Content uploaded by Adriaan Brebels
Author content
All content in this area was uploaded by Adriaan Brebels on Mar 24, 2016
Content may be subject to copyright.
(
)
()m
t tt
e yf= −
()
m
ft
1,
1
1
,
n
ii
in
i
MAE e mean e
n=
=
= =
∑
1,
,
i
in
MdAE median e
=
=
() ()
22
11,
1
,
n
ii
iin
MSE e mean e
n==
= =
∑
World Applied Sciences Journal 24 (Information Technologies in Modern Industry, Education & Society): 171-176, 2013
ISSN 1818-4952
© IDOSI Publications, 2013
DOI: 10.5829/idosi.wasj.2013.24.itmies.80032
Corresponding Author: Shcherbakov, Volgograd State Technical University, Lenin avenue, 28, 400005, Volgograd, Russia.
171
A Survey of Forecast Error Measures
Maxim Vladimirovich Shcherbakov, Adriaan Brebels,
Nataliya Lvovna Shcherbakova, Anton Pavlovich Tyukov,
Timur Alexandrovich Janovsky and Valeriy Anatol’evich Kamaev
Volgograd State Technical University, Volgograd, Russia
Submitted: Aug 7, 2013; Accepted: Sep 18, 2013; Published: Sep 25, 2013
Abstract: This article reviews the common used forecast error measurements. All error measurements have been
joined in the seven groups: absolute forecasting errors, measures based on percentage errors, symmetric errors,
measures based on relative errors, scaled errors, relative measures and other error measures. The formulas are
presented and drawbacks are discussed for every accuracy measurements. To reduce the impact of outliers, an
Integral Normalized Mean Square Error have been proposed. Due to the fact that each error measure has the
disadvantages that can lead to inaccurate evaluation of the forecasting results, it is impossible to choose only
one measure, the recommendations for selecting the appropriate error measurements are given.
Key words: Forecasting Forecast accuracy Forecast error measurements
INTRODUCTION
Different criteria such as forecast error measurements,
the speed of calculation, interpretability and others have where - y is the measured value at time t, - predicted
been used to assess the quality of forecasting [1-6].
Forecast error measures or forecast accuracy are the most
important in solving practical problems [6]. Typically, the
common used forecast error measurements are applied for
estimating the quality of forecasting methods and for
choosing the best forecasting mechanism in case of
multiple objects. A set of "traditional" error measurements
in every domain is applied despite on their drawbacks.
These error measurements are used as presets in domains
despite on drawbacks.
This paper provides an analysis of existing and quite
common forecast error measures that are used in
forecasting [4, 7-10]. Measures are divided into groups
according to the calculating method an value of error for
certain time t. The calculating formula, the description of
the drawbacks, the names of assessments are considered
for each error measure.
A Review
Absolute Forecasting Error: The first group is based on
the absolute error calculation. It includes estimates based
on the calculation of the value ei
(1)
t
value at time t, obtained from the use of the forecast
model m. Hereinafter referred to as the index of the model
(m) will be omitted.
Mean Absolute Error, MAE is given by:
(2)
where n –forecast horizon, mean(•) – a mean operation.
Median Absolute Error, MdAE is obtained using the
following formula
(3)
where mean(•) – operation for calculation of a median.
Mean Square Error, MSE is calculated by the formula
(4)
() ()
22
11,
1n
ii
iin
RMSE e mean e
n==
= =
∑
t
t
t
e
py
=
()
1,
1
1100 100
n
ii
in
i
MAPE p mean p
n=
=
= ⋅= ⋅
∑
(
)
1,
100 i
in
MdAPE median p
=
= ⋅
()
2
1,
10
0,
i
in
RMSPE mean p
=
= ⋅
()
2
1,
10
0,
i
in
RMdSPE median p
=
= ⋅
()
.
t
t
tt
e
syf
=+
()
1,
1
1200 20
0,
n
ii
in
i
sMAPE s mean s
n=
=
= ⋅= ⋅
∑
(
)
1,
20
0.
i
in
sMdAPE median s
=
= ⋅
()
1
1
,
/2
n
ii
ii i
i
yf
msMAPE nyf S
=
−
=++
∑
11
11
,.
11
11
11
ii
S yy y y
ik k
ii
ii
kk
−−
= −=
−−
−−
= =
∑∑
World Appl. Sci. J., 24 (Information Technologies in Modern Industry, Education & Society): 171-176, 2013
172
hence, Root Mean Square Error, RMSE is calculated as:
(5) We note the following shortcomings.
These error measures are the most popular in various is equal to zero.
domains [8, 9]. However, absolute error measures have the Non-symmetrical issue - the error values differ
following shortcomings. whether the predicted value is bigger or smaller than
The main drawback is the scale dependency [9]. Outliers have significant impact on the result,
Therefore if the forecast task includes objects with particularly if outlier has a value much bigger then
different scales or magnitudes then absolute error the maximal value of the "normal" cases [4].
measures could not be applied. The error measures are biased. This can lead to an
The next drawback is the high influence of outliers in incorrect evaluation of the forecasting models
data on the forecast performance evaluation [11]. performance [15].
So, if data contain an outliers with maximal value
(this is common case in real world tasks), then Symmetric Errors: The criteria which have been included
absolute error measures provide conservative values. in this group are calculated based on the value:
RMSE,MSE have a low reliability: the results could
be different depending on different fraction of data
[4]. (11)
Measures Based on Percentage Errors: Percentage errors The group includes next measures. Symmetric
are calculated based on the value PMean Absolute Percentage Error, sMAPE is calculated
t
(6)
Also these errors are the most common in forecasting
domain. The group of percentage based errors includes and the median mean absolute percentage error
the following errors.
Mean Absolute Percentage Error, MAPE
(7) To avoid the problems associated with the division
Median Absolute Percentage Error, MdAPE is more has been proposed. Their denominators have an
resistant to outliers and calculated according to the additional member:
formula
(14)
(8)
Root Mean Square Percentage Error, RMSPE is
calculated according to:
(9)
and the median percentage error of the quadratic
(10)
Appearance division by zero when the actual value
the actual [12-14].
according to
(12)
(13)
by zero, a modified sMAPE - Modified sMAPE, msMAPE
where .
Developing the idea for the inclusion of an additional
terms, more sophisticated measures was presented [16]:
KL-N, KL-N1, KL-N2, KL-DE1, KL-DE2, IQR
()
*
,
t
t
tt
e
r
yf
=−
*
ft
*
t
tl
fy
−
=
1,
,
i
in
MRAE mean r
=
=
1,
,
i
in
MdRAE median r
=
=
1
2
.
1
1
t
tn
ii
i
e
q
yy
n−
=
=
−
−∑
1,
,
i
in
MASE mean q
=
=
(
)
2
1,
.
i
in
RMSSE mean q
=
=
*
,
MAE
RMAE
MAE
=
*
.
RMSE
RRMSE
RMSE
=
World Appl. Sci. J., 24 (Information Technologies in Modern Industry, Education & Society): 171-176, 2013
173
The following disadvantages should be noted. If naive model has been chosen then division by zero
Despite its name, this error is also non-symmetric
[13].
Furthermore, if the actual value is equal to forecasted
value, but with opposite sign, or both of these values
are zero, then a divide by zero error occurs.
These criteria are affected by outliers in analogous
with the percentage errors.
If more complex estimations have been used, the
problem of interpretability of results occurs and this
fact slows their spread in practice [4].
Measures Based on Relative Errors: The basis for
calculation of errors in this group is the value determined
as follows:
(15)
where - the predictive value obtained using a reference
model prediction (benchmark model). The main practice is
to use a naive model as a reference model
,
(16)
where l - the value of the lag and l = 1.
The group includes the next measures. Mean Relative
Absolute Error, MRAE is given by the formula
(17)
Median Relative Absolute Error, MRAE is calculated
according to
(18)
and Geometric Mean Relative Absolute Error, GMRAE),
which is calculated similarly to (17), but instead of mean(•)
the geometric mean is obtained gmean(•).
It should be noted the following shortcomings.
Based the formulas (15-18), division by zero error
occurs, if the predictive value obtained by reference
model is equal to the actual value.
error occurs in case of continuous sequence of
identical values of the time series.
Scaled Error: As a basis for calculating the value of the
scaled errors q is given by
i
(19)
This group contains Mean Absolute Scaled Error,
MASE proposed in [9]. It is calculated according to:
(20)
Another evaluation of this group is Root Mean
Square Scaled Error, RMSSE is calculated by the formula
[10]:
(22)
These measures is symmetrical and resistant to
outliers. However, we can point to two drawbacks.
If the forecast horizon real values are equal to each
other, then division by zero occurs.
Besides it is possible to observe a weak bias
estimates if you do the experiments by analogy with
[15].
Relative Measures: This group contains of measures
calculated as a ratio of mentioned above error measures
obtained by estimated forecasting models and reference
models. Relative Mean Absolute Error, RelMAE is
calculated by the formula.
(23)
where MAE and MAE the mean absolute error for the
*
analyzed forecasting model and the reference model
respectively, calculated using the formula (2).
Relative Root Mean Square Error, RelRMSE is
calculated similarly to (23), except that the right side is
calculated by (5)
(24)
*
lo
g.
RMSE
LMR
RMSE
=
{
}
(
)
*
( ) 100
%.
PB MAE mean I MAE MAE=⋅<
*
0
,;
() 1.
if MAE MAE
I MAE <
=
(
)
2
1,
1
,
i
in
nRMSE mean e
y=
=
y
(
)
2
1,
1
1
.
i
n
in
i
i
inRSE mean e
y=
=
=
∑
()
()
2
1
2
1
,
n
i
i
n
i
i
e
inRSE
yy
=
=
=
−
∑
∑
1
1
n
yy
k
n
k
=
=
∑
()
2
1
_1
n
i
i
e MAE
Std AE n
=
−
=−
∑
()
2
1
_1
n
i
i
p MAPE
Std APE n
=
−
=−
∑
World Appl. Sci. J., 24 (Information Technologies in Modern Industry, Education & Society): 171-176, 2013
174
In some situations it is reasonable to calculate the over the entire interval or time horizon or defined
logarithm of the ratio (23). In this case, the measure is
called the Log Mean Squared Error Ratio, (LMR)
(25)
Syntetos et al. proposed a more complex assessment
of the relative geometric standard deviation Relative
Geometric Root Mean Square Error, RGRMSE [17].
The next group of measures counts the number of
cases where the error of the model prediction error is
greater than the reference model. For instance, PB (MAE)
- Percentage Better (MAE), calculated by the formula:
(26)
where I{•} - the operator that yields the value of zero or
one, in accordance with the expression:
(27)
By analogy with PB (MAE), Percentage Better (MSE)
can be defined.
The disadvantages of these measures are the
following.
Division by zero error occurs if the reference forecast
error is equal to zero.
These criteria determine the number of cases when
the analyzed forecasting model superior to the base
but do not evaluate the value of difference.
Other Error Measures: This group includes measures
proposed in various studies to avoid the shortcomings of
existing and common measures.
To avoid the scale dependency, Normalized Root
Mean Square Error (nRMSE) has been proposed,
calculated by the formula:
(28)
where - the normalization factor, which is usually equal
to either the maximum measured value on the forecast
horizon, or the difference between the maximum and
minimum values. Normalization factor can be calculated
short interval of observation [18]. However, this
estimate is affected by influence of outliers, if outlier has
a value much bigger the maximal "normal" value. To
reduce the impact of outliers, Integral Normalized Mean
Square Error [19] have been proposed, calculated by the
formula:
(29)
Some research contains the the ways of NRMSE
calculation as [16]:
(30)
where .
Other measures are called normalized std_APE and
std_MAPE [20, 21] and calculated by the formula
(31)
and
(32)
respectively.
As a drawback, you can specify a division by zero
error if normalization factor is equal to zero.
Recommendations How to Choose Error Measures:
One of the most difficult issues is the question of
choosing the most appropriate measures out of the
groups. Due to the fact that each error measure has the
disadvantages that can lead to inaccurate evaluation of
the forecasting results, it is impossible to choose only one
measure [5].
World Appl. Sci. J., 24 (Information Technologies in Modern Industry, Education & Society): 171-176, 2013
175
We provide the following guidelines for choosing the ACKNOWLEDGMENTS
error measures.
If forecast performance is evaluated for time series research (Grants #12-07-31017, 12-01-00684).
with the same scale and the data preprocessing
procedures were performed (data cleaning, anomaly REFERENCES
detection), it is reasonable to choose MAE, MdAE,
RMSE. In case of different scales, these error 1. Tyukov, A., A. Brebels, M. Shcherbakov and
measures are not applicable. The following V. Kamaev, 2012. A concept of web-based energy
recommendations are provided for mutli-scales cases. data quality assurance and control system. In the
In spite of the fact that percentage errors are Proceedings of the 14th International Conference on
commonly used in real world forecast tasks, but due Information Integration and Web-based Applications
to the non-symmetry, they are not recommended. & Services, pp: 267-271.
If the range of the values lies in the positive 2. Kamaev, V.A., M.V. Shcherbakov, D.P. Panchenko,
half-plane and there are no outliers in the data, it is N.L. Shcherbakova and A. Brebels, 2012. Using
advisable to use symmetric error measures. Connectionist Systems for Electric Energy
If the data are "dirty", i.e. contain outliers, it is Consumption Forecasting in Shopping Centers.
advisable to apply the scaled measures such as Automation and Remote Control, 73(6): 1075-1084.
MASE, inRSE. In this case (i) the horizon should be 3. Owoeye, D., M. Shcherbakov and V. Kamaev, 2013.
large enough, (ii) no identical values should be, (iii) A photovoltaic output backcast and forecast method
the normalized factor should be not equal to zero. based on cloud cover and historical data. In the
If predicted data have seasonal or cyclical patterns, Proceedings of the The Sixth IASTED Asian
it is advisable to use the normalized error measures, Conference on Power and Energy Systems (AsiaPES
wherein the normalization factors could be calculated 2013), pp: 28-31.
within the interval equal to the cycle or season. 4. Armstrong, J.S. and F. Collopy, 1992. Error measures
If there is no results of prior analysis and a-prior for generalizing about forecasting methods: Empirical
information about the quality of the data, it comparisons. International Journal of Forecasting,
reasonable to use the defined set of error measures. 8(1): 69-80.
After calculating, the results are analyzed with 5. Mahmoud, E., 1984. Accuracy in forecasting: A
respect to division by zero errors and contradiction survey. Journal of Forecasting, 3(2): 139-159.
cases: 6. Yokuma, J.T. and J.S. Armstrong, 1995. Beyond
For the same time series the results for model maccuracy: Comparison of criteria used to select
1
is better than m, based on the one error forecasting methods. International Journal of
2
measure, but opposite for another one; Forecasting, 11(4): 591-597.
For different time series the results for model m7. Armstrong, J.S., 2001. Evaluating forecasting
1
is better in most cases, but worst for a few of methods. In Principles of forecasting: a handbook for
cases. researchers and practitioners. Norwell, MA: Kluwer
CONCLUSION 8. Gooijer, J.G.D. and R.J. Hyndman, 2006. 25 years of
The review contains the error measures for time series Forecasting, 22(3): 443-473.
forecasting models. All these measures are grouped into 9. Hyndman, R.J. and A.B. Koehler, 2006. Another look
seven groups: absolute forecasting error, percentage at measures of forecast accuracy. International
forecasting error, symmetrical forecasting error, measures Journal of Forecasting, 22(4): 679-688.
based on relative errors, scaled errors, relative errors and 10. Theodosiou, M., 2011. Forecasting monthly
other (modified). For each error measure the way of and quarterly time series using STL
calculation is presented. Also shortcomings are defined decomposition. International Journal of Forecasting,
for each of group. 27(4): 1178-1195.
Authors would like to thank RFBR for support of the
Academic Publishers, pp: 443-512.
time series forecasting. International Journal of
World Appl. Sci. J., 24 (Information Technologies in Modern Industry, Education & Society): 171-176, 2013
176
11. Shcherbakov, M.V. and A. Brebels, 2011. Outliers and 18. Tyukov, A., M. Shcherbakov and A. Brebels, 2011.
anomalies detection based on neural networks Automatic two way synchronization between server
forecast procedure. In the Proceedings of the 31 and multiple clients for HVAC system. In the
st
Annual International Symposium on Forecasting, Proceedings of The 13th International Conference on
ISF-2011, pp: 21-22. Information Integration and Web-based Applications
12. Goodwin, P. and R. Lawton, 1999. On the asymmetry & Services, pp: 467-470.
of the symmetric MAPE. International Journal of 19. Brebels, A., M.V. Shcherbakov and V.A. Kamaev,
Forecasting, 15(4): 405-408. 2010. Mathematical and statistical framework for
13. Koehler, A.B., 2001. The asymmetry of the sAPE comparison of neural network models with other
measure and other comments on the M3-competition. algorithms for prediction of Energy consumption in
International Journal of Forecasting, 17: 570-574. shopping centres. In the Proceedings of the 37 Int.
14. Makridakis, S., 1993. Accuracy measures: Theoretical Conf. Information Technology in Science Education
and practical concerns. International Journal of Telecommunication and Business, suppl. to Journal
Forecasting, 9: 527-529. Open Education, pp: 96-97.
15. Kolassa, S. and R. Martin, 2011. Percentage errors 20. Casella, G. and R. Berger, 1990. Statistical inference.
can ruin your day (and rolling the dice shows how). 2nd ed. Duxbury Press, pp: 686.
Foresight, (Fall): 21-27. 21. Kusiak, A., M. Li and Z. Zhang, 2010. A data-driven
16. Assessing Forecast Accuracy Measures. Date View approach for steam load prediction in buildings.
01.08.2013 http:// www.stat.iastate.edu/ preprint/ Applied Energy, 87(3): 925-933.
articles/2004-10.pdf
17. Syntetos, A.A. and J.E. Boylan, 2005. The accuracy
of intermittent demand estimates. International
Journal of Forecasting, 21(2): 303-314.