ror
(MA
PE
), a
nd r
oo
t m
ean
squa
re p
erce
ntag
e er
ror
(RM
SP
E)
are
defi
ned
in e
quat
ions
(13
), (
14),
and
(15
) in
th
e te
xt.
b T
he s
hift
ind
ex i
s co
mpu
ted
as f
ollo
ws:
/
=
[(yg
/yo)
1 '8 — 1
] -H
[{y
ajy-
i 2)1/
12—
1].
whe
re y
8 =
act
ual
ave
rage
pai
d cl
aim
co
sts
inth
e la
st q
uar
ter
of t
he
fore
cast
per
iod,
y0
=
actu
al c
laim
cos
ts i
n th
e q
uar
ter
prec
edin
g th
e fo
reca
st p
erio
d, a
nd y
. 12
=
act
ual
cla
imco
sts
thir
teen
qu
arte
rs p
rio
r to
th
e fo
reca
st p
erio
d.c
Thi
s fo
reca
st p
erio
d ru
ns
from
19
77.3
th
rou
gh
1978
.2.
n > o o 3
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
TA
BL
E 4
: E
XPO
NE
NT
IAL
TR
EN
D A
ND
LIN
EA
R
OL
S P
RE
DIC
TIO
N
EQ
UA
TIO
NS
FOR
AU
TO
MO
BIL
E B
OD
ILY
IN
JUR
Y L
IAB
ILIT
Y
AV
ER
AG
E P
AID
CLA
IM C
OST
S
Ke
y:
yt
=
aver
age
pai
d cl
aim
co
sts
in q
ua
rte
r t;
yt
=
aver
age
pai
d cl
aim
co
sts
for
the
yea
r en
din
g in
qu
art
er
t;t=
ti
me,
in
qu
art
ers
;a
nd
wt
=
pri
va
te s
ecto
r w
ages
per
man
-ho
ur
in q
ua
rte
r t.
No
te:
Th
e es
tim
atio
n p
erio
d fo
r ea
ch t
ime
tre
nd
equ
atio
n is
th
e tw
elv
e q
ua
rte
rs p
rece
din
g th
e fi
rst
fore
cast
p
erio
d.
Fo
r ea
ch
lin
ear
eco
no
met
ric
equ
atio
n,
the
esti
mat
ion
per
iod
beg
ins
in
1964
.1 a
nd
end
s in
th
e q
ua
rte
r p
rece
din
g th
e fi
rst
fore
cast
per
iod
. T
he
nu
mb
ers
in p
aren
thes
es b
elo
w t
he
esti
mat
ed
coef
fici
ents
ar
e ^-
stat
isti
cs.
a N
um
be
rs
in
this
co
lum
n ar
e D
urb
in-W
atso
n d
stat
isti
cs.
Rej
ecti
on
of t
he
hy
po
thes
is
of
no
seri
al c
orr
elat
ion
(at
the
5 p
erce
nt
lev
el)
is i
nd
icat
ed
by
an a
ster
isk
. T
he
test
fo
r 19
73.3
is
inco
ncl
usi
ve.
b T
his
co
lum
n p
rese
nts
th
e G
OL
DF
EL
D-Q
UA
ND
T (1
965)
tes
t st
atis
tics
fo
r h
eter
osk
edas
tici
ty.
Th
e h
yp
oth
esis
of
h
om
osk
edas
tici
tyis
rej
ecte
d if
th
e st
atis
tic
val
ue
exce
eds
Fai
Vl, V
2.
Th
e d
egre
es o
f fr
eed
om
p
aram
eter
s, v
x an
d v 2
, ar
e g
iven
in
par
enth
eses
bel
ow
th
e st
a-ti
stic
s. A
t th
e 5
per
cen
t le
vel
th
e h
yp
oth
esis
is
reje
cted
fo
r ea
ch o
f th
e la
st f
ou
r eq
uat
ion
s (1
97
4.3
th
rou
gh
19
77
.3).
o to
Eq
ua
tio
n fo
rF
ore
cast
Beg
inn
ing
in
1971
-3
1972
-3
1973
-3
1974
-3
1975
-3
1976
-3
1977
-3
a
6-99
43(1
868.
45)
7-O
458
(844
-845
)
7-13
O3
(5O
5-19
3)
7.19
64(6
03.5
96)
7-19
37(1
210.9
5)
7-2
514
(112
5-23
)
7-33
°4(1
162.
57)
lny t
=
a
+ (3
/
P
0.0
14
7(2
8.93
8)
0.0
171
(15.
070)
0.0
116
(6-0
53)
0.0
09
4(5
-772
)
0.0
174
(21-
525)
0.0
200
(22.
812)
0.0
202
(23-
549)
R* .988
.958
.786
.769
•979
.981
.982
yt
a
- 14
7603
(-2
.92
9)
-85
-92
3(-
1.8
84
)
— 1
4.90
8(-
0.3
60
)
27.2
017
(0.7
12)
104.4
25
(2.9
07)
107.
879
(3-5
90)
112.
189
(4-3
47)
= i +
fo
34
0.0
42
(24-
356)
321.8
92
(26.
432)
30
1.2
99
(28.
304)
28
9.3
16
(30.
657)
26
8.3
13
(31-
695)
26
7.4
02
(39
-6i8
)
26
6.2
64
(48.
278)
i?2
•955
• 95
6
•957
•959
.958
.970
•978
DW
*
2.0
88
1.82
7
1.46
1
1-35
6*
1.2
08*
1-25
5*
1.3
09*
GQ
b
1-35
(10,
10)
1-93
(12,
12)
2.29
(H.1
4)
3-25
(16,1
6)
2-93
(18,
18)
3-37
(20,2
0)
2.67
(22,2
2)
d 0 r>rMMIN en t-1 ALW1 O <
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
AUTOMOBILE INSURANCE CLAIM COSTS 103
TABLE 5 : AUTOREGRESSIVE EQUATIONS FOR AUTOMOBILE BODILY INJURY LIABILITYAVERAGE PAID CLAIM COSTS:
EQUATIONS IN LOWER PANEL ADJUSTED FOR HETEROSKEDASTICITY
Equation forForecast —Beginning in
1971-3
1972-3
1973-3
1974-3
1975-3
1976-3
1977-3
1971-3
1972-3
1973-3
1974-3
1975-3
19763
1977-3
a
— 110.946(-1-559)
N.E.
N.E.
127.914(1-556)
288.274(2-435)
235-516(3.021)
221.158(3-538)
- 142-4O5(-2.253)
N.E.
N.E.
106.858(1-252)
234.868
231-793(2.849)
220.898
= PW
344-294(17-381)
313-435(41-365)
313.612(39.736)
284.169(16.527)
252.929(12.912)
260.034(19.667)
263.160(25.981)
353-898(18.752)
317-833(35-971)
317-593(37-473)
290.600
(15-467)
264.464(11.848)
261.226(18.589)
262.653(29-542)
i ( i - * ) + &(«A
Yi
-41-1317(-3-O37)
— 40.3019
(-2-773)
-41.3901(-3-070)
-38.1477(—2.902)
-45.5498(-3-449)
-46.9951(-3.621)
-49.726(-4.069)
-38.1636(-3-239)
-42.1668(-3-578)
-42.2427(-3-789)
— 40.8001(-3-276)
-45-5516(-3-6655)
— 46.6890(-3-645)
— 51.6890(-4.190)
Y2
N.E.
N.E.
N.E.
N.E.
N.E.
N.E.
N.E.
N.E.
N.E.
N.E.
N.E.
N.E.
N.E.
N.E.
t) + Tx5l + Yr
Ys
-74.0772(-4-774)-70.9714(-4-473)
-70.3248(-5-O15)
-81.546(-5.908)
-90.9309(-6.458)
-83.5657(-6.206)
-84.430(-6.732)
— 79.2642(—6.201)
-84.5401(-7.021)
-83.5804(-7-436)
-82.9638(-6.396)
— 89.5063(-6.896)
-84.1028
(-6-343)
-83.8710(-6.609)
S2 + Y3S3
P
.440689(2.145)
•527511(3-129)
.586522(4.208)
.620412
(5-oi5)
•712105(6.303)
•654109
(5-897)
•638744(6.030)
.414891(2.464)
•56165(4-371)
•57197(4-8i5)
.63102(5-i69)
.71128(6.390)
.66127(6.016)
•61937(5.717)
i?2
•9763
.9711
.9767
.9807
.9822
.9862
.9901
.966
•969
•978
.872
.852
.980
•995
Key: yt = average paid claim costs in quarter t; wt = private sector wages per man-hour in quarter t; and S( = 1.0 in quarter i of each year, o otherwise.
Note: The estimation period for each equation begins in 1964.1 and ends in the quarterpreceding the first forecast period. The number in parentheses below each estimatedcoefficient is the ratio of the coefficient to its estimated standard deviation. N.E. = foundto be insignificant in earlier versions of the equations and thus omitted from the finalversion.
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
TA
BL
E 6
: R
ES
UL
TS
O
F E
IGH
T Q
UA
RT
ER
F
OR
EC
AS
TS
O
F A
UT
OM
OB
ILE
B
OD
ILY
IN
JUR
Y L
IAB
ILIT
Y A
VE
RA
GE
PA
ID C
LA
IM C
OS
TS
Fir
st P
erio
dof
For
ecas
t
1971
-3
1972
.3
1973
-3
1974
-3
1975
-3
1976
.319
77.3
c
Abs
olut
e A
vera
ges
Tot
al P
redi
cted
Cha
nge
Err
ors a
ET
10
.1%
-8.4
-4.6
-4.6
-5-o 2.4
3-9
5-6%
OL
S A
R
3-4%
5
-6%
0.1
0.4
3.2
6.1
5-3
6.3
0.7
-0
.4
2-5
2.8
i-5
2.6
-7
A 0
/ 0
=0
/
2-4
/o
3-5
/o
AR
H
7-o%
1.6
7.6
7-3
1-4
2.9
2-5
4-3%
ET
Mea
n A
bsol
ute
Per
cent
. E
rror
s
OL
S
2.6%
3-7
%
5-9
2.2 6.8
2.2
0.5
0.8
4.6
4-5
5-3
2.4
2.0 1.8
2 n
0/
1
c0/
3-%
3
-5/0
AR
4-3%
2.4 4-3
4.6 1-5
1.2 i-7
2.9
%
a
AR
H
5-2%
2.7
5-1
5-3
1.9
1.2 1.6
3-3%
Roo
t M
ean
Squ
are
Per
cent
. E
rror
s a
ET
2.8
%
6.0
2.6 7.2
2.4
0.7 1.1 3-3%
OL
S
4-5%
5-7
5-4
5-6
2.8
2.4
2.2 4-1
%
AR
5-o%
3-4
5°
5-i
2.0 1.6
2.0 3-4%
AR
H
6.0
%
3-6
5-7
5-6
2.2 1.6
2.0 3-8%
Shi
ftIn
dex
*>
63
%9
0
12
6
16
1
17
1 85 80
(—i d % d d g w d
Key
: E
T =
ex
pone
ntia
l tr
end
mod
el,
OL
S =
or
dina
ry l
east
squ
ares
mod
el,
AR
=
aut
oreg
ress
ive
mod
el w
itho
ut
hete
rosk
edas
tici
tyad
just
men
t,
AR
H
=
auto
regr
essi
ve m
odel
wit
h he
tero
sked
asti
city
ad
just
men
t.
No
te:
The
eco
nom
etri
c fo
reca
sts
(col
umns
lab
elle
d O
LS
, AR
, an
d A
RH
) ar
e ba
sed
on p
redi
cted
val
ues
of t
he
priv
ate
sect
or w
age
rate
prom
ulga
ted
by
Wh
arto
n E
cono
met
ric
For
ecas
ting
Ass
ocia
tes,
Inc
. T
he e
quat
ions
wer
e es
tim
ated
usi
ng r
evis
ed v
alue
s of
the
wag
era
te s
erie
s, w
hich
dif
fer
slig
htly
in
mos
t ca
ses
from
th
e pu
blis
hed
valu
es a
vail
able
at t
he
tim
e th
e W
har
ton
fore
cast
s w
ere
mad
e. T
hus,
the
pred
icte
d va
lue
of
the
wag
e ra
te w
t, fo
r us
e in
the
clai
m c
ost
fore
cast
s is
wt
= w
oc •
wtw
lwop
> w
here
woc =
th
e cu
rren
t (r
evis
ed)
valu
e of
w f
or t
he
peri
od p
rece
ding
th
e fi
rst
fore
cast
per
iod,
wop =
th
e (p
reli
min
ary)
val
ue o
f w
for
the
peri
od p
rece
ding
th
e fi
rst
fore
cast
per
iod
at
the
tim
e th
e W
har
ton
fore
cast
was
mad
e, a
nd w
t w =
th
e W
har
ton
fore
cast
of
w fo
r pe
riod
t.
a T
otal
pre
dict
ed c
hang
e er
ror
(TP
CE
), m
ean
abso
lute
per
cent
age
erro
r (M
AP
E),
and
roo
t m
ean
squa
re p
erce
ntag
e er
ror
(RM
SP
E)
are
defi
ned
in e
quat
ions
(13
), (
14),
and
(15
) in
the
tex
t.b
The
shi
ft i
ndex
is
com
pute
d as
foll
ows:
I
=
[(y s
/y0)
1/s
— 1
] -r-
\_{
yjy
-n)lin
—
1].
whe
re y
s =
act
ual
aver
age
paid
cla
im c
osts
in
the
last
qu
arte
r of
the
fore
cast
per
iod,
y0 =
ac
tual
cla
im c
osts
in
the
qu
arte
r pr
eced
ing
the
fore
cast
per
iod,
and
;y_
12 =
act
ual
clai
mco
sts
thir
teen
qua
rter
s pr
ior
to t
he
fore
cast
per
iod.
c T
his
fore
cast
per
iod
runs
fro
m
1977
.3 t
hrou
gh 1
978.
2.
O
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
AUTOMOBILE INSURANCE CLAIM COSTS 105
with adjustment for autocorrelation and for both autocorrelation and hetero-skedasticity, using Method 2 15. These results are shown in Table 5.
The forecasting errors for the four sets of bodily injury equations are pre-sented in Table 6. Using the TPCE as the measure of accuracy, the OLSequations give the best results in five of the seven forecast periods and theexponential trend model is superior in two. In one of these periods, however,the trend line underpredicts while the econometric models overpredict. In theother, the trend line is better than the OLS equation by only o.l percentagepoint. When MAPE and RMSPE are used to gauge forecasting accuracy, theexponential trend equation is superior to the econometric equations in four ofthe seven periods. However, two of the periods when at least one of the econ-ometric equations is better are those with the highest unanticipated inflation.
6. SUMMARY AND CONCLUSIONS
The results of this study indicate that econometric models can be used to fore-cast automobile insurance paid claim costs more accurately than the exponentialtrend models now in use in the United States. The superiority of the econometricmodels is especially evident during periods of high unanticipated inflation.Although adjustments for seasonality, autocorrelation, and heteroskedasticityoften improve forecasting accuracy, even the OLS models perform quite well.Adoption of the econometric approach may enable insurance companies toeliminate some of the uncertainty associated with claims cost inflation.
REFERENCES
COOK, C. (1970). Trend and Loss Development Factors, Proceedings of the CasualtyActuarial Society, 52, 1-14.
DURBIN, J. (May 1970). Testing for Serial Correlation in Least-Squares Regression WhenSome of the Regressors Are Lagged Dependent Variables, Econometrica, 38, no. 3,410-421.
GLEJSER, H. (1969). A New Test for Heteroskedasticity, Journal of the American Statis-tical Association, 64, 316-323.
GOLDFELD, S. M. and R. E. QUANDT. (1965). Some Tests for Homoskedasticity, Journalof the American Statistical Association, 60, 539-547.
HALL, B., HALL, R., and S. BECKETTI. (1977). Time Series Processor, Version 3.5, Pri-vately published. Stanford, Calif.
HARRISON, M. J. and B. P. M. MCCABE. (1979). A Test for Heteroskedasticity Based onOrdinary Least Squares Residuals, Journal of the American Statistical Association,74, 494-499-
HARVEY, A. C. and G. D. A. PHILLIPS. (1974). A Comparison of the Power of Some Tests forHeteroskedasticity in the General Linear Model, Journal of Econometrics, 2, 307-316.
KLEIN, L. R. and E. BURMEISTER. (1976). Econometric Model Performance: Comparative Si-mulation Studies of the U.S. Economy, University of Pennsylvania Press, Philadelphia.
KMENTA, J. (1971). Elements of Econometrics, The Macmillan Co., New York.SPENCER, B. G. (1975). The Small Sample Bias of Durbin's Test for Serial Correlation,
Journal of Econometrics, 3, 249-254.WAELBROECK, J., ed. (1976). The Models of Project LINK, North-Holland Publishing Co.,
Amsterdam.15 For the first three sample periods, P{xt) = ~X\x\\ while for the last four periods
H*t) = T0+7V4
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,
AP
PE
ND
IX:
PRIN
CIP
AL
DA
TA S
ER
IES
Qua
rter
1954-1
1954-2
1954
-319
54-4
1955-1
1955-2
1955
-319
55-4
1956.1
1956.2
1956
-319
56.4
1957-1
1957-2
1957
-319
57-4
19
58
11958.2
1958
-319
58.4
1959-1
1959-2
1959
-319
59-4
1960.1
1960.2
1960.3
1960.4
1961.1
1961.2
1961.3
1961.4
1962.1
1962.2
PD
1.0
0.9
90
30.9
709
1.0
097
1.0
097
1.0
291
1.0
194
1.0
680
1.0
680
1.0
777
1.0
777
1.1
456
1-1
553
1.1
845
1.1
845
1.2
427
1.2
330
1.2
427
1.2
427
1.2
718
1.2
524
1.2
913
1.2
816
1-33
981.
3204
1.3
204
1-33
O1
1.36
891.3
010
1-34
951-3
592
1.4
078
1.36
891.
3981
WR
1-95
1.96
1.98
2.0
2.0
12.0
3
2.0
62.0
8
2.1
2
2-1
52.1
92.2
22.2
52.2
8
2.3
1
2-33
2-35
2-37
2.4
0
2.4
22.4
42.4
62.4
8
2-5
1
2-55
2.5
8
2-5
92.6
22.6
3
2.6
5
2.67
2.6
9
2-73
2-75
Qua
rter
1962.3
1962.4
1963.1
1963.2
1963
319
634
1964.1
1964.2
1964
-31964.4
1965.1
1965.2
1965
319
65.4
1966.1
1966.2
1966
.319
66.4
1967.1
1967.2
1967
319
67.4
1968.1
1968.2
1968
.319
68.4
1969.1
1969.2
1969
319
69.4
1970.1
1970.2
1970.3
1970.4
PD
1.4
078
1.44
661.4
272
1.4
660
1-47
571-5
146
1-48
541-
5437
1-55
341.6
214
1.6
117
1.67
961.6
990
1.7
961
1.78
641.
8447
1.84
471-9
515
1.9
417
1.9
903
2.0
291
2.1
068
2.1
262
2.2
136
2.2
330
2.3
592
2-3
592
2.4
466
2.4
951
2.6
117
2.5
631
2.67
962-
7379
2.8
350
BI
1.0
1-0
341
0.9
330
1.0
364
1.0
4
1.0
811
1.0
129
1.0
999
1.0
764
1.1
340
1.1
011
1-1
915
1.2
045
1.2
468
1-2
233
1.2
667
1.2
785
1.3
619
1.2
808
1.3
690
1359
61.4
101
1.4
148
1.4
160
1.43
481.5
018
1.4
606
1.5
018
WR
2.7
6
2.7
92.8
22.8
4
2.8
62.9
0
2.94
72.9
72
3.0
10
3.0
26
3.0
40
3.06
73.1
06
3-1
41
3-18
53-2
41
3.2
92
3-33
13-
384
3-43
33-
472
3-51
53.6
09
3.66
93.7
21
3.78
93-8
42
3-9O
43
.96
64-
°35
4.0
98
4.16
84-
253
4.30
6
Qu
arte
r
1971-1
19
71
2
1971-3
1971.4
1972.1
1972.2
1972-3
1972.4
1973-1
1973-2
1973
-319
73-4
1974-1
1974-2
1974
-319
74-4
1975-1
1975-2
1975
-319
75-4
1976.1
1976.2
1976
-319
76.4
1977-1
1977-2
1977
-319
77-4
1978.1
1978.2
PD
2.78
642.9
029
2-9
515
3.0
291
2.9
417
2.9
320
3-o
i94
3-1
165
3.08
743-
i845
3-18
453-
3495
3-23
3°3.2
621
3.40
783.6
117
3.67
963-
7476
3.80
58
;3-
9417
3.98
064.
0874
4.18
45
:4-
4175
4.40
78
:4-5
922
4-75
734.9
612
4-8
83
55-
1456
BI
•527
6•7
O74
1-6
273
1.6
592
L.6
228
•578i
•585
2.6
804
1.7
004
1-74
741-
6439
1-77
O9
1.86
961-9
389
1.8
284
1.9
130
1.88
481-
9965
2.03
64
2-16
332.
1293
2.17
742.
1598
2.29
962.
2985
2-39
252.
3678
2.4
618
2.3
960
2.4
982
WR
4-39
74.
468
4-54
O4-
57O
4.6
79
4-73
74-
797
4.85
85.0
06
5.0
90
5-17
95.2
72
5-38
45-
538
5-67
55-
837
5-99
96.
097
6.1
91
6.30
56-
439
6.57
86.
707
6.8
42
7.0
07
7.1
18
7.2
40
7-35
57613
7-75
8
O ON
'—,
> <; 5 0 a a g s; w > 0 > t-1 < z 0 W f
Key
: BI
and P
D =
aut
omob
ile
bodi
ly i
njur
y an
d p
rop
erty
dam
age
liab
ilit
y pa
id c
laim
cos
ts,
resp
ecti
vely
. The
se s
erie
s ar
e ex
pres
sed
as i
ndic
es w
ith
the
earl
iest
per
iod
= 1.0
. WR
= t
he
pri
vat
e se
ctor
wag
e ra
te.
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0515036100006681Downloaded from https://www.cambridge.org/core. IP address: 54.39.106.173, on 09 Jun 2020 at 11:54:38, subject to the Cambridge Core terms of use,