Journal of Forecasting, Vol.

12, 13-35 (1993)

Forecasting Time Series with Outliers

CHUNG CHEN
Syracuse University, Syracuse, N Y, USA

LON-MU LIU
The University of Illinois at Chicago, Chicago, IL, USA

ABSTRACT
Time-series data are often contaminated with outliers due to the influence
of unusual and non-repetitive events. Forecast accuracy in such situations
is reduced due to (1) a carry-over effect of the outlier on the point forecast
and (2) a bias in the estimates of model parameters. Hillmer (1984) and
Ledolter (1989) studied the effect of additive outliers o n forecasts. It was
found that forecast intervals are quite sensitive to additive outliers, but
that point forecasts are largely unaffected unless the outlier occurs near the
forecast origin. In such a situation the carry-over effect of the outlier can
be quite substantial. In this study, we investigate the issues of forecasting
when outliers occur near or at the forecast origin. We propose a strategy
which first estimates the model parameters and outlier effects using &he
procedure of Chen and Liu (1993) to reduce the bias in the parameter
estimates, and then uses a lower critical value to detect outliers near the
forecast origin in the forecasting stage. One aspect of this study is on the
carry-over effects of outliers on forecasts. Four types of outliers are
considered: innovational outlier, additive outlier, temporary change, and
level shift. The effects due to a misidentification of an outlier type are
examined. The performance of the outlier detection procedure is studied
for cases where outliers are near the end of the series. In such cases, we
demonstrate that statistical procedures may not be able to effectively
determine the outlier types due to insufficient information. Some strategies
are recommended to reduce potential difficulties caused by incorrectly
detected outlier types. These findings may serve as a justification for
forecasting in conjunction with judgment. Two real examples are
employed to illustrate the issues discussed.

KEY WORDS Misidentification Forecast errors Innovation outlier

Additive outIier Temporary change Level shift
Outlier test statistics Critical value

INTRODUCTION

Time-series data are often contaminated with outliers. The impact of outliers on the estimates
of model parameters and innovation variance has been extensively studied since Fox (1972)
proposed the concepts of additive outliers (AO) and innovational outliers (10) in time-series
0277-6693/93/0 10013-23$16.5 Received January 1992
0 1993 by John Wiley & Sons, Ltd. Accepted March 1992
14 Journal of Forecasting Vol. 12, Iss. No. I

modeling. More details on other developments in the detection and estimation of outliers can
be found in Guttman and Tiao (1978), Abraham and Box (1979), Hillmer et af. (1983), Hillmer
(1984), Chang et al. (1988), Tsay (1989), Ledolter (1989), Chen and Tiao (1990), Chen and Liu
(1993), and others. These procedures follow the line of a two-step operation: identifying the
locations and the types of outliers, and then adjusting the effects of the outliers for the purpose
of model parameter estimation.
To alleviate the influence of outliers in time-series modeling, Martin (1981) considered a
robust procedure for the estimation of model parameters in the presence of outliers. In this
approach, outlying observations are down-weighted through various types of a $-function in
the estimation process. However, all types of outliers are treated in the same fashion. This
partially explains why some robust methods perform well in some cases and perform poorly
in others, as reported in a recent simulation study conducted by Chuang and Abraham (1989).
Another disadvantage of this approach is that no information may be obtained regarding the
nature of an individual outlier. This may cause substantial loss of forecast accuracy since we
cannot effectively incorporate the impact of outliers on forecasts.
On the other hand, Smith and West (1983) proposed an on-line monitoring procedure to
detect the presence of abrupt level changes, slope changes and (additive) outliers within a state-
space model formulation. They classified each observation of a time series into one of four
states (steady state, change in level, change in slope, and outlier) in terms of the patterns of
three variances in the state-space model. To apply their proposed procedure, users need to
specify the variances and the prior probabilities of each state as the input information.
Inference on the outliers and structural changes is based on the posterior probabilities of
various states estimated by usual Bayesian approaches. The approach of Smith and West is
appropriate when sufficient prior knowledge and experiences about the data and potential
structural changes are available.
The main goal of this study is to investigate the outlier issues in time-series forecasting using
a detection and adjustment approach. Unlike outliers in time-series modeling, the study of
outliers in time-series forecasting is much less extensive and a number of issues still need to
be addressed. Hillmer (1984) studied the monitoring and adjustment of forecasts in the
presence of outliers under ARIMA models. Ledolter (1989) considered the effect of additive
outliers on the forecasts from ARIMA models. Both Hillmer (1984) and Ledolter (1989)
concluded that forecast intervals are quite sensitive to additive outliers. However, they found
that point forecasts are largely unaffected unless the outlier occurs near the forecast origin. In
general, forecasts are affected by outliers through (1) a carry-over effect of the outlier on the
point forecast and (2) a bias in the estimates of model parameters. Our study focuses on
forecasting issues when the procedure of Chen and Liu (1993) is used for the steps of parameter
estimation and outlier detection. In particular, we examine the situations that outliers occur
near or at the forecast origin.
In the studies by Hillmer (1984) and Ledolter (1989) only additive and innovational outliers
are considered. Since innovational outliers do not affect forecasts (see below), their primary
interest is on the effect of additive outliers. In this paper we also consider two other types of
outliers, namely, temporary change (TC) and level shift (LS). Our interest is to determine the
effects of these outliers on forecasts when they occur at the end of a time series (i.e. at the
forecast origin), or near the end of a time series. We shall also explore the loss of accuracy
in forecasts when the type of an outlier is misidentified or mis-specified. In studying the
accuracy of forecasts, we find that it is quite misleading to simply employ the overall post-
sample accuracy measures (e.g. root mean square error, or absolute average percentage error),
and more informative measures should be considered. By separating forecast errors during the
C. Chen and L - M . Liu Forecasting Time Series with Outliers 15

post-sample period according to the occurrences of outliers, new insights into the comparison
of forecast performance are revealed.
As mentioned earlier, the type of an outlier that occurs at the end of a series cannot be
determined empirically. Therefore, some judgment on the nature of the outlier is needed. Our
study also indicates that the loss of forecast accuracy due to undetected outliers near the
forecast origin may be substantial. To forecast in the presence of outliers, one needs to
combine the forecast of the uncontaminated series and the forecast of the outlier effects. We
propose a strategy which first jointly estimates model parameters and outlier effects using the
procedure of Chen and Liu (1993) and then uses a lower critical value to detect outliers near
the forecast origin in the forecasting stage. To improve the effectiveness of outlier detection
near the end of a time series, we also find it is helpful to restrict the types of outliers for the
observations near the forecasting origin.
This paper is organized as follows. In the next section we present the models and the types
of outliers considered in this study. In the third section we derive forecast errors and mean
square errors of forecasts if outliers that occur near or at the end of a time series are
misidentified or mis-specified. Simulation studies are conducted in the fourth section to study
the performance of the outlier-detection procedure when an outlier occurs near the end of the
series. Two examples are employed in the fifth section to illustrate forecasting time series with
outliers proposed in this paper. A summary and conclusions of this paper are provided in the
final section.

ARIMA MODELS WITH OUTLIERS

In this paper we shall assume that a time series can be modeled by an autoregressive-integrated
moving average, ARIMAb, d , q), model (Box and Jenkins, 1970):
+(B)(I - B)dZt = B(B)at, t = 1,2, ...,n (1)
where n is the number of observations for the series, B is the back shift operator such that
# ( B ) = 1 - 4,B - --- - &,Bp and B(B) = 1 - BIB- - 8,B“ are two polynomials
B i z t = Zt-i,
in B with orders p and q, d is a non-negative integer and al’s are independently and identically
distributed with mean zero and constant variance . a We also assume that + ( B ) and B(B) have
:
no common factors and that all of their roots lie outside the unit circle. If the time series is
seasonal with a periodicity s, we shall use the usual multiplicative ARIMA(p, d , q ) x ( P ,D,Q)
model as the underlying model:
+ ( B ) + ( B s ) ( l- B ) d ( l - B s ) D Z r = B ( B ) B ( B S ) a t , t = 1,2, ..., n (2)
where @ ( E s )= 1 - + l B S - ... - and 6 ( B s )= 1 - &BS- ... - S Q B ~ ’ satisfy certain
conditions similar to those of + ( B ) and B(B) defined in equation (1) and D is a non-negative
integer. Both models in equations (1) and (2) may include a constant term on the right-hand
side. To simplify the presentation, however, only models without a constant term will be
employed.
There are several ways in which an outlier can affect a time series. In this paper, four types
of outliers shall be considered. They are innovational outlier (10), additive outlier (AO), level
shift (LS) and temporary change (TC). Assuming only one outlier occurred in the series at
t = t I and l t ( r l ) = 1 when t = t I ,I f ( t l )= 0 otherwise, the observed series 1 Yr]can be expressed
as follows:
Yr = Zr + wL(B)Ir(tt) (3)
16 Journal of Forecasting Vol. 12, Iss. No. I

with

L(B)= 1 for A 0

for LS

and
1
L ( B ) = ___ for TC (0 < 6 < 1) (7)
(I - 6B)
For a seasonal time series, the L ( B ) for an innovational outlier needs to modified according
to equation (2).
It is useful to note that, except for the case of an 10, the effect of an outlier on the observed
series is independent of the model. Also, the A 0 and LS are the boundary cases of a TC, in
which 6 = 0 and 6 = 1 respectively. For the case of a TC, the outlier produces an initial effect
w at time l l and this effect dies out gradually with time. The parameter 6 is designed to model
the pace of the dynamic dampening effect. In practice, the value of 6 can be specified by the
analyst. To properly distinguish the nature of a TC from that of an A 0 or LS, we recommend
that 6 = 0.7 be used to identify a TC. The performance of such a specification can be found
in Chen and Liu (1993). In the case of an AO, the outlier causes an immediate and one-shot
effect on the observed series. An LS produces an abrupt and permanent step change in the
series.
The effect of an I 0 on a time series is more intricate than those of the other types of outliers.
According to equation (4), we see the effect of an I 0 is dependent upon the $j weights of the
ARIMA model of ZI. For a stationary series, an I 0 will produce a temporary effect since the
$j’s decay to zero exponentially. However, the pattern of { $ j ) for a nonstationary series can
be quite different. Depending on the model of Z t , an I 0 may produce
(1) An initial effect at the time of its occurrence and a level shift from the following period
when ZI follows an ARIMA(0, 1, 1) model;
(2) An initial effect at the time of its occurrence which will gradually converge to a permanent
level shift when ZI follows an ARIMA(1, 1, 1) model;
(3) A seasonal level shift when Z , follows a pure seasonal ARIMA(0, 1, 1) model, e.g. a level
shift at every January of each year and so on; and
(4) An annual trend change when Z, follows a multiplicative seasonal
ARIMA(0, 1, 1) x (0, 1 , 1) model.
This suggests that the effects of LS and TC out.liers may sometimes be approximated by an
1 0 effect.

Effects of an outlier on the forecasts from ARIMA models with known coefficients
To better understand the effect of an outlier on forecasts, we first consider model (3) and
assume that the coefficients of the ARIMA model and the outlier effect are known. The I-step-
ahead minimum mean square error (MMSE) forecast at time origin t = n can be obtained by
combining the forecasts of the two components
Yn(0 = Zn(1) + a L ( B ) I n + / ( f l ) (8)
C. Chen and L . M . Liu Forecasting Time Series with Outliers 17

where Z,,(l) is the I-step-ahead MMSE forecast of Z n + /at time origin n. In general, this
forecast is
z~(I=
) a ["Zn + *J"Zn- 1 + (9)
where Z,'s can be obtained based on the observations Yt's, the model parameters and the
outlier effects. The a-weights in the above computation are obtained from the following
recursive formula (see Box and Jenkins, 1970, p. 142):

(10)

and T ; ' ) = aj are the a-weights obtained from models (1) or (2). Alternatively, the forecast
Zn(I) can also be computed using
Z,,(I) = $ran+ $ / + I an-I + (1 1)
where $j's are the w-weights obtained from models (1) or (2) (see Box and Jenkins, 1970, p.
128). Under the above ideal situation (i.e. both ARIMA coefficients and outlier effects are
known), the I-step-ahead forecast error is
e n ( [ )= Yn+/ - Yn(I)= Zn+/ - Zn(I) (12)
Typically, information concerning the presence or the type of outliers is unknown. One
approach to address this issue is to apply the procedure developed by Chen and Liu (1993) to
detect the outlier and to obtain the model parameter estimates. The series adjusted for outlier
effects can then be regarded as an uncontaminated series and its forecast can be calculated
based on the usual recursive formula associated with the ARIMA model. In this case, the I-
step-ahead MMSE forecasts at time origin t = n can be derived based on the empirically
detected outlier as
= 2 n ( I ) + tL(B)In+/(il)
P~((I) (13)
where & ( I ) is the I-step-ahead MMSE forecast based on the adjusted observations and 2 , L ( B )
and fl are the estimated outlier effect, type and location, respectively. In other words, forecasts
in the presence of outliers can be obtained by adding the forecast of the uncontaminated series
(i.e. the outlier-adjusted series) and the extrapolation of the outlier effect. It is important to
emphasize that ,%(I) represents the forecast of the uncontaminated series (i.e. the adjusted
series). In general, this uncontaminated series is obtained via outlier adjustment based on the
location, type and size of the detected outlier. The empirically determined outlier type and
location, L^(B)and it, are the same as L ( B ) and t l if the type of outlier and its location are
detected correctly, otherwise they can be different. Thus, the forecast error in this case is
Cn(/) = Yn+/ - pn(O
= zn+/Z,(I) + (wL(B)In+r(tl)- G L ( B ) I n + / ( i I ) )
- (14)
As discussed in Chen and Liu (1993), an outlier occurred in the middle of the series usually
can be detected accurately. However, an outlier near the end of the series is difficult to identify
due to insufficient data. In particular, when an outlier occurs at the last period of the series,
the test statistics used to determine the outlier type are identical for the four types of outliers
in equations (4) to (7) and hence the outlier type cannot be empirically determined in this case.
Therefore, forecasters are required to make a judgmental decision as to the nature of the
outlier at the forecast origin. In other situations, an outlier may occur at one or two periods
before the forecast origin. In such cases, an outlier-detection procedure may not be able to
18 Journal of Forecasting VoI. 12, Iss. No. I

identify the type of the outlier correctly. In the following section we investigate the behavior
of forecast errors when an outlier occurs near or at the forecast origin assuming that a
detection procedure is applied. The cases to be considered include situations that the outliers
may or may not be identified; and if identified, the types may or may not be correct.

FORECAST ERRORS WHEN AN OUTLIER OCCURS NEAR OR AT THE

FORECAST ORIGIN

In this section we study the forecast errors when an outlier occurs near or at the forecast origin.
We first investigate how forecasts should be calculated when an outlier occurs. We then study
the forecast errors and loss of efficiency when misidentification takes place.

The forecast origin is identified as an outlier

Assuming the coefficients of the ARIMA models are known, the effect of an outlier described
in equation (3) can be computed based on the filtered series 4, where
r i r = n(B)Yt (15)
and n(B) is the n-polynomial obtained based on models (1) or (2) (Box and Jenkins, 1970).
At the forecast origin t = n, the estimate of the outlier effect (for all types) and its test statistic
are
d = 6,, (16)
and
.i= GI& (17)
where &, is a robust estimate of ua. More details on the estimation of outlier effects can be
found in Chen and Liu (1993). In this situation, even though we can test if the observation
at t = n is an outlier, a judgmental decision is required to determine the type of the outlier.
Depending upon the choice of the outlier type, the I-step-ahead forecast p,,(I) can be computed
as follows based on equation (13):

The notation .??,,([)NO(which is equal to p,,(I)m) represents the I-step-ahead MMSE forecast
of Yn+rwithout any outlier adjustment, and .??,,(I)TPis the I-step-ahead MMSE forecast of Zn+l
based on the adjusted observations if an outlier exists. In the case that an outlier occurs at the
forecast origin, the estimated outlier effect for any outlier type is the same (G), and its impact
can only be observed at t = n. Therefore, a judgmental decision of the outlier type in this case
only affects the extrapolation of the outlier effects to the future observations, and has no effect
on the adjusted observation at the forecast origin.
The results of equation (18) show that the outlier-adjusted forecasts consist of the sum of
the forecasts of the uncontaminated series and the outlier effects on future observations. The
former component represents the homogeneous behavior of the series and the latter denotes
C. Chen and L.-M. Liu Forecasting Time Series with Outliers 19

the impact of the outlier on the future observations. When the forecast origin is detected as
an outlier, typical outlier detection procedures are able to provide an estimate for the first
component, but may have difficulty to provide complete information for the second.
To obtain a better understanding of the effect of an outlier adjustment on forecasts, we
express the outlier-adjusted I-step-ahead forecast Pn(I) relative to typical forecasts without any
outlier adjustment :

The above expression can be easily obtained using results of equations (lo), (11) and (16).
The results of equation (19) show the differences among forecasts with outlier adjustment
and those if outliers are ignored. It is useful to note that under the I 0 situation, the I-step-
ahead forecast is identical to the forecast if no outlier adjustment is employed. This finding
is in congruence with that obtained by Ledolter (1989), and is an alternative way of showing
the adaptive nature of the ARIMA forecasts. That is, when an outlier produces an effect that
follows the pattern of the $-weight memory there is no need to conduct any adjustment for
the outlier to obtain a more accurate forecast. This is part of the reason that traditional
ARIMA models may still perform well in forecasting even if the time series is subjected to
outliers and interventions.
To obtain the statistical properties of the forecast errors, we need to express the forecast of
the uncontaminated series in terms of the forecast of the underlying process as described in
equation (9). The adjusted series can be obtained by substituting the detected outliers into
equation (3). Hence it can be shown that
Z n ( I )+ $I(W - G ) when an outlier occurs at t = n
z n (OTP = (20)
Zn(I) - $6 when no outlier occurs at t = n
with TP = AO, TC, 1 0 , LS and
Zn(I)+ $ 1 ~ when an outlier occurs at t = n
zn(l)NO = [zn(r) (21)
when no outlier occurs at t = n
where Z n ( l ) is defined in equation (9). Based on equations (3), (17) and (21), we can derive
the forecast errors contingent upon the specification of outlier type in terms of en(l)defined
in equation (12) and the outlier effect. These forecast errors are listed in Table I. In this table
the diagonal terms provide the forecast error when the outlier type is correctly specified, and
the off-diagonal terms provide the forecast error when the outlier type is mis-specified. In
particular, the first row gives the forecast errors when forecasts are made with no adjustment
for outliers. It then provides the bias due to an undetected outlier at the forecast origin. Except
for the LS outlier, the major effect for an outlier is on short-term forecasting, i.e. when I is
small. The effect associated with a mis-specified outlier type depends on the underlying
memory pattern ($11 and the actual outlier type.
Using the fact that

;=[ w
an
+ an when an outlier occurs at t = n
when no outlier occurs at t = n
20 Journal of Forecasting VOI. 12, Iss. No. I

+++++

h
c
Lo
3

I I + + +

I l l
3 3 '3 '3 '3
3 3 w w w
I I + + +

h N
h
- n
Lo- 3 "3
I. I.
-33
3 w w
'3 t3 '3
+++

h
- n
Lo-
I I

+++
+++
C. Chen and L.-M. Liu Forecasting Time Series with Outliers 21

we can easily obtain the mean square errors (MSE) of the forecast in correspondence to those
shown in Table I. The first row of this table shows the effects of various outliers when outlier
detection and adjustment is not employed. We observe that except for 10, the MSE of
forecasts increases if an outlier occurs at the end of a series.
By examining Table I1 row by row it is easy to see that correct outlier detection produces
the smallest mean square error. For a given row, the ratio of MSEs to the smallest MSE of
the row provides information on the loss of efficiency due to an incorrect judgmental decision
for using a specific outlier type. The columns of Table I1 list the MSE of forecasts under five
different conditions. For the cases of I 0 and no outlier, if there is a mis-specification of an
AO, TC or LS, the MSE is increased. When an I 0 is misidentified as a non-outlier, or an
observation is spuriously identified as an 10, no loss of MSE is involved. This result may
provide a basis to argue that perhaps there is no need to consider detecting an I 0 in the context
of forecasting (Ledolter, 1989). However, this may not be a sound suggestion considering that
when an I 0 is misidentified as AO, TC or LS, there is always a loss of efficiency in the MSE.
In general, the loss of efficiency in MSE due to a mis-specified outlier type depends on (1)
$-weights of the underlying process, (2) the size of the outlier effect w and (3) outlier types.
When an AO, TC or LS is mis-specified as an I 0 or no outlier, or when an A 0 is mis-specified
as a TC or an LS, the MSE of forecasts is always greater than that of a correct detection if
I u I >a,. When a TC or an LS is mis-specified as an AO, TC or LS, the MSE increases if

Outlier occurs k periods before the forecast origin

When the outlier occurs k periods before the forecast origin n , the observation Yl can be
exmessed as

The residuals obtained by filtering the series Yl with the known model parameters can be
expressed as
Un-i =
a,, - i
1 i> k
+ wa(B)L(B)I,-;(n- k ) 0 < i < k
A misidentified outlier may result from a correct location detection but a wrong type
specification, from a correct type identification but a wrong location identification, or from
both incorrect type and location. In this investigation of the effect of a misidentified outlier
on the forecast, we only consider the case that the location of an outlier is correctly identified
but the type of an outlier may be mis-specified. The results of other cases are complicated and
do not provide further insight on the outlier effects on the forecast and hence are not presented
here.
When an outlier is detected at k periods before the forecast origin n , the estimates for
different types of outliers are
WIO = Un-k, ? I 0 = (-&O/&I

&A0 = ( k - k - 7rIlL-k+I - ... - aka,,)/SS(a)Ao, FA0 = hAO/ (<a/SS(?r)AO)

&K= (an-/(+ (6 - T I ) L i , , - k + , + ... ?TC = GTC/ ( & / S ~ ( ~ ) T C )

+ ( s k - s k - ' a l - .*.- ak)a,)/SS(a)Tc,
GLS = (an-k + (1 ~ i ) L i ~ - k ++i ... + (1 - H I - ...
- - a k ) i , , ) / S S ( a ) ~ s ,? L s = cj~s/(6n/SS(~)Ls)
(25)
22 Journal of Forecasting Vol. 12, Iss. No. I

where
SS(a)*o = 1 + a: + ' * . + 7ri
SS(7r)Tc = 1 + (6 - + + (sk - 6 k - 1 T l - - 7rk IZ
**'

SS(7r)LS = 1 + (1 - + ..' + (1 - 7rl - **. - 7rk)

2

Using equation (23), the adjusted observations according to the outlier type TP can be
expressed as:
i>k
Yn-i - ~ T p L ( ~ ) z , - i (-
nk) i <k

Using the above result and equation (9), the I-step-ahead forecast of the adjusted series of the
outlier type T P can be expressed as

%(f)Tp = zn(I) + ,z
i=O
K
T!;)I [wL(B)In-i(n - k ) - GTPL^(B)ln-i(n - k ) ] (27)

where &(I) is the I-step-ahead MMSE forecast of the underlying process [ Z , ] .As in equation
(13), we can express ~,,(I)TPas
P Zn(0-r~+ cj~~L^(B)l,+,(n
F ~ ( I ) T= -k) (28)
Thus the forecast error defined in equation (14) can be rewritten as
zn(f) = Yn+f- pn(f)TP
= - zn(0l

+ [uL(B)I,+,(n- k) - G~pL^(B)ln+/(n- k ) ] (29)

The three components (each enclosed in a pair of brackets) in the above expression
respectively represent (1) the intrinsic uncertainty of forecast errors associated with the
underlying time series process, (2) sampling errors in the adjustment of outlier effects on the
observations and (3) the error associated with the extrapolation of the outlier effect.
Substituting t2t-i of equation (24) into equations (25) and (29), we can derive the MSE of
forecast errors as a function of the detected outlier and the true outlier similar to those shown
in Table 11.
In summary, we have explicitly demonstrated the effect of an outlier that occurs at or near
the forecast origin. The potential loss of efficiency due to either ignoring the presence of an
outlier or misidentifying the outlier type is also discussed. In practice, a judgmental decision
is required to determine the outlier type when it occurs at the forecast origin. When outliers
occur a few periods before the forecast origin, the usual outlier detection procedure may not
be very effective due to limited available information. A strategy of reducing the critical value
for detecting outliers at or near the end of the series is proposed in the next section. A
simulation study is also conducted to demonstrate the performance of the proposed method.
To reduce the probability of misidentifying outlier type near the end of a time series, it may
be useful to restrict the choice of outlier types to an appropriate subset.
C. Chen and L.-M. Liu Forecasting Time Series with Outliers 23

OUTLIER DETECTION NEAR OR AT THE FORECAST ORIGIN

As pointed out by Chen and Liu (1993), the estimates of model parameters can be affected by
the choice of the critical value for outlier detection in the joint estimation of model parameters
and outlier effects. If the critical value is too large, the power of outlier detection will be low.
In the extreme situation that a very large critical value is used, it is equivalent to no outlier
detection. When a smaller critical value is employed for outlier detection, the power of outlier
detection is increased. However, the chance of Type I error (detecting spurious outliers) is also
increased. As a result, the estimates of model parameters based on the joint estimation
procedure may also be biased due to over-adjustment of spurious outliers. For the purpose of
forecasting, we would like to obtain accurate estimates of model parameters, and also to have
high power of outlier detection during the forecasting period. To achieve these two goals, we
recommend a two-step strategy. (1) for model estimation, use a moderate critical value in the
joint estimation procedure; and (2) for forecasting, use the model estimated in (l), but employ
a smaller critical value for outlier detection during forecasting.
In this section we present a study of the performance of the above strategy with respect to
a range of critical values. In addition to studying the power of outlier detection, it is also
desirable to obtain empirical information concerning the nature of type misidentification of
outliers. The simulation study considers four major factors: (1) the model of the underlying
process, (2) the size of the outlier effect, (3) the location of the outlier, and (4) the magnitude
of the critical value. For each outlier type, a time series of length 100 is generated based on
one of three models: AR(1), MA(1) and IMA(1,l). The standard deviation of the innovation
series is set to 1. Two outlier effect sizes (o= 3 and o = 4) and three outlier locations ( t l = 50,
t1 = 98 and tl = 100) are considered. Assuming the model is correctly specified, the simulated
series (which is contaminated with an outlier) is fitted by the joint estimation procedure of
Chen and Liu (1993) with the critical value C = 2.75 for outlier detection. Based on the model
estimated in this manner, outlier detection is applied to the same series again using six possible
critical values: C = 2.00, C = 2.25, C = 2.50, C = 2.75, C = 3.00 and C = 3.25. For each
specification of outlier type, underlying model, outlier effect size and outlier location, the
above procedure is repeated 500 times. The main goal of this simulation is to examine the
performance of the outlier detection under various situations. The results of this simulation
study are summarized in three tables. Table I11 presents the probabilities of the correct
detection of the outlier location. Tables IV and V provide more detailed classification of the
types of outlier detected.

Power of outlier detection and its implication

The results in Table I11 are organized into six groups according to the critical value for outlier
detection. For each outlier type, the first row shows the probability of correct detection of
location (regardless of type) when the outlier occurred at t l = 50 and the second row gives those
when tl = 98. The last row in each group presents the results when t l = 100. It is important to
note that when an outlier occurs at the end of the series (tl = 100) there is no statistical
information to determine the outlier type. Except for the case of a level shift, the power of
detection is not greatly affected by the underlying models. As expected, the power for the cases
of the smaller outlier (a= 3) is lower than that of the larger outlier (a = 4). Another study
based on a multiplicative seasonal model (the airline model) gives very similar results. The
power of detecting an outlier is slightly lower than that of the IMA(0, 1 , l ) case, but very close
to that of the MA(1). With respect to the location of outliers, we find no major differences
when an outlier occurs in the middle of the series (tl = 50) o r near the end ( t l = 98) based on
 h)
P

Table 111. Detection power of outlier location

AR(1) MA(1) IMA(1,l) AR(1) MA(1) IMA(1,l) AR(1) MA(1) IMA(1, 1) AR(1) MA(1) IMA(1,l)
w=3 w=4 w=3 w=4 T
3
e
c=2.00 C = 2.75
I0 0.90 0.85 0.88 0.98 0.96 0.97 I0 0.69 0.61 0.73 0.93 0.88 0.91
0.84 0.83 0.87 0.97 0.97 0.98 0.68 0.62 0.72 0.92 0.89 0.92
0.92 0.94
23
A0 0.91 0.82 0.91 0.99 0.95 0.98 A0 0.78 0.70 0.78 0.97
0.91 0.81 0.89 0.98 0.94 0.99 0.78 0.71 0.73 0.95 0.92 0.95
0.92 0.95
82
TC 0.90 0.84 0.89 0.97 0.97 0.98 TC 0.68 0.63 0.78 0.92
0.84 0.84 0.86 0.97 0.98 0.97 0.67 0.76 0.74 0.93 0.92 0.92 09
LS 0.78 0.64 0.88 0.91 0.75 0.98 LS 0.62 0.60 0.79 0.84 0.72 0.96
0.84 0.81 0.86 0.97 0.95 0.96 0.73 0.64 0.78 0.94 0.90 0.95
T = 100 0.77 0.77 0.79 0.94 0.91 0.97 T = 100 0.58 0.55 0.61 0.89 0.83 0.91
C = 2.25 c=3.00
I0 0.88 0.84 0.88 0.98 0.96 0.97 I0 0.58 0.54 0.61 0.89 0.84 0.88
0.83 0.81 0.87 0.97 0.95 0.98 0.61 0.52 0.63 0.89 0.84 0.87
A0 0.90 0.80 0.90 0.99 0.95 0.98 A0 0.72 0.67 0.69 0.96 0.90 0.92
0.90 0.80 0.88 0.98 0.93 0.99 0.71 0.67 0.64 0.92 0.91 0.91
TC 0.87 0.81 0.89 0.97 0.97 0.98 TC 0.60 0.57 0.68 0.89 0.89 0.92
0.82 0.83 0.86 0.97 0.97 0.97 0.56 0.75 0.67 0.89 0.90 0.88
LS 0.78 0.62 0.88 0.91 0.73 0.98 LS 0.58 0.61 0.73 0.84 0.72 0.94
0.83 0.80 0.86 0.97 0.94 0.96 0.67 0.60 0.72 0.93 0.83 0.94
T = 100 0.75 0.74 0.79 0.94 0.90 0.97 T = 100 0.50 0.48 0.50 0.83 0.79 0.86
C = 2.50 C = 3.25
I0 0.82 0.73 0.82 0.96 0.93 0.95 I0 0.47 0.45 0.52 0.82 0.76 0.83
0.76 0.73 0.82 0.96 0.93 0.97 0.52 0.41 0.51 0.85 0.77 0.82
A0 0.87 0.77 0.86 0.99 0.95 0.97 A0 0.62 0.64 0.60 0.93 0.89 0.88
0.85 0.77 0.82 0.97 0.93 0.98 0.59 0.61 0.52 0.87 0.89 0.85 T
TC 0.78 0.75 0.86 0.95 0.96 0.97 TC 0.48 0.46 0.59 0.81 0.83 0.89 0
1
0.78 0.80 0.82 0.96 0.95 0.95 0.47 0.73 0.56 0.84 0.86 0.82
LS 0.71 0.60 0.86 0.89 0.72 0.97 LS 0.54 0.60 0.67 0.82 0.73 0.94 -c
0.79 0.71 0.84 0.96 0.92 0.96 0.59 0.57 0.65 0.89 0.77 0.91
T = 100 0.69 0.65 0.71 0.92 0.87 0.95 T = 100 0.38 0.38 0.40 0.78 0.71 0.79 $
For 10, AO, TC, and LS cases, first row t = 50, second row t = 98.
C. Chen and L.-M. Liu Forecasting Time Series with Outliers 25

Table IV. Power of outlier detection: a simulation study when an outlier occurs at t , = 50 and n = 100 (The first
number in each cell represents the probability associated with w = 3 and the second number associated with w = 4)
~ ~~ ~ ~

AR(1) model (6= 0.6) MA(1) model (0 = 0.6) IMA( 1,l) model (0 = 0.6)

Detected type Detected type Detected type

Actual type 10 A0 TC LS I0 A0 TC LS I0 A0 TC LS
~ ~~

10 0.348 0.156 0.394 0.000 0.752 0.100 0.000 0.000 0.414 0.202 0.136 0.132
0.414 0.130 0.438 0.000 0.842 0.120 0.002 0.000 0.562 0.170 0.140 0.094
A0 0.148 0.686 0.074 0.000 0.162 0.630 0.026 0.000 0.222 0.604 0.066 0.014
0.106 0.854 0.032 0.000 0.098 0.828 0.028 0.000 0.216 0.744 0.024 0.000
C = 2.00
TC 0.294 0.120 0.482 0.000 0.264 0.120 0.456 0.000 0.138 0.096 0.576 0.080
0.332 0.060 0.580 0.002 0.029 0.062 0.622 0.000 0.128 0.032 0.776 0.044
LS 0.110 0.060 0.058 0.552 0.016 0.076 0.044 0.504 0.104 0.016 0.082 0.678
0.138 0.058 0.046 0.672 0.014 0.094 0.044 0.594 0.064 0.000 0.036 0.878
I0 0.344 0.148 0.390 0.000 0.736 0.102 0,000 0.000 0.410 0.200 0.134 0.132
0.414 0.130 0.438 0.000 0.836 0.120 0.002 0.000 0.562 0.170 0.140 0.094
A0 0.148 0.684 0.072 0.000 0.148 0.624 0.028 0.000 0.216 0.602 0.064 0.014
0.106 0.854 0.032 O.OO0 0.088 0.832 0.034 0.000 0.216 0.744 0.024 0.000
C = 2.25 TC 0.284 0.118 0.470 O.OO0 0.252 0.116 0.446 0.000 0.138 0.096 0.576 0.078
0.332 0.058 0.580 0.002 0.290 0.062 0.622 0.000 0.128 0.032 0.776 0.044
LS 0.108 0.060 0.054 0.554 0.012 0.064 0.038 0.504 0.102 0.016 0.082 0.678
0.138 0.056 0.046 0.674 0.004 0.086 0.038 0.604 0.062 0.000 0.036 0.878
10 0.314 0.136 0.372 O.OO0 0.636 0.092 O.OO0 0.000 0.382 0.188 0.124 0.126
0.408 0.126 0.426 0.000 0.810 0.120 0.002 0.000 0.554 0.160 0.134 0.104
A0 0.142 0.662 0.064 0.000 0.122 0.608 0.036 0.000 0.208 0.580 0.062 0.012
0.106 0.850 0.030 0.000 0.082 0.828 0.038 0.000 0.214 0.736 0.024 0.000
C = 2.50
TC 0.264 0.092 0.428 0.000 0.228 0.108 0.410 0.000 0.130 0.088 0.564 0.078
0.326 0.058 0.568 0.000 0.280 0.062 0.618 0.000 0.126 0.032 0.774 0.042
LS 0.094 0.050 0.048 0.522 0.004 0.046 0.032 0.514 0.092 0.014 0.078 0.674
0.130 0.042 0.042 0.678 0.002 0.076 0.030 0.608 0.058 0.000 0.030 0.878
10 0.272 0.092 0.324 0.000 0.536 0.076 0.000 0.000 0.340 0.162 0.110 0.120
0.388 0.122 0.416 0.000 0.762 0.116 0.000 0.000 0.534 0.140 0.132 0.108
A0 0.124 0.610 0.048 0.000 0.104 0.566 0.032 0.000 0.186 0.524 0.060 0.008
0.102 0.840 0.028 0.002 0.064 0.816 0.038 0.000 0.204 0.718 0.022 0.000
C = 2.75
TC 0.228 0.068 0.378 0.002 0.188 0.088 0.352 0.000 0.108 0.068 0.528 0.074
0.314 0.056 0.546 0.000 0.266 0.060 0.598 0.000 0.122 0.026 0.764 0.042
LS 0.082 0.036 0.038 0.468 0.002 0.038 0.022 0.536 0.082 0.004 0.072 0.628
0.128 0.038 0.038 0.638 0.004 0.076 0.018 0.620 0.058 0.000 0.024 0.878
10 0.226 0.068 0.290 0.000 0.468 0.074 0.000 0.000 0.294 0.118 0.096 0.106
0.374 0.116 0.404 0.000 0.724 0.118 0.000 0.000 0.516 0.132 0.128 0.106
A0 0.110 0.564 0.044 0.000 0.086 0.552 0.036 0.000 0.160 0.474 0.054 0.006
0.100 0.828 0.026 0.002 0.056 0.810 0.038 0.000 0.200 0.696 0.022 0.002
c=3.00
TC 0.204 0.058 0.336 0.002 0.166 0.070 0.330 0.000 0.090 0.050 0.480 0.058
0.308 0.048 0.534 0.000 0.256 0.060 0.574 0.002 0.112 0.024 0.748 0.038
LS 0.074 0.024 0.028 0.452 0.000 0.032 0.020 0.562 0.070 0.000 0.062 0.600
0.122 0.034 0.034 0.646 0.002 0.070 0.018 0.626 0.050 0.000 0.026 0.868
10 0.176 0.046 0.244 0.000 0.388 0.064 0.000 0.000 0.252 0.100 0.080 0.090
0.346 0.098 0.380 0.000 0.646 0.114 0.000 0.000 0.488 0.126 0.124 0.090
A0 0.088 0.498 0.038 0.000 0.076 0.524 0.036 0.000 0.132 0.312 0.050 0.006
0.096 0.804 0.026 0.002 0.054 0.796 0.038 0.000 0.192 0.668 0.022 0.002
C = 3.25
TC 0.156 0.048 0.278 0.002 0.128 0.060 0.276 0.000 0.082 0.046 0.412 0.054
0.290 0.040 0.484 0.000 0.234 0.054 0.540 0.002 0.104 0.020 0.726 0.038
LS 0.066 0.010 0.022 0.440 0.000 0.026 0.014 0.564 0.058 0.000 0.050 0.558
0.120 0.026 0.028 0.642 0.000 0.064 0.018 0.652 0.046 0.000 0.026 0.864
26 Journal of Forecasting Vol. 12, Zss. No. I

Table V. Power of outlier detection: a simulation study when an outlier occurs at tl = 98 and n = 100 (The first
number in each cell represents the probability associated with w = 3 and the second number associated with w = 4)

AR(1) model (6= 0.6) MA(]) model (0 = 0.6) IMA(1,I) model ( O = 0.6)

Detected type Detected type Detected type

Actual type I0 A0 TC LS I0 A0 TC LS 10 A0 TC LS

10 0.328 0.176 0.254 0.084 0.680 0.138 0.006 0.002 0.306 0.178 0.292 0.092
0.390 0.134 0.358 0.092 0.778 0.186 0.002 0.002 0.470 0.180 0.278 0.052
A0 0.144 0.698 0.052 0.012 0.104 0.612 0.078 0.014 0.198 0.636 0.046 0.008
0.136 0.828 0.012 0.004 0.084 0.758 0.092 0.004 0.196 0.768 0.028 0.00
c = 2.00
TC 0.238 0.116 0.302 0.184 0.112 0.102 0.474 0.148 0.178 0.060 0.442 0.202
0.302 0.058 0.446 0.168 0.226 0.034 0.572 0.144 0.230 0.024 0.554 0.164
LS 0.056 0.028 0.192 0.560 0.288 0.032 0.178 0.308 0.066 0.004 0.154 0.638
0.034 0.010 0.138 0.790 0.474 0.028 0.142 0.310 0.054 0.004 0.134 0.766
10 0.322 0.174 0.254 0.082 0.666 0.136 0.004 0.002 0.306 0.178 0.292 0.092
0.390 0.134 0.358 0.092 0.762 0.186 0.002 0.002 0.470 0.178 0.278 0.052
A0 0.140 0.694 0.052 0.010 0.098 0.610 0.080 0.014 0.194 0.632 0.046 0.008
0.134 0.828 0.012 0.004 0.080 0.758 0.092 0.004 0.196 0.768 0.028 0.000
c = 2.25
TC 0.230 0.112 0.298 0.178 0.100 0.102 0.476 0.148 0.176 0.058 0.420 0.202
0.302 0.056 0.446 0.168 0.222 0.034 0.572 0.144 0.230 0.024 0.554 0.164
LS 0.056 0.026 0.190 0.558 0.280 0.032 0.178 0.308 0.066 0.004 0.154 0.638
0.034 0.008 0.138 0.790 0.464 0.028 0.144 0.308 0.054 0.004 0.132 0.766
10 0.298 0.146 0.242 0.076 0.604 0.126 0.000 0.002 0.290 0.164 0.284 0.086
0.382 0.130 0.354 0.092 0.744 0.188 0.002 0.000 0.468 0.176 0.270 0.052
A0 0.128 0.666 0,050 0.010 0.082 0.596 0.078 0.014 0.184 0.590 0.042 0.006
0.128 0.826 0.012 0.004 0.074 0.758 0.092 0.004 0.196 0.760 0.026 0.000
c = 2.50
TC 0.220 0.100 0.288 0.170 0.074 0.094 0.480 0.148 0.166 0.048 0.406 0.202
0.296 0.054 0.440 0.166 0.196 0.034 0.576 0.144 0.224 0.018 0.544 0.164
LS 0.050 0.016 0.184 0.540 0.198 0.026 0.178 0.312 0.062 0.002 0.150 0.626
0.032 0.000 0.142 0.788 0.442 0.022 0.144 0.312 0.032 0.000 0.142 0.788
10 0.266 0.128 0.218 0.070 0.512 0.102 0.000 0.002 0.260 0.132 0.256 0.074
0.370 0.122 0.344 0.086 0.700 0.188 0.002 0.000 0.462 0.156 0.258 0.048
A0 0.116 0.610 0.048 0.010 0.068 0.562 0.066 0.010 0.162 0.534 0.032 0.006
0.122 0.816 0.012 0.002 0.064 0.756 0.092 0.004 0.186 0.742 0.026 0.000
C = 2.75
TC 0.180 0.078 0.256 0.156 0.044 0.090 0.480 0.148 0. I48 0.042 0.364 0.188
0.290 0.046 0.430 0.162 0.168 0.034 0.574 0.146 0.220 0.014 0.536 0.154
LS 0.042 0.014 0.170 0.504 0.130 0.024 0.178 0.312 0.056 0.000 0.138 0.582
0.030 0.004 0.130 0.776 0.420 0.024 0.144 0.312 0.046 0.000 0.134 0.768
I0 0.238 0.104 0.206 0.064 0.424 0.092 0.000 0.000 0.226 0.112 0.226 0.068
0.366 0.1 14 0.336 0.078 0.652 0.182 0.002 0.000 0.448 0.134 0.250 0.042
A0 0.098 0.558 0.042 0.010 0.060 0.540 0.062 0.010 0.136 0.472 0.030 0.006
0.120 0.790 0.012 0.000 0.060 0.752 0.090 0.006 0.180 0.706 0.026 0.000
C = 3.00
TC 0.148 0.066 0.210 0.134 0.030 0.088 0.480 0.150 0.126 0.036 0.338 0.170
0.280 0.036 0.422 0.152 0.144 0.034 0.574 0.146 0.208 0.014 0.508 0.150
LS 0.038 0.010 0.158 0.462 0.088 0.022 0.178 0.312 0.050 0.000 0.126 0.546
0.028 0.002 0.126 0.772 0.358 0.020 0.144 0.312 0.044 0.000 0.132 0.764
10 0.196 0.080 0.180 0.060 0.338 0.072 0.000 0.000 0.192 0.076 0.186 0.058
0.348 0.106 0.324 0.070 0.608 0.162 0.002 0.000 0.412 0.124 0.242 0.040
A0 0.084 0.466 0.034 0.010 0.046 0.494 0.060 0.008 0.122 0.372 0.022 0.006
0.110 0.754 0.010 O.Oo0 0.060 0.742 0.086 0.006 0.168 0.662 0.024 0.000
C = 3.25
TC 0.136 0.048 0.182 0.104 0.014 0.086 0.480 0.152 0.098 0.026 0.286 0. I50
0.266 0.032 0.398 0.142 0.108 0.032 0.574 0.146 0.194 0.012 0.472 0.144
LS 0.030 0.004 0.140 0.416 0.062 0.018 0.178 0.312 0.046 0.000 0.112 0.494
0.022 0.002 0.124 0.742 0.294 0.016 0.144 0.312 0.040 O.Oo0 0.126 0.748
C. Chen and L - M . Liu Forecasting Time Series with Outliers 27

the models we explored. However, this does not imply that the same result holds true for all
models. For more complicated models, we expect that the power will be reduced if the outlier
occurs near the end of a series.
In the case when an outlier occurs at the end of the series we observe a noticeable power
reduction. This is more pronounced for the cases of the smaller outlier effect (w = 3) than that
of the larger effect (w=4). These findings suggest that (1) the proposed procedure is quite
effective for reasonably large outliers (w = 4) even if the outlier occurs near or at the end of
the series; and (2) a smaller critical value may be needed to detect outliers of smaller sizes
(W < 3), particularly when the outlier is at the end of the series. Based on the results in Table
111, we find that when the size of the outlier effect is 3 (w = 3), the probability to detect the
outlier at the end of a series is only about 50%. The probability increases to about 70% if the
critical value is reduced to C = 2.50. Furthermore, based on the results discussed in the
previous section and summarized in Table 11, we find that if the effect of a n outlier is greater
than the standard deviation of the innovation series (ua), then the MSE of the forecasts will
be improved if the outlier can be correctly detected. In practice, we may only be interested in
outliers at least greater than 20,. With these considerations in mind, a smaller critical value
(say, between 2.0 and 2.5) is desirable for outlier detection near the forecast origin.
There are a few points worth further explanation regarding the results in Table 111. In the
case of a level shift, two unusual patterns are observed: (1) the power of detection in the MA(1)
case is lower, and (2) the power of detection for outliers in the middle of the series is lower
than those with outliers near the end of the series. The former result is mainly due to the strong
implication of a level shift on parameter estimates when its effect is not adjusted. In the latter
case, a more detailed examination of the detected results shows that there is a higher tendency
to obtain a detection of a level shift missed by one time period when it occurs in the middle
of the series. By including the detected outliers that miss by one time period, we find the similar
pattern that the power of detecting a level shift is again higher in the middle of the series than
near the end.
Misidentification of outlier type and its implication
To take full advantage of outlier adjustment in forecasting, we need to identify both the
location and the type correctly. Tables IV and V summarize the results of more detailed
classification of outliers detected. Table IV presents the results for an outlier that occurs in the
middle of a time series (tl = 50) and Table V those for an outlier that occurs near the end of
the series ( t , = 98). In each panel, the results are displayed into eighteen groups associated with
six critical values and three models. Within each group, we present the frequency of the
detected outlier type for each case of actual outlier type. For instance, in the first group of
Table IV the first row and the first column gives the probability that an I 0 is identified as an
I 0 with critical value C = 2.00 and an AR(1) underlying model. Generally, the following
observations are obtained:
(1) The probability of type misidentification is greater when the critical value is small;
(2) The probability of misidentification increases when an outlier occurs near the end of the
series;
(3) A 0 has lower chance to be misidentified in comparison with 10, TC or LS;
(4) When an A 0 is misidentified, it is typically classified as an I 0 (depending on the
underlying model, misidentification ranges from 10% to 20%);
( 5 ) LS has a low chance to be misidentified if the outlier occurred in the middle of the series,
and has a higher probability to be misidentified as I 0 or T C if the outlier occurs near the
end of the series:
28 Journal of Forecasting voi. 12, rss. NO. I

(6) TC and I 0 are not clearly distinguishable, particularly for AR( 1) and IMA( 1, l) models;
and
(7) The power of detection is a decreasing function of the critical value and an increasing
function of the size of the outlier effect.
The above findings indicate the usefulness as well as some potential limitations in the
application of outlier adjustment in forecasting. In addition, it provides information for the
strategy of forecasting time series with outlier adjustment. Since the probability of
misidentification of outlier types increases when the critical value decreases, particularly if the
outliers occur closer to the end of the series, it is more advantageous if the types of outliers
under detection are restricted to a more limited subset. Since the gain of forecast efficiency is
high for an A 0 and an A 0 is less likely to be misidentified, it is advisable that A 0 is always
considered. As discussed in the third section, I 0 has an adaptive nature similar to typical
ARIMA innovations, thus it should also be considered. It is not advisable to always include
TC and LS outliers during the forecasting process since such outliers cannot be accurately
determined in terms of their types and effects when they occur near the end of a time series.
However, if strong evidence exists for such outlier types, they should be incorporated into the
forecasting process accordingly. The finding here provides a justification for forecasting in
conjunction with judgment.

ILLUSTRATIVE EXAMPLES

In this section we use two examples to illustrate the application of outlier detection and
adjustment in forecasting. The first example is the monthly totals (in thousands) of
international airline passengers from January 1949 through December 1960 in Box and Jenkins
(1970). The second example is the quarterly imports of goods and services of Taiwan between
the first quarter of 1968 and the fourth quarter of 1990. In the analysis of these examples we
employ the SCA Statistical System (Liu e t a / . , 1986, and Chen ef a/., 1990) for joint estimation
of model parameters and outlier effects. Other examples related to forecasting with outliers can
be found in Hillmer (1984), Tsay (1988), Liu (1991), and Liu and Lin (1991). Before we present
these two examples, we first consider the criteria for forecast performance that may be
employed.

Forecast performance
In this research we are particularly interested in the improvement of forecast performance due
to outlier adjustment. To evaluate forecast performance, we may employ the root mean
squared error (RMSE) for the post-sample periods. Generally, the RMSE is defined as
r- __

where Z , is the one-step-ahead forecast of Zr based on an estimated model and m is the number
of forecasts used in the comparison. Assuming that the estimated model is representative of
the forecasting period, the post-sample RMSE should be in consonant with the sample
standard deviation (&) of the estimated model. While the gross RMSE defined above is
appropriate if no outliers exist, this value may be greatly increased if any outliers exist during
the post-sample period. As a result, comparisons of forecast performance based on the gross
RMSE are often misleading and inclusive. We shall denote the gross RMSE defined in equation
C. Chen and L.-M. Liu Forecasting Time Series with Outliers 29

(30) as RMSE,. To obtain better insights into the effects of outlier adjustment on forecasting,
we consider three additional variations of RMSE.
Based on the results in the third section of this paper, it is easy to see that at a time point
that an outlier occurs, the forecast cannot be improved, no matter if outlier adjustment is
employed or not. However, the forecasts after such a time point can be greatly affected
depending upon whether outlier adjustment is applied properly or not. In particular, the
forecasts immediately following an outlier are subject to the greatest impact. With these points
in mind, we consider the following variations of RMSE:
RMSEo: Post-sample RMSE computed using the time periods where outliers occurred
RMSE,: Post-sample RMSE computed using the time periods immediately next to the
outliers
RMSE,: Post-sample RMSE computed using the time periods excluding outliers and those
immediately following the outliers
In the evaluation of forecast performance we shall compare the quadruplet (RMSE,, RMSE,,
RMSE,, RMSE,).

Airline passenger example

In this example we consider the Series G of Box and Jenkins (1970), the monthly totals (in
thousands) of international airline passengers from January 1949 through December 1960.
There are 144 observations in this series. The first 132 observations are employed for model
fitting and the last 12 observations are reserved for post-sample comparisons. In order to
obtain a more homogeneous variance, we will use the natural logarithm of the monthly totals.
This log transformed series is denoted as Zt and is plotted in Figure 1.
Box and Jenkins (1970) considered a multiplicative seasonal model, ARIMA
(0,1,1) x (0, 1,1)12,for this time series. The model can be written as:
( 1 - B)(I - B ’ ~ ) z=, (1 - elB)(i - BaB”)at (31)
Using the first 132 observations, the exact maximum likelihood estimates of model parameters
in the above equation are listed in the third column of Table VI and the parameter and outlier
estimates based on the joint estimation procedure of Chen and Liu (1993) are listed in the

f49 f 51 ’53 f 55

12 24 36 48 GO 72 84 96 108 120 132 144

Figure 1. Airline passenger data of Box and Jenkins (1970): natural logarithm of monthly totals (in
thousands) for the period 1/1949 to 12/1960
30 Journal of Forecasting Vol. 12, Iss. No. 1

Table VI. Estimates of model parameters and outlier effects for the airline passenger
data using ARIMA model (31)

Estimates and their t-values

Parameter With outlier adjustment Without outlier adjustment

01 0.3180 (3.63) 0.3488 (4.05)

02 0.4824 (6.24) 0.5624 (7.65)
Ua 0.0332 0.0362

Time Type Estimate t-value

29 A0 0.095 4.08
Outliers 54 -3.55
LS -0.097
detected 62
A0 -0.080 -3.44

second column of the table. The critical value 3.0 is used for outlier detection in the joint
estimation procedure.
As shown in Table VI, we find that three outliers are detected, none of them near the
forecast origin ( t = 132). By adjusting for these outliers, & is reduced by about 9%. Using the
estimated model shown in Table VI we can compute one-step-ahead forecasts for the forecast
origins 132 through 143. During the forecasting period, the critical value 2.5 is used for outlier
detection. There is only one outlier (an additive outlier at t = 135) detected during the post-
sample period. The estimate of the outlier effect and its t-value are -0.093 and -2.79
respectively. It is important to note that this outlier will not be detected if the critical value
for outlier detection is still set to 3.0 (see Hillmer, 1984). Assuming that the outlier is
‘correctly’ specified as an A 0 during forecasting, the one-step-ahead forecasts with outlier
adjustment are listed in Table VII. The forecasts based on traditional ARIMA model without
outlier adjustment are also listed in the Table.
Based on the results presented in Table VII, we obtain the quadruplet (RMSE,, RMSE,,
RMSE,, RMSE,) for forecasting with outlier adjustment to be (0.0343, 0.0215, 0.0927,

Table VII. Forecasts of the airline data in the post-sample period

~ ~ ~ _ _ _ ~

Without outlier adjustment Without outlier adjustment

Actual One-step-ahead Forecast One-step-ahead Forecast
t value forecast error forecast error

133 6.0331 6.0410 - 0.0079 6.0386 - 0.0055

134 5.9687 5.9846 -0.0159 5.9851 - 0.0164
135” 6.0379 6.1306 - 0.0927 6.1311 - 0.0932
136 6.1334 6.1037 0.0297 6.0440 0.0894
137 6.1570 6.1715 -0.0145 6.1429 0.0141
138 6.2823 6.3052 - 0.0229 6.2971 -0.0148
139 6.4329 6.4214 0.01 15 6.4160 0.0169
140 6.4069 6.4446 - 0.0377 6.4397 - 0.0328
141 6.2305 6.2343 - 0.0038 6.2391 - 0.0086
142 6.1334 5.1018 0.0316 6.1030 0.0304
143 5.9661 5.9960 - 0.0299 5.9945 - 0.0284
144 6.0684 6.0810 - 0.0126 6.0825 -0.0141

” A n A 0 occurs at t = 135 with t = 2.79.

C . Chen and L.-M. Liu Forecasting Time Series with Outliers 31

0.0297), while that for forecasting without outlier adjustment to be (0.0416 ,0.0202, 0.0932,
0.0894). Comparing these two sets of RMSEs, we find that the gross RMSE for the forecasts
with outlier adjustment is about 17.5% less than that without outlier adjustment, which is
almost entirely caused by the reduction of forecast error for t = 136. The RMSE, and RMSE,
for forecasting with and without adjustment are about the same in this case. From the above
RMSEs, we observe that the decomposed RMSEs in general provide more informative
knowledge on forecasting accuracy than the gross RMSE alone.
We also examine the impact on forecasts if the outlier type at t = 135 is not specified
correctly. Since in this case the outlier at t = 135 is always detected as an A 0 as soon as the
data at t = 136 are available, the mis-specification of the outlier type at t = 135 will only affect
the one-step-ahead forecast for t = 136. The rest of forecasts are the same as those shown in
Table VII. The one-step-ahead forecast for t = 136 is 6.0582 if the outlier at t = 135 is assumed
to be a TC, 6.0404 if the outlier is assumed to be an I 0 (based on the parameter estimates with
outlier adjustment) and 6.0110 if the outlier is assumed to be an LS. Based on these results,
we find that the forecast performance is similar to typical ARIMA forecasting without outlier
adjustment if the outlier type at t = 135 is mis-specified as a T C or an 10, and is worse than
typical ARIMA forecasting if the outlier type is mis-specified as an LS. Since an LS outlier
has a permanent impact on forecasts as shown in the third section of this paper, it should be
used with such understanding. In general, a judgmental decision is required to specify the
outlier type at the forecast origin. When the outlier type is appropriately specified, outlier
adjustment will improve the accuracy of forecasts.

Quarterly total imports of Taiwan, ROC

The second example is concerned with quarterly imports of goods and services of Taiwan in
constant billion New Taiwan Dollars (NT$). The data were obtained from the Directorate
General of Budget, Accounting and Statistics (DGBAS) of the Republic of China in Taiwan.
Figure 2 displays this series ( M , ) between the first quarter of 1968 and the fourth quarter of
1990. The imports of goods and services of a country are generally influenced by a number
of economic variables, in particular, the value of its currency (i.e. the exchange rate). The
currency of Taiwan (NT$) appreciated greatly with respect to US dollars (US$) since 1987, and
has stimulated Taiwan’s imports of goods and services since then. This impact of the exchange
rate on Taiwan’s imports can be clearly observed in Figure 2. In the following analyses, the
data between 1/68 and 4/87 are used for model building, and the data between 1/88 and 4/90
are reserved for post-sample forecasting comparison.
Following the model identification procedure outlined in Box and Jenkins (1970), the
following seasonal ARIMA model appears to be appropriate for the import series
(1 - B)(1 - B4)Mt = (1 - 8B4)ar (32)
Using the first 80 observations, the exact maximum likelihood estimates of model parameters
of equation (32) are listed in the third column of Table VIII, and the parameter and outlier
estimates based on the joint estimation procedure of Chen and Liu (1993) are listed in the
second column of the table. A critical value of 2.5 is used for outlier detection in the joint
estimatiori procedure since the length of the series is short.
Using model estimation with outlier detection, we find two outliers as shown in Table VIII.
The first outlier is an A 0 which occurred at t = 64. The second is an I 0 at t = 78. Under model
(32), an I 0 implies an annual trend change. This annual trend change was mainly caused by
the drastic appreciation of NT$ versus US$. By adjusting for these outliers, & is reduced by
about 10%.
32 Journal of Forecasting Vol. 12, Iss. No. I

'68 '70 '72 '74 '76 '78 '80 '82 '84 '86 '88 '90
5601 1 " ' ' ' ' ' ' 1 ' ' ' ' ' " ' ~ ' ' ' '

420

280

140

Figure 2. Quarterly total imports (in constant billion NT$) of Taiwan, ROC, for the period 111968 to
41 1990

Using the parameter estimates shown in Table VIII, we compute one-step-ahead forecasts
for the forecast origins between t = 81 and t = 91. The critical value 2.0 is used for outlier
detection during the post-sample period if outlier adjustment is employed. For forecasting with
outlier adjustment, we consider the outlier types during the forecasting period can be (1) A 0
and I 0 only; (2) AO, I 0 and TC; (3) AO, I 0 and LS; and (4) AO, 10, TC and LS. The results
of one-step-ahead forecasts under each scenario are listed in Table IX. In the process of
forecasting, we first assume that outliers occurred at the forecast origins are all additive
outliers. Such a choice is considered t o be conservative in the sense that we d o not expect an
outlier to affect future observations. The results obtained by assuming innovational outliers at
the forecast origins are discussed later.
When only A 0 and I 0 are allowed during forecasting period, we find that the forecast
origins at t = 82 and t = 85 are identified as outliers. These two outliers are both identified as
additive outliers when the entire series is estimated with outlier adjustment. Hence in the
computation of RMSE, and RMSE, for different models and outlier types, we treat these two
points as true outliers. The results of RMSEs for different scenarios are also listed in Table
IX. From these results, we find that RMSEs of one-step-ahead forecasts with outlier
adjustment are generally smaller than those generated by typical ARIMA forecasting. The only
exception is the case that all outlier types (AO, 10, TC and LS) are allowed during the
forecasting period.

Table VIII. Estimates of model parameters and outlier effects for Taiwan import data
using ARIMA model (32)
~

Estimates and their t-values

~~ ~ ~

Parameter With outlier adjustment Without outlier adjustment

0 0.7394 (9.38) 0.6879 (6.96)

0, 1 1.4540 12.7124

Time Type Estimate t-value

Outliers 64 A0 20.133 2.12
detected 78 10 42.819 3.14
C. Chen and L.-M. Liu Forecasting Time Series with Outliers 33

Table IX. One-step-ahead forecasts of Taiwain import data in the post-sample period

No Forecasts with outlier adjustment

Actual outliers
Time value allowed AO, I 0 AO, ro, TC AO, 10,LS AO, 10,TC, LS

81 392.060 381.2888 382.8683 382.8683 382.8683 382.8683

82 394.993 428.3065 425.3440 425.3440b 425.3440 425.3440
83 432.538 396.622 1 427.2278 a 397.1408 396.8768 396.8768
84 441.280 439.043 5 435.7460 435.7460 435.7460 435.7460
85 416.387 445.6038 446.2357 446.23 57 446.2357b 446.2357b
86 468.695 442.2381 480.21 13= 480.21 13 470.9777 427.1635
87 479.167 481.5315 47 1.4624 41 1.4624 47 1.4624 47 1.4624
88 472.004 486.3705 483.8389 488.6172 483.8389 488.61 72
89 466.250 467.2108 475.5242 415.5242 475.5242 475.5242
90 489.856 500.3570 498.6801 498.680 1 498.6801 498.680 1
91 475.215 501.9548 494.9600 494.9600 494.9600 494.9600
92 495.769 477.9352 476.8002 476.8002 476.8002 476.8002

RMSE, 21.273 16.333 19.497 18.967 22.680

RMSE, 13.589 12.390 13.058 12.390 13.058
RMSE, 31.332 30.101 30.101 30.101 30.101
RMSE, 31.543 8.967 26.321 25.268 38.708

"The one-step-ahead forecast for t = 83 is 396.8767, and that for t = 86 is 450.3643, if the outliers at t = 82 and t = 85
are assumed to be 10.
The corresponding observations are not identified as outliers when they are used as forecast origins, but are identified
as outliers later.

In the case when only A 0 and I 0 are allowed during the forecasting period and all outliers
at the forecasting origins are treated as AOs we obtain the most impressive improvement for
all types of RMSEs, particularly RMSEn. If we specified the outlier type at t = 82 and t = 85
incorrectly as an I 0 (rather than an AO), then RMSEn and RMSE, are increased to 28.353 and
19.681, respectively, and RMSE, and RMSE, remain the same. This result indicates the
importance of specifying a correct outlier type when an outlier occurs at the forecast origin.
However, in this case, forecasting with outlier adjustment generally produces better results
than typical ARIMA forecasting even if the outliers at the forecast origins are specified
incorrectly.
The inferior results produced by other cases of forecasting with outlier adjustment are
mainly caused by the difficulty in determining the locations and the types of outliers. As a
result, the outliers at t = 82 or t = 85 may not be detected and the forecasts for t = 83 and t = 86
are not improved. In addition, the locations and the types of outliers often change as the
forecast origin increases. These results are consistent with our findings in the simulation studies
of the fourth section of this paper. Some of the difficulties arise since (1) a TC or LS outlier
needs more data for the determination of its type and estimate, and cannot be obtained
accurately at the end of the series; and (2) depending upon the model, a TC or LS may not
be distinguishable from an 10. As a result of these complications, it is important to limit the
types of outliers allowed during the forecasting period.

SUMMARY AND CONCLUSIONS

In this paper we investigate the effects of outliers on forecasting. Four types of outliers are
considered. It has been shown that an outlier occurring at the forecast origin has the greatest
34 Journal of Forecasting Vol. 12, Iss. No. 1

impact on forecasts. As an outlier occurs further away from the forecast origin, its effect on
forecasts becomes smaller. We have also shown analytically that outlier adjustment can
improve forecast accuracy if the effect of the outlier is greater than one standard deviation of
the innovation series. Furthermore, we have demonstrated through a simulation study that the
power of outlier detection is reduced when the outlier occurs at the end of the series (i.e. the
forecast origin). Based on these results, we have proposed a strategy for forecasting time series
with outliers.
We have also investigated the potential loss of forecast accuracy due to the misidentification
of an outlier type. Generally, forecasting with outlier adjustment seems to perform better than
forecasting without outlier adjustment even though sometimes outlier type may be
misidentified. Based on the simulation study, we find some potential difficulty in the
identification of outlier types at the end of a series. Further study is needed in this area.
We have proposed four different post-sample RMSEs to reflect the forecast performance of
a model with and without the consideration of outlier adjustment. These measures provide us
with more informative ways to evaluate the forecasting performance of a model. The proposed
forecasting approach is applied to two real examples in which some practical forecasting issues
are discussed and the usefulness of the proposed approach is demonstrated.

ACKNOWLEDGEMENTS

We wish to thank Gregory B. Hudak and John L. Harris for their helpful comments and
suggestions. This work was supported in part by the School of Management Research Funds
of Syracuse University and Scientific Computing Associates.

REFERENCES

Abraham, B. and Box, G. E. P., ‘Bayesian analysis of some outlier problems in time series’, Biometrika,
66 (1979), 229-36.
Box, G. E. P. and Jenkins, G. W., Time Series Analysis: Forecasting and Control, San Francisco:
Holden Day, 1970.
Chang, I., Tiao, G. C. and Chen, C., ‘Estimation of time series parameters in the presence of outliers’,
Technometrics, 30 (1988), 193-204.
Chen, C. and Liu, L.-M., ‘Joint estimation of model parameters and outlier effects in time series’, To
appear in the Journal of the American Statistical Association (1993).
Chen, C., Liu, L.-M. and Hudak, G. B., Outlier Detection and Adjustment in Time Series Modelling
and Forecasting, Scientific Computing Associates, PO Box 625, DeKalb, Illinois 601 15, USA, 1990.
Chen, C. and Tiao, G. C., ‘Random level shift time series models, ARIMA approximation, and level
shift detection’, Journal of Business and Economic Statistics, 8 (1990), 170-86.
Chuang, A. and Abraham, B., ‘Comparison of parameter estimation methods in time series with outliers:
a simulation study’, ASA Proceedings, Business & Economic Statistics Section, pp. 83-92, 1989.
Fox, A. J., ‘Outliers in time series’, Journal of the Royal Statistical Society, Series B, 34 (1972), 350-63.
Guttman, I. and Tiao, G . C., ‘Effect of correlation on the estimation of a mean in the presence of
spurious observations’, Canadian Journal of Statistics, 6 (1978), 229-47.
Hillmer, S. C., Bell, W. R. and Tiao, G. C., ‘Modelling considerations in the seasonal adjustment of
economic time series’, Applied Time Series Analysis of Economic Data, 74-100, Washington, DC: US
Bureau of the Census, 1983.
Hillmer, S. C., ‘Monitoring and adjusting forecasts in the presence of additive outliers’, Journal of
Forecasting, 3 (1984), 205-15.
Ledolter, J . , ‘The effect of additive outliers on the forecasts from ARIMA models’, International
Journal of Forecasting, 5 ( 1 989), 23 1-40.
C. Chen and L.-M. Liu Forecasting Time Series with Outliers 35

Liu, L.-M., Hudak, G., Box, G. E. P . , Muller, M. E. and Tiao, G. C., The SCA Statistical System:
Reference Manual for Forecasting and Time Series Analysis, Scientific Computing Associates, P O Box
625, DeKalb, Illinois 60115, USA, 1986.
Liu, L.-M., ‘Dynamic relationship analysis of U.S. gasoline and crude oil prices’, Journal of Forecasting,
10 (1991), 521-47.
Liu, L.-M. and Lin, M.-W., ‘Forecasting residential consumption of natural gas using monthly and
quarterly time series’, International Journal of Forecasting, 7 (1991), 3-16.
Martin, D., ‘Robust methods for time series’, in Findley, D. F. (ed.), Applied Time Series Analysis I f ,
pp. 683-759, New York: Academic Press, 1981.
Smith, A. F. M. and West, M., ‘Monitoring renal transplants: an appIication of the multiprocess Kalman
filter’, Biometrics, 39 (1983), 867-78.
Tsay, R. S., ‘Outliers, level shifts, and variance changes in time series’, Journal of Forecasting, 7 (1988),
1-20.

Authors’ biographies:
Chung Chen is Associate Professor of Managerial Statistics of Department of Quantitative Methods at
the Syracuse University. He received his PhD from the University of Wisconsin-Madison in 1984. His
research interests include time-series modeling, forecasting, detecting outliers and structural changes,
causality testing and statistical procedures in decision support. Some of his publications appear in Review
of Economics and Statistics, Technometrics, Journal of Business and Economic Statistics, and Journal
of Accounting and Economics.
Lon-Mu Liu is Professor of Statistics, Department of Information and Decision Sciences, University of
Illinois at Chicago. H e has a PhD in Statistics from the University of Wisconsin-Madison, 1978. He is
one of the principal developers of the SCA Statistical System, and has published papers in Journal of
American Statistical Association, Journal of Business and Economic Statistics, Journal of Forecasting,
International Journal of Forecasting, Journal of Econometrics, Management Science, Communications
in Statistics, and others. His current research interest includes time-series analysis and forecasting,
econometric modeling, and computer software engineering. He was elected an IS1 member in 1988.

Authors’ address:
Chung Chen, Department of Quantitative Methods, School of Management, Syracuse University,
Syracuse, NY 13244-2130, USA.
Lon-Mu Liu, Department of Information and Decision Sciences, College of Business Administration
(M/C 294), The University of Illinois at Chicago, Box 4348, Chicago, IL 60680, USA.