Expert Systems with Applications 38 (2011) 13810–13815

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

Robust control charts for time series data

Christophe Croux a,⇑, Sarah Gelper b, Koen Mahieu c
a
K.U.Leuven & Tilburg University, Naamsestraat 69, B3000 Leuven, Belgium
b
Rotterdam School of Management, P.O. Box 1738, 3000 DR Rotterdam, The Netherlands
c
K.U.Leuven, Naamsestraat 69, B3000 Leuven, Belgium

a r t i c l e i n f o a b s t r a c t

Keywords: This article presents a control chart for time series data, based on the one-step-ahead forecast errors of
Control chart the Holt–Winters forecasting method. We use robust techniques to prevent that outliers affect the esti-
Holt–Winters mation of the control limits of the chart. Moreover, robustness is important to maintain the reliability of
Non-stationary time series the control chart after the occurrence of alarm observations. The properties of the new control chart are
Outlier detection
examined in a simulation study and on a real data example.
Robustness
Statistical process control
Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction individual data. Since the control chart should be resistant to out-
liers, a robust version of both the Holt–Winters forecasting proce-
Control charts are used to detect anomalies in processes. They dure and the X-chart needs to be used.
are most often used to monitor production-related processes. Several types of control charts for time series data have been
Samples taken from such processes are assumed to be independent proposed in the literature. To deal with serial correlation in sta-
and identically distributed. Many business-related processes, for tionary processes ARMA-charts (Jiang, Tsui, & Woodall, 2000) were
instance sales volumes or product prices, behave very differently. proposed, or data transformations (Wang, 2005). CuScore charts
They typically contain a trend, local or global, and serial correlation. (Box & Ramírez, 1992) can be of use for non-stationary series;
In this paper we propose a control chart aimed at detecting aberrant see Nembhard and Changpetch (2007) for a recent application.
observations, or outliers, in such time series. The control chart Vander Wiel (1996) proposed control charts for processes that
needs to be robust, meaning that the presence of outliers in the ser- wander, based on integrated moving average models. The Holt–
ies should not harm its performance. Outliers may be present in the Winters method is an important example of this approach. As
‘‘training period’’, being the part of the series used to determine the the other proposals, the approach of Vander Wiel (1996) is sensi-
control limits, or in the ‘‘test period’’, being the part of the series tive to outliers in the training sample and test sample. This prob-
where outliers should be flagged as alarm observations. lem is remediated by following the robust approach taken in this
The control chart we propose belongs to the class of special- paper.
cause control charts (Alwan & Roberts, 1988), where forecast errors Robust versions of the standard X and R chart are given in Rocke
of a prediction method are subject to regular control chart tech- (1989, 1992) and Tatum (1997), where the mean and scale of the
niques. If an observation has a large deviation from its predicted process are estimated robustly for setting up the control limits.
value, an unexpected event occurred, and an alarm should be More recently, robust control charts for multivariate processes
given. A large difference between the observed and predicted value were proposed, see Vargas (2003), Alfaro and Ortega (2009), and
implies an unusually large forecasting error, and the corresponding Stefatos and Hamza (2009). This paper contributes to this litera-
observation will be flagged as an outlier in the control chart of the ture by providing a robust control chart for time series typically
forecast errors. In this paper, we apply the Holt–Winters forecast encountered in business environments, containing stochastic
method, which is a widely used and simple procedure to forecast trends, outliers, and strong serial correlation.
time series containing trends and serial correlation. Moreover, it In Section 2 control charts based on the Holt–Winters method
is known to yield good forecast performance for many types of are briefly explained, while Section 3 describes the robust version.
times series encountered in business and industry (see Gardner The Holt–Winters method, both the standard and the robust ver-
(2006) & De Gooijer & Hyndman (2006) for recent applications). sion, has the advantage of being easy to implement by means of
Every single forecast error is then monitored on an X-chart for a simple updating scheme. Section 4 presents a simulation study
that demonstrates the advantages of a robust approach. First, the
⇑ Corresponding author. robust version gives reliable control limits in cases where the
E-mail address: [email protected] (C. Croux). training sample contains outliers. Secondly, if an isolated outlier

0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2011.04.184
 C. Croux et al. / Expert Systems with Applications 38 (2011) 13810–13815 13811

occurs in the test sample, the non-robust control chart might yield with za/2 the (1  a/2)-quantile of the standard normal distribution.
a sequence of false alarms right after the occurrence of the outlier. For every new observations yt in the test period, where t > n, one
The robust approach does not suffer from this drawback. Section 5 updates the recursion relations (2) and computes the new predic-
presents an application with real data. Finally, some conclusions tion error et as in (1). The value of et is then plotted in the control
are made in Section 6. chart. If it falls out of the control limits, then the observed value
deviates substantially from the predicted value, indicating an
2. Control charts using the Holt–Winters method unexpected change in the process. Note that the observations in
the test period influence future prediction errors via the Holt–Win-
In this section, we construct a control chart for a time series {yt}. ters forecast formulas in (2), but do not alter the control limits.
At every time point we make a one-step-ahead forecast by the It is of crucial importance that the process is completely in con-
Holt–Winters forecasting algorithm. The one-step-ahead forecast trol during the training period, otherwise the scale estimate S is in-
errors of this procedure are plotted on a control chart. Control lim- flated by the presence of outliers in the training period. If outliers
its to monitor new incoming observations are constructed from the are present in the training sample, it is necessary to follow a robust
training sample. approach, as discussed in the next section.

2.1. The forecasting algorithm 3. Control charts using the robust Holt–Winters method

Consider a time series observed up to time point t  1. The The control chart presented in the previous section is sensitive
Holt–Winters method makes a prediction of the series at time t, to outlying observations in several ways. First, the Holt–Winters
denoted by y^t jt  1. Once the true value yt is observed, the one- forecasts are adversely affected by outliers, since the estimated lo-
step-ahead forecast error is given by cal level and trend in (2) depend on past values of the observed ser-
^tjt1 :
et ¼ yt  y ð1Þ ies, and thus also on possible outliers. Second, if outliers are
present in the training sample, the selected smoothing parameters
Denote the local level of the series by {at} and the local trend by k1 and k2 might be biased. Finally, due to an inflated scale of the
{bt}. The Holt–Winters algorithm estimates these unknown values forecast errors, outliers may lead to wider control limits.
by

a^ t ¼ k1 yt þ ð1  k1 Þða^ t1 þ b^t1 Þ; 3.1. The forecasting algorithm

ð2Þ
^ t ¼ k2 ða
b ^t  a
^ t1 Þ þ ð1  k2 Þb^t1
The shortcomings of the standard Holt–Winters algorithm have
resulting in the forecast been addressed by Gelper, Fried, and Croux (2009). They present a
robust version of Holt–Winters method, consisting of the applica-
^tjt1 ¼ a
y ^t1 :
^ t1 þ b ð3Þ tion of the standard procedure on a cleaned version of the observed
series. Furthermore, they stress the importance of a robust selec-
The parameters k1 and k2 in (2) are smoothing parameters, with val-
tion of the smoothing parameters and a robust estimation of the
ues between zero and one. The larger the smoothing parameters,
starting values of the algorithm. We briefly review their approach.
the less the local level and/or local trend series are smoothed, and
A cleaned version yt of the time series yt is obtained as
the more weight the current value of yt has on the prediction of
!
the next value. yt  a ^t1
^ t1  b
The recursive relations (2) are started up after a short startup yt ¼ wk r^ t þ a^ t1 þ b^t1 ; ð6Þ
r^ t
period of length m. A linear trend line is fitted to the data in the
startup period for estimating the local level and trend at t = m. where wk(y) = max (k, min (y, k)) is the Huber w-function with
The smoothing parameters k1 and k2 are selected by minimizing boundary value k, and r^ t is a local scale estimate of the one-step-
the sum of squared forecast errors over a training period of length n: ahead forecast errors. We set k = 2, such that the influence of fore-
X
n cast errors larger than two times the local scale estimate is
ðkopt opt
1 ; k2 Þ ¼ argminðk1 ;k2 Þ
^tjt1 Þ2 :
ðyt  y ð4Þ bounded. Note that for k tending to infinity, one gets the standard
t¼mþ1 Holt-Winters method again. The local scale estimate is computed
Note that the observations in the startup period are not used for recursively as
!
determining the smoothing parameters. ^t1
yt  a
^ t1  b
r^ 2t ¼ kr qk r^ 2t1 þ ð1  kr Þr^ 2t1 ; ð7Þ
r^ t1
2.2. Control chart
where qk() is a bounded loss function with boundary value k = 2,
As motivated in Section 1, the one-step-ahead forecast errors see Gelper et al. (2009). Forecast errors larger than two times the
et ¼ yt  y ^t jt  1 are monitored to detect anomalies in the outcome local scale estimate r
^ t1 will have a bounded influence on the next
of the process. We proceed as for a regular X-chart, and assume scale estimate. We allow for a slowly varying scale, which is trans-
normality of the forecast errors. The control limits are computed lated in a choice of kr rather close to zero, for example kr = 0.3.
from the forecast errors et in the training period, with t = m + 1, The recursive scheme of the robust Holt–Winters algorithm
. . . , n. Denote S2 the mean of the squared prediction errors in the then becomes, analogous to (2),
training period, excluding the startup values:
Pn 2 a^ t ¼ k1 yt þ ð1  k1 Þða^ t1 þ b^t1 Þ;
t¼mþ1 et ð8Þ
S2 ¼ : ^t ¼ k2 ða ^t  a^ t1 Þ þ ð1  k2 Þb^t1 ;
nm b
Since the target value for the forecast errors is zero, the following where yt follows Eq. (6).
(1  a) Upper and Lower Control Limits are obtained: There are two further issues that need to be addressed. First, the
startup values a^ m and b ^m need to be computed using a robust
UCL ¼ za=2  S;
ð5Þ regression fit to the data in the startup period. For this robust fit
LCL ¼ za=2  S;
the repeated median estimators is used. The median absolute
13812 C. Croux et al. / Expert Systems with Applications 38 (2011) 13810–13815

deviation of the residuals with respect to this robust fit yields the rate is like the size of the control charts, but for a process that is not
starting value r
^ 2m for the scale estimates. The second issue is the ro- fully under control.
bust selection of the smoothing parameters. Instead of minimizing
the sum of squares of the one-step-ahead forecast errors over the 4.1. Simulation design
training sample, we minimize a robust performance measure
Xn   We generate series {yt} from a local linear trend model. This
  2 ^tjt1
yt  y
kopt
1 ; kopt
2 ¼ argminðk1 ;k2 Þ 0 qk : ð9Þ model generates non-stationary and correlated time series, and al-
t¼mþ1
s0 lows for trends. It is a reasonable model for many series encoun-
tered in the practice of business and economics. At this model
This performance measure corresponds to a s-scale estimator, see
the forecast errors of the Holt–Winters method form an i.i.d. se-
for example Maronna, Martin, and Yohai (2006), which combines
quence from a normal distribution (e.g. Chatfield & Yar (1988)).
a high robustness with a high statistical efficiency. The auxiliary
The local linear trend model is defined as
scale estimate s0 in (9) is given by s0 ¼ medm þ 1 6 t 6 njyt   
^t jt  1j, the median absolute deviation of the prediction errors.
y Y t ¼ at þ et ; et  N 0; r2e ; ð13Þ
The loss function used in (9) is
where at, the local level, and bt, the local trend, are given by
qk ðyÞ ¼ minðk2 ; y2 Þ ¼ wk ðyÞ2 ; ð10Þ 8  
< a ¼ a þ b þ g ; g  N 0; r2 ;
t t1 t1 t t g
ð14Þ
with k = 2. : b ¼ b þ m ; m  N0; r2 :
t t1 t t m

3.2. Control chart The noise terms et, gt and mt are independent, and serially uncorre-
lated. In the simulation study, we take re = 1 and rg = rm = 0.1, lead-
Similar as in Section 2.2, we plot the one-step-ahead forecast ing to a smaller variance for the local level and trend compared to
^t jt  1, for t > n, in the control chart, but now with
errors et ¼ yt  y the variance in the measurement Eq. (13).
forecasts obtained from the robust Holt–Winters algorithm. The The first n observations from a generated series constitute the
control limits are determined from a robust scale estimate of the training sample, including a startup period of length m = 10. The
one-step-ahead forecast errors from the training period. We use next 200 generated observations constitute the test sample. From
again the s-estimator of scale the training data the control limits are constructed as outlined in
Xn   Sections 2 and 3. Outliers, hence alarm observations, are induced
s20 e
s2 ¼ c k qk t ; ð11Þ in the test sample by adding the value kre to 10% of randomly se-
n  m t¼mþ1 sT lected observations in the test sample. For simulating the power,
we take k = 5, corresponding to outliers of moderate size. To study
where s0 = medm+16t6njetj, and qk as in (10). The consistency factor
the robustness of the size and power with respect to outliers in the
ck in the above equation ensures that s2 is a consistent estimator
training sample, the outliers are generated in the same way as for
of the population variance of a normal distribution. For k = 2, we
the test sample. The percentage of outliers in the training data are
have ck = 1.404. We construct approximate (1  a) control limits as
taken as 0% (clean), 2%, and 5%, respectively.
UCL ¼ za=2  s;
ð12Þ 4.2. Simulation results
LCL ¼ za=2  s;

with za/2 as the (1  a/2)-quantile of the standard normal distribu- The simulated power and size are summarized in Table 1 for the
tion. Hence, the formula for the control limits (5) and (12) are sim- control charts based on the standard and the robust Holt–Winters
ilar, except that the robust method uses a robust s-scale estimator
instead of a standard deviation (centered at zero). The use of the s- Table 1
scale estimator guarantees the resistance to outliers in the training Size and power of the control chart, with a = 5%, for the standard (HW) and the robust
Holt–Winters (RHW) procedures, for different sizes of the training sample (50 and
period.
100), and for different percentages of outliers in training sample.
In the next section a simulation study is carried out to verify
whether the constructed control limits for the standard and the ro- n Clean 2% 5%
bust Holt–Winters method have comparable properties if the HW RHW HW RHW HW RHW
training sample is clean. In the presence of outliers in the training Size
sample, we expect the robust method to perform better. 50 .073 .086 .044 .073 .026 .067
100 .059 .063 .027 .052 .009 .037

4. Simulation study Power

50 .903 .900 .843 .874 .784 .850
In this simulation study we take a closer look at the size, the 100 .903 .902 .842 .881 .757 .853
power and the false detection rate of the proposed robust control
chart, and compare them to the same characteristics of the non-ro-
bust version. The size of a control chart is the probability that an
Table 2
observation in the test sample falls out of the control limits when False detection rate of the control chart, with a = 5%, for the standard (HW) and the
the process is actually in control, i.e. when no outliers are present robust Holt-Winters (RHW) algorithm, and for different lengths n of the training
in the test sample. It is supposed to be approximately 5%, when sample. The training data are clean, and outliers are created in the test data by
taking a = 0.05 in (5) and (12). The power is the probability that shifting 10% of the observations upwards by k units.
an alarm observation is detected. In a simulation study, the power Shift HW RHW
equals the proportion of generated outliers in the test sample that size
k=5 k = 10 k = 15 k = 20 k=5 k = 10 k = 15 k = 20
are detected. The percentage of observations in the test sample
n = 50 0.097 0.167 0.206 0.232 0.084 0.088 0.083 0.085
that are not generated as outliers, but that are falsely identified
n = 100 0.093 0.161 0.203 0.229 0.079 0.080 0.081 0.082
as alarm observations, is the false detection rate. The false detection
 C. Croux et al. / Expert Systems with Applications 38 (2011) 13810–13815 13813

Standard H−W
a

2.5
Housing starts
1.5
0.5
1.0
forecast errors
0.0
−1.0
−2.0

1970 1975 1980 1985 1990 1995

Robust H−W
b
2.5
Housing starts
1.5
0.5
1.0
forecast errors
0.0
−1.0
−2.0

1970 1975 1980 1985 1990 1995

c Standard H−W
2.5
Housing completions
1.5
0.5
0.0 0.2 0.4
forecast errors
−0.4

1970 1975 1980 1985 1990 1995

d Robust H−W
2.5
Housing completions

1.5
0.5
0.0 0.2 0.4
forecast errors

−0.4

1970 1975 1980 1985 1990 1995

Fig. 1. In each graph, the upper panel presents the observed data (dots) and the smoothed series (solid line), while the lower panel presents the control chart, with forecast
errors (points) and control limits (dashed lines). The housing starts series – containing outliers in the training period – is analyzed with standard Holt–Winters in (a) and with
robust Holt–Winters method in (b). Similarly for the housing completions series. This series does not contain outliers in the training period, explaining the resemblance
between (c) and (d).
13814 C. Croux et al. / Expert Systems with Applications 38 (2011) 13810–13815

algorithm. The reported numbers are averages over 1000 simula- We first take a closer look at the housing starts in Fig. 1(a) and
tion runs. Training samples of length n = 50 and n = 100 are consid- (b). The smoothing parameters (k1, k2) selected from the training
ered. The simulation study, although rather modest in the number period are (0.3, 0.2) for the standard method and (0.4, 0.2) for the
of considered simulation settings, is quite time consuming, since robust method. These parameters are reasonable, since they are
the optimal selection of the smoothing parameters, see (4) and close to zero and the parameter corresponding to the trend is
(9), needs to be performed for every single time series. smaller than the one for the level. The standard method tends to
As we see from Table 1 there is a small size distortion in the select smaller parameters than the robust version, since outliers
clean case, when no outliers are present in the training sample, may bias the parameters towards zero. The direct effect of the
and this for both methods. For the larger sample size, the simulated large outliers in the standard Holt–Winters case can be seen at
size is already much closer to 5%. In the presence of outliers in the the time points subsequent to the large outliers. The forecasts
training sample, the size of the non-robust chart is close to zero in for these time points are attracted by the outlier and therefore,
most cases, while the size of the robust method is kept much more in this case, the values are too small. The robust method clearly
stable. The breakdown of the size for the non-robust method is due yields better forecasts in the outlier’s subsequent time points. Also
to two reasons. The main reason is that the scale S in (5) is inflated in terms of mean squared one-step-ahead forecast errors (MSFE),
due to the outliers, resulting in too wide control limits. Further- computed over the test period, the robust method outperforms
more, the outliers cause a bias in the selection of the smoothing the standard one with an MSFE of 0.0144 against 0.0175. The con-
parameters, leading to a less optimal smoothing. trol chart shows that the control limits are inflated by the outliers
Consider now the power in Table 1. If the training sample is in the non-robust case, and all observations in the test sample re-
clean, the two methods have a comparable power to detect the main within the control bounds. This leads to a loss of power of
generated alarm observations. When 2% or 5% of outliers are pres- the non-robust method to detect outliers. For example, the alarm
ent in the training sample, they do affect the detection power of observation in January 1984 is only detected by the robust
the control charts, and we see that the robust Holt–Winters clearly method.
outperforms the standard one. The difference in power is about The housing completions in Fig. 1(a) and (b), do not show
10% in presence of only 5% of moderate outliers in the training important outliers in the training period. The selected smoothing
sample. When increasing k, the size of the outliers, and the per- parameters for this series are (0.5, 0.2) in the standard case and
centage of outliers in the training sample, these differences be- (0.6, 0.3) in the robust case. Similar to the housing starts, the stan-
come even much more pronounced. dard parameters are somewhat smaller than the robust ones. The
Next, we compare the false detection rate of the two control MSFE for both methods are comparable, with 6.01  103 in the
charts, see Table 2. We follow the same simulation scheme as be- standard case and 6.09  103 in the robust case. Since the series
fore, but only consider clean training data to ensure that the size of do not contain important outliers, both the standard and the robust
the two methods remains comparable. We vary the magnitude of control charts look similar. Nevertheless, the robust control chart
the outliers by taking k = 5, 10, 15 and 20. A first observation is that has slightly tighter control limits. The percentage of observations
the false detection rate does not seem to depend much on n. Fur- exceeding the control limits is 2% for the standard method, which
thermore, for k = 5 the robust method is only slightly better than seems overly optimistic, and 6.1% for the robust method, closer to
the non-robust method. For larger values of k, in contrast, the ro- the expected size of a = 5%.
bust procedure has a much lower false detection rate. This is due We conclude that it is worthwhile to use the robust method. If
to the following reason. For large values of k, an outlying observa- there are no major outliers, the performance of both methods is
tion in the test sample affects the outcome of the next forecast in quite similar. However, if major outliers are present, in particular
the Holt–Winters procedure. Hence it is likely to induce a large in the training period, the robust method is more reliable.
forecasting error for the next observation as well, even if the latter
one is not an outlier. The propagation effect of outliers does not oc-
cur with the robust approach, since it uses ‘‘cleaned observations’’ 6. Conclusion
in the update formulas (8).
In this paper, we propose to use the one-step-ahead forecast er-
rors of the Holt–Winters forecasting method as the input of a con-
trol chart for monitoring non-stationary processes. Holt–Winters
5. Real data example forecasting is a frequently used method for univariate time series
with a stochastic trend and a seasonality pattern. In this study,
We illustrate the proposed method by means of a real data exam- we did not explicitly deal with seasonality, but the method can
ple. We use the Housing Data from the book of Diebold (2001), con- easily be adapted for that purpose. The main advantages are that
taining monthly data for housing starts and completions from the method is very easy to implement and still performs well com-
January 1968 until June 1996. Fig. 1 shows both series and the cor- pared to other more complicated forecasting techniques. We com-
responding control charts obtained by the standard and the robust pared the control charts based on the standard Holt–Winters
Holt–Winters procedure. The upper panel of every graph shows method with a robust version.
the observed data yt, together with the smoothed series y ^tjt1 . The The simulation study shows that with clean training data, the
lower panel of every graph contains the control chart. As explained standard and robust control charts perform almost equally well
in Section 2, the prediction errors et are plotted in the control chart. in terms of size and power. On the other hand, when we add a
The vertical dashed lines in the chart indicate the end of the startup small amount of moderate outliers to the training sample the
period and of the training period. Control limits are indicated for the non robust method looses power, and suffers from severe size dis-
test period. We used a startup period of length m = 20, a training tortion. Even when the training data are clean, the false detection
period of length n = 150. The other 192 observations serve as the rate of the robust procedures is much lower if major outliers are
test sample. We immediately notice two large outliers, one near present in the test sample. Finally, the application to real data
1971 and another near 1977, in the training sample of the housing shows that the proposed robust method is easy to put in practice.
starts. Moreover, both time series may contain other smaller outli- The graphical displays in Fig. 1, representing both the observed ser-
ers. By applying robust exponential smoothing, we expect that the ies and the control chart in the same graph, are an easy-to-use vi-
results remain stable in the presence of these outliers. sual aid for managers to monitor business processes.
 C. Croux et al. / Expert Systems with Applications 38 (2011) 13810–13815 13815

There are several open questions, which we consider as areas Diebold, F. X. (2001). Elements of forecasting, second ed. South-Western.
Gardner, E. (2006). Exponential smoothing: The state of the art – part ii.
for future research. If a structural break appears in the time series,
International Journal of Forecasting, 22(4), 637–666.
the control chart should be readjusted and new control limits need Gelper, S., Fried, R., & Croux, C. (2009). Robust forecasting with exponential and
to be computed. Another limitation of the presented approach is Holt–Winters smoothing. Journal of Forecasting, 29, 285–300.
the normality assumption on the forecast errors (excluding outli- Jiang, W., Tsui, K.-L., & Woodall, W. H. (2000). A new SPC monitoring method: The
arma chart. Technometrics, 42(4), 399–410.
ers). Different distributional assumptions will lead to different for- Maronna, R. A., Martin, R. D., & Yohai, V. J. (2006). Robust statistics: Theory and
mula for the control limits. We refer to Yang and Chen (2009) and methods. Wiley.
Chen and Yeh (2009) for recent contributions to these problems, Nembhard, H. B., & Changpetch, P. (2007). Directed monitoring using cuscore charts
for seasonal time series. Quality and Reliability Engineering International, 23,
although in different settings. 219–232.
Rocke, D. M. (1989). Robust control charts. Technometrics, 31(2), 173–184.
References Rocke, D. M. (1992). X q and rq charts: Robust control charts. The Statistician, 41(1),
97–104.
Stefatos, G., & Hamza, A. B. (2009). Fault detection using robust multivariate control
Alfaro, J., & Ortega, J. (2009). A comparison of robust alternatives to Hotelling’s T2
chart. Expert Systems with Applications, 36(3), 5888–5894.
control chart. Journal of Applied Statistics, 36(12), 1385–1396.
Tatum, L. G. (1997). Robust estimation of the process standard deviation for control
Alwan, L. C., & Roberts, H. V. (1988). Time-series modeling for statistical process
charts. Technometrics, 39(2), 127–141.
control. Journal of Business & Economic Statistics, 6(1), 87–95.
Vander Wiel, S. (1996). Monitoring processes that wander using integrated moving
Box, G., & Ramírez, J. (1992). Cumulative score charts. Quality and Reliability
everage models. Technometrics, 38(2).
Engineering, 8(1), 17–27.
Vargas, J. A. (2003). Robust estimation in multivariate control charts for individual
Chatfield, C., & Yar, M. (1988). Holt–Winters forecasting: Some practical issues. The
observations. Journal of Quality Technology, 35(4).
Statistician, 37(2), 129–140.
Wang, F. (2005). A simple data transformation of auto-correlated data for SPC.
Chen, F. L., & Yeh, C. H. (2009). Economic statistical design of non-uniform sampling
International Journal of Production Research, 43(5), 981–989.
scheme X bar control charts under non-normality and Gamma shock using
Yang, S.-F., & Chen, W.-Y. (2009). Controlling over-adjusted process means and
genetic algorithm. Expert Systems with Applications, 36(5), 9488–9497.
variances using VSI cause selecting control charts. Expert Systems with
De Gooijer, J., & Hyndman, R. (2006). 25 years of time series forecasting.
Applications, 36(3), 7170–7182.
International Journal of Forecasting, 22(3), 443–473.