Applied Energy 236 (2019) 196–210

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

On the impact of outlier filtering on the electricity price forecasting accuracy T

a,⁎ b,c
Dmitriy O. Afanasyev , Elena A. Fedorova
a
JSC Greenatom, 1st Nagatinskiy pas. 10 bld. 1, Moscow, Russia
b
Financial University under the Government of the Russian Federation, 49 Leningradskiy av., Moscow, Russia
c
National Research University Higher School of Economics, 20 Myasnitskaya st., Moscow, Russia

HIGHLIGHTS

• Outlier filtering leads often to the electricity price forecasting accuracy gain.
• The marginal forecasting performance of outlier filters are comparable.
• Proposed combined filter on prices is efficient while does not require a priori setup.

ARTICLE INFO ABSTRACT

Keywords: Increasing the accuracy of short-term electricity price forecasting allows day-ahead power market participants to
Electricity price forecasting obtain a positive economic effect by bidding close to the equilibrium price. However the electricity price time-
Outlier filtering series is generally infested with extreme values due to high price volatility. This paper discusses the impact of
Committee machine outlier filtering on forecasting accuracy based on a recently introduced seasonal component autoregressive
Model confidence set
model. We consider such methods of outlier detection (with a priori defined cut-off parameter) as threshold,
Long-term trend-seasonal component
standard deviation, percentage, recursive, and moving filter on prices. It is shown that such data pre-processing
often leads to the forecasting accuracy gain while the error decrease (relative to the approach without filtering)
in a number of cases may reach 1.8–1.9% of the average weekly price (in absolute values). For an a priori defined
cut-off parameter, the simple threshold and standard deviation filter on prices outperform other considered
methods, and yield to the accuracy gain in 63% and 67% of cases, correspondingly. At the same time, in case of
the out-of-sample filter parameter grid-optimization all of the methods demonstrate comparable prediction
power (equal to the marginal performance). But, practically speaking, such optimization is time-consuming and
cannot be carried out on unavailable future data. As an competitive alternative, we propose a combined filter on
prices based on a committee machine which uses the results of individual non-optimized algorithms and is not
time-consuming, but gives accuracy comparable to the best one obtained for each of the studied electricity
markets and leads to forecast gain in 63% of the considered cases.

1. Introduction mean-reverting of electricity prices; daily, weekly, and yearly season-

In the modern economy, the accuracy of price forecasting in com-
modity markets is often a key determinant in managerial decision
making for company’s operating and strategic units, with the electricity
industry not being an exception here. But, as compared to other com-
modities, electricity is a unique one in its attributes. Specifically, it has
the following features: non-storable nature of electricity (i.e., generally,
the moment of electricity production coincides with the moment of its
consumption); different technological structures of electricity produc-
tion in different markets (which results in different average prices);
ality of electricity price and demand; low price elasticity of demand
(specifically in peak hours); high price volatility, as well as occurrence
of price spikes [1–3].
Such diversity of electricity unique features resulted over the past
15–20 years in development of a whole bunch of different mathematical
tools and models for its price forecasting, with a primary focus on the
short-term price forecasting in a so-called day-ahead market (DAM),
where the price is determined 24 h before the physical delivery based
on balancing of the demand and the supply (see, for example, [4–6],
and etc.). A comprehensive reviews of recent developments in this field

⁎
Corresponding author.
E-mail addresses: dmafanasyev@gmail.com (D.O. Afanasyev), ecolena@mail.ru (E.A. Fedorova).
URL: http://dmafanasyev.ru/en/ (D.O. Afanasyev).

https://doi.org/10.1016/j.apenergy.2018.11.076
Received 26 May 2018; Received in revised form 13 November 2018; Accepted 22 November 2018
Available online 29 November 2018
0306-2619/ © 2018 Elsevier Ltd. All rights reserved.
D.O. Afanasyev, E.A. Fedorova
can be found in Aggarwal et al. [7], Weron [8].
One of the principal research questions which is still not sufficiently
studied in the modern literature [8] is accounting for the long-term
seasonal component (LTSC) in the price of electricity, specifically in the
context of short-term forecasting. The several studies attempt to fill in
this gap by separately modeling and forecasting of the long-term de-
terministic and short-term stochastic price components [9–11] or even
working on the level of more detailed independent time-scales [12,13].
In particular, Nowotarski and Weron [9] introduces a seasonal com-
ponent autoregressive with exogenous factors (SCARX) model which
combines extrapolation of estimated trend-seasonal component based
on a ”persistent” assumption and deviations forecasts obtained from an
autoregressive model (ARX) specified earlier in Misiorek et al. [14],
Weron and Misiorek [15]. The authors conclude that accounting for
LTSC and a proper choice of the method and parameters for the LTSC
estimation yield a significant decrease of the forecast error as compared
both to the traditional ARX model and to the naïve approach based on
”similarities” of prices in same weekdays [16]. Overall the approach
based on the separate LTSC and short-term dynamics modeling has
great potential for electricity price forecasting [11].
But in the mentioned forecasting researches (accounting the LTSC)
the authors pay little attention to such an important issue as sharp
short-term and, generally speaking, poorly predictable extreme values
of the electricity price, i.e. spikes or drops. These outliers usually occur
due to accidents at power plants, congestions of the energy transmission
networks, and climatic anomalies [17,2]. In the number of papers,
concerning the problem of outliers in electricity price, the authors note
that, in the presence of such extreme values, subtracted deterministic
and stochastic components may be substantially distorted [18,19]. As a
result, the adequacy of models for such components (especially, linear
models where SCARX belongs) is very doubtful, with the estimates of
parameters of these models being potentially biased. In the case of in-
sample modeling such biases may cause, for example, under-evaluation
of value-based risk measures, while in the case of forecasting this may
potentially corrupt the price forecast accuracy.
There are two principal approaches to dealing with outliers in the
energy literature. First one is outlier modeling, i.e. incorporating spikes
into the electricity price modeling framework. Early studies use
econometric tools for this purpose. Cuaresma et al. [20], Cartea and
Figueroa [21] consider spikes as a stochastic jump process (Bernoulli
and Poisson). De Jong [22], Janczura and Weron [23] introduce an
independent regime for outliers into a 2-state Markov regime-switching
model (MRS). Janczura et al. [19] pay attention not only to positive
(spikes) but also to negative outliers (drops) for which they use a 3-state
MRS model. As an alternative, data mining (or machine learning)
techniques were introduced for the extreme price forecasting problem
in the last decade. Amjady and Keynia [24] use a probabilistic neural
network for prediction of spike occurrence, while Zhao et al. [25],
Voronin and Partanen [26] utilize the k-nearest neighbors (kNN) ap-
proach to forecast the actual values of spikes.
Second approach is outlier filtering: the identified extreme values
that exceed a selected threshold are replaced by ”normal” price values,
and the obtained outlier-free time-series are modeled then. There is
quite a lot of identification techniques discussed in the literature: the
fixed threshold filter on prices [27,28], the standard deviation filter on
prices [21], the recursive filter on prices [29], the moving filter on
prices or the varying threshold filter [30], the percentage filter on
prices [18], as well as filters for price increments [19].
While the first approach considers outliers in a more ”natural” form,
it requires introduction of non-linear model specifications and ad-
vanced data mining techniques, which increases both estimation com-
plexity and time. For example, Voronin and Partanen [26] use a com-
plex pattern recognition technique for outlier modeling that takes about
40 h for the initial model to setup (for the first day forecast) and about
50 min for each of the next days forecasting as reported in their paper.
Practically speaking, it is much more preferable to get results faster
197
Applied Energy 236 (2019) 196–210
while utilizing less computational resources. The second approach
based on preliminary outlier filtering is quite competitive to the first
one in this sense while also being suitable for mean price forecasting
[16,31]1. At the same time, outlier filtering in the current energy lit-
erature adopted mostly for in-sample modeling frameworks [18,19]
while in the forecasting research it has quite a limited application.
Moreover, as mentioned above, there is a lot of filtering techniques
used in the previous modeling studies, and there is no consensus on
which spike identification methods should be used and which price
values are to be treated as spikes. The application of various methods
may lead to principally different results, with the number of observa-
tions classified as outliers being dramatically different for the same
time-series. From this perspective, it is important to test empirically
which methods from the early used for the in-sample framework
modeling are suitable in the forecasting context as well.
It is also worth noting that previous studies use mostly an a priori
defined setup for filtering methods. At the same time, correct value of
the outlier filter threshold is a key point to gain in forecasting accuracy.
Indeed, on the one hand, it is quite natural to assume that neglecting
pre-filtering of extreme values may have a negative impact on the
forecasting accuracy. But, on the other hand, too aggressive filtering of
spikes may also substantially distort the original time-series and result
in predictions virtually unrelated to the original historical dynamic of
the price. To the best of our knowledge, there are no studies that raise
the question of an optimal cut-off threshold selection in the context of
forecasting accuracy. Besides, the intuition of the committee machine
approach from the data mining area suggests the use of a combination
of a priori defined outlier filters as an additional alternative.
Taking into account the mentioned gaps in the literature, in this
research we pose the question of outlier filtering impact on electricity
price forecasting accuracy. As a basis model for our study we chose the
SCARX model. Firstly, this is due to the fact that this model separately
estimates the LTSC and the short-term dynamics of the electricity price,
with both of the components being prone to substantial influence of
outliers; and, secondly, due to the fact that the SCARX model, by
changing its smoothing parameter, allows to obtain a whole range of
quite independent models, and, thus, to have more objective results of
testing using them. Let us also note that while this research is focused
on the SCARX model, the most important findings can be extended to
other models, because the filtering and modeling steps of the proposed
methodology are conducted independently.
Our study may be considered as a development of the ideas ex-
pressed in Truck et al. [18], Janczura et al. [19] and Nowotarski and
Weron [9], with our own contribution to the existing literature being
the following:
1. In the context of forecasting, we compare the performance of 5 most
commonly used outlier filtering methods (with a priori defined
setup), early applied for the in-sample modeling. Based on the
weekly-weighted mean absolute error and the model confidence set
performance comparison approaches, we show that the threshold
filter on prices (with the cut-off log-price being equal to 0.5) and the
standard deviation filter on prices (with the multiplier being equal
to 3) are efficient and fast tools to achieve accuracy gain and can be
applied on different power markets despite the a priori defined cut-
off parameter.
2. We optimize cut-off parameters of the filters in the sense of the out-
of-sample forecasting accuracy and show that each of the 5 con-
sidered methods has its inherent saddle point where the forecasting
error is minimized while the discrepancies in prediction
1
For our research, the total computational time of outlier filtering, model
calibration, and forecasting on the test period of 728 days was about 40 min for
4 markets on a hardware configuration almost similar to the one used in
Voronin and Partanen [26].
D.O. Afanasyev, E.A. Fedorova
performance are insignificant for different filters. In other words,
with a proper setup all the considered filters have nearly equivalent
marginal forecasting performance. But it should also be kept in mind
that choosing an optimal values for the filter threshold is itself a
time-consuming task which cannot be practically solved for un-
available future data.
3. As a competitive alternative to the unrealistic out-of-sample opti-
mization of filters’ parameters and as a way to mitigate the risk of
incorrect a priori defined cut-off parameter, we propose a combined,
committee machine based filter which implies independent appli-
cation of basic algorithms (outlier filters) with subsequent result
combination via the majority voting rule, and is an efficient tool in
the sense of forecasting accuracy and computational time.
The rest of the paper is organized as follows. In Section 2 we provide
our research methodology: the SCARX model, methods for outlier fil-
tering, the numerical experiment design, and approaches to perfor-
mance evaluation. In Section 3 we discuss the data used in the empirical
validation of the filters. Section 4 considers the obtained results, as well
as proposes some practical recommendations for outlier filtering ap-
plication. Section 5 concludes.
2. Research methodology
2.1. Seasonal component AutoRegressive with exogenous factors model
The procedure of short-term forecasting of electricity price with use
of the seasonal component autoregressive with exogenous factors
(SCARX) model proposed by Nowotarski and Weron [9] consists of the
following steps.
At the first step, the time-series of hourly electricity prices
pt , t = 1, …, T is taken the logarithm and is additively decomposed into
a long-term seasonal (trend-cyclical) component Lt and a short-term
seasonally stochastic component St :
logpt = Pt = Lt + St (1)
At this step, different methods for Lt estimation can be used. But,
generally speaking, selection of a relevant filter is a non-trivial problem.
There is no consensus in the energy economics literature on the best of
such filters, while the question itself remains disputable [32,33]. But,
taking into account that in the current research we are primarily in-
terested in the influence of outlier filtering on the accuracy of fore-
casting (but not in the effect of smoothing method selection), as well as
considering the recent results of Nowotarski and Weron [9], Xiao et al.
[34], Yang et al. [35], hereafter we use the wavelet-decomposition for
LTSC estimation.
The wavelet-decomposition is a decomposition of a time-series over
an orthogonal basis – a shifted and scaled version of a wavelet (t ) . The
original time-series is presented as a combination of successive
approximations of the coarse (approximating or smoothing) Am
component and the refined (detailing) Dm component with their
following adjustment by an iterative procedure:
Pt = A1 + D1 = A2 + D2 + D1 = …=Am + Dm + Dm 1 + Dm 2 + …+ D1.
The basis wavelet for the approximating Am component is called the
”father” wavelet, and the one for the detailing Dm component – the
”mother” wavelet. By changing the scale parameter m we get different
degrees of smoothness of the original time-series.
Following Nowotarski et al. [31], we use the Daubechies wavelet
family of order 24. To study the influence of outlier pre-filtering on the
accuracy of forecasting on a ”sufficient” number of different models, we
consider a wide range of values of the scale parameter m = 8, …, 13. In
addition, such approach is caused by the fact that it is a priori unknown
which of the values will yield the estimate of LTSC as close as possible
to the ”true” LTSC. The choice of this specific range of values is dictated
by the following considerations. The value of corresponds to smoothing
at a 10-day scale (28 = 256 h 10 days) and ensures exclusion of
198
Applied Energy 236 (2019) 196–210
weekly (and certainly, daily) seasonality from the estimate of LTSC. The
value of m = 10 corresponds to smoothing at a 1 year scale
(213 = 8192 h 341 days 1 year), which coincides with the width of
the calibrating window set by us to estimate the model parameters (see
the numeric experiment design in Section 2.4).
At the second step, for the short-term seasonally stochastic compo-
nent St parameters of the autoregressive specification proposed by
Misiorek et al. [14] are estimated:
St = 1 St 24 + 2 St 48 + 7 St 168 + 8 mSt + 1 Zt + di Dti +
t
i = 1,6,7
(2)
where St 24/48/168 are autoregressive components reflecting the influ-
ence of the prices in previous days; mSt is the price signal for the market
participants, which is equal to the minimum price in the previous day’s
trading session; Zt is an exogenous factor which in our case is equal to a
day-ahead forecast of electricity consumption (load) before time mo-
ment t ;Dti are dummy variables accounting for weekly seasonality
(i = 1, 6, 7 for Monday, Saturday, and Sunday correspondingly); t is a
normally, independently, and identically distributed error term.
The modeling is run for each of the day hour separately, i.e. there
are 24 models in total. Such approach is widely spread in the literature
on electricity price forecasting [20,8]. On the one hand, it allows not to
take explicit account of the intra-day patterns, which significantly
simplifies calibration of the model and reduces the number of coeffi-
cients to be estimated; and on the other hand, the approach is dictated
by the fact that different day hours exhibit principally different price
dynamic [8]. Using each of the estimated models, one-step-ahead
forecasts ST + 1 are calculated (where T is the time moment of the last
available observation), which ultimately results in a forecast for the full
day ST + h, h = 1, …, 24 . It should be noted that the parameters estima-
tion of Eq. (2) was done by the ordinary least squares.
Eq. (2) can also be applied directly to the logarithmic prices Pt (after
they have been centered) as it originally proposed by Misiorek et al.
[14], i.e. without special accounting for their long-term dynamic. For
the purposes of comparison, we will also consider this approach, with
the corresponding model being denoted as ARX.
At the third step, the long-term seasonal component Lt is assumed to
be persistent in the short-term scales, i.e. it remains almost unchanged
within each specific hour. As a result, its forecast value is equal to the
value in the corresponding hour of the previous day, i.e.
LT + h = LT + h 24 . Based on this assumption, day-ahead forecasts
LT + h , h = 1, …, 24 are calculated.
Finally, at the fourth step, the early obtained forecasts ST + h and LT + h
are summed up for the corresponding hours h = 1, …, 24 (see Eq. (1)),
and the inverse logarithmic transformation is calculated which gives
the final forecast pT + h . From now on, we will denote the considered
model based on the wavelet-decomposition with scale parameter m as
SCARXm.
2.2. Price outlier filtering
The above-described SCARX model can be applied both directly to
the original time-series (source, SRC), and to its outlier-free version
(where outliers are filtered from short-term component St ). Potentially,
outlier filtering may result in accuracy gain. At the same time, fanatic
substitution of price outliers for ”normal” prices will quite substantially
distort the original time-series, which may have an adverse effect on the
forecast error. Empirical investigation of this effect is an interesting
problem, which may allow to obtain some new findings in short-term
forecasting of electricity price.
2.2.1. Short-term weekly and daily seasonalities
But before switching to discussion of time-series St outlier filtering
methods, let us consider the following important fact. As to the elec-
tricity price, its short-term seasonality and its spikes are of principally
D.O. Afanasyev, E.A. Fedorova
different nature. While the former occurs due to business and climatic
cycles, the latter are typically caused by accidents at power plants,
congestions of the energy transmission networks, and climatic anoma-
lies, i.e. have a stochastic nature. Besides, technically speaking, the
presence of the short-term seasonal component in time-series St may
result in inadequate performance of outlier filtering methods (for ex-
ample, sample standard deviation is incorrectly calculated; see the
description of the SFP filter below). Taking this into account, before
running an outlier filtering procedure, the short-term component St
should be pre-filtered for weekly and daily seasonality. To this end,
different methods can be applied (see, for example, [32]). But, since in
this study we primarily focus on the influence of outlier filtering on
forecast accuracy, then, following Trück et al. [18], Janczura et al.
[19], we use quite a simple approach: we calculate median prices for
each separate hour of the week (overall, 168 median price values),
demean these median prices (to have their sum equal to 0), and after
that subtract them from St . But doing this, we eliminate weekly sea-
sonality. For the obtained residual time-series, we repeat the same
procedure, but for each hour of the day (overall, 24 median price va-
lues), to eliminate daily seasonality. Thus obtained stochastic compo-
nent Xt of the electricity price is then subject to an outlier filtering
procedure.
2.2.2. Methods of price outlier filtering
As of nowadays, to solve the problem of filtering out atypical values
of stochastic component Xt in the price of electricity, there are quite a
lot of different approaches in the literature. In this research we consider
the following ones.
Filtering based on threshold filter on prices (TFP) was earlier used in
Lapuerta and Moselle [27], Fanone et al. [28] considers as outliers the
price values that are greater than some arbitrarily set threshold. In
these studies, the authors pay primary attention to positive spikes, al-
though, generally speaking, negative spikes are also present in the
electricity prices, while the upper and lower thresholds can be assigned
different values. In our study, based on simple visual analysis, we use
symmetrical thresholds for the logarithmic prices (−0.5; 0.5), and the
subset of price spikes Xto is detected by the TFP filter as follows:
Xto = {Xt : Xt 0.5} (3)
Filtering based on standard deviation filter on prices (SFP) implements
price substitution for those prices whose absolute deviation from the
sample mean X is greater than a multiple of sample standard deviation
. Generally speaking, the multiplier can be chosen arbitrarily, but in
the literature its most popular value is 3 [29,21,18]. Thus, the subset of
spikes Xto detected with the use of the SFP filter is:
Xto = {Xt : Xt X 3· } (4)
Recursive filter on prices (RFP) works similarly to SFP but its proce-
dure is repeated iteratively for the whole time-series until there is no
spike left. At the first step, sample mean X1 and sample standard de-
viation 1 are calculated. After that, using Eq. (4), the set Xto,1 is iden-
tified, and the values of this set are substituted for ”normal” prices. If set
Xto,1 is not empty (contains at least one spike), then the obtained time-
series (with partially removed spikes) is given as input to the next
iteration. Otherwise, the algorithm is stopped. Such filtering was earlier
used in Clewlow and Strickland [29], Cartea and Figueroa [21] with the
only difference that the authors worked with price increments instead
of price levels.
Moving filter on prices (MFP) was proposed in Borovkova and
Permana [30] and is also similar to SFP, with its procedure being run
not for the whole time-series under study but in a rolling window of
fixed width. At the first step, the original time-series is split into
N = T / M parts, where M is the width of the window2. Then, to the first
2
It is worth noting that in the case when the length of the final part of the
199
Applied Energy 236 (2019) 196–210
piece of the original time-series the SFP filter is applied (see Eq. (4)),
after which the window is shifted M observations forward, and the
filtering procedure is repeated in the next window. Taking into account
that the MFP filter is of more local kind that the SFP filter, it is sensible
for MFP to choose fewer number of standard deviations as the threshold
value. Borovkova and Permana [30] propose to use to this end the value
of the 90% predictive interval as implied by the normal distribution
(i.e., 1.64· ). In this study we do similarly, but take the 95% predictive
interval (i.e., 1.96· ), with the width of the window being equal to four
weeks (or 672 h). Thus, the subset of outliers Xto , as obtained by MFP
with a rolling window of width M, is found as follows:
Xto = Xn: Xn Xn 1.96· n,
n = 1, … , N
n n 1 · M + 1, n· M
(5)
Filtering based on percentage filter on prices (PFP) considers as out-
liers a specific fraction of minimum and maximum electricity prices.
The filter was earlier used in Trück et al. [18], Janczura et al. [19]. As it
was the case for TFP, the above-mentioned fraction can be set different
for positive and negative spikes. But, following Janczura et al. [19], in
this research we apply a symmetrical threshold equal to 2.5%. Thus, the
subset of outliers Xto obtained by PFP is found as follows:
Xto = {Xt : Xt Xt2.5/100 } {Xt : Xt Xt97.5/100} (6)
where and
Xt2.5/100 are the percentiles of time-series Xt of orders
Xt97.5/100
2.5 and 97.5, correspondingly.
Besides the above-mentioned, in this research we also propose a
combined filter on prices (CFP) which, to the best of our knowledge, was
not earlier applied in energy studies on outlier filtering. Here we ex-
plain the idea of this filter.
The problem of outlier identification can formally be considered as a
binary classification problem, specifically, as a problem of attribution of
each observation to one of two classes – outlier or regular price value.
Nowadays, there is quite a lot of classifiers for solving this problem
[36,37]. But each of them may lead to very different results [38].
Moreover, as an extreme case, identically classified sets of observations
may not intersect at all for different classifiers. Similarly, in the context
of our study, each specific outlier filter k may produce a subset of price
spikes Xto, k that is dramatically different from the results of other
methods. First of all, this happens because of the presence of a priori
defined threshold parameter for each specific filter.
The committee machine approach potentially allows to overcome
this issue. The main idea of this approach is to combine independent
results of basic algorithms (filters, as in our case), which often exhibits
much higher overall classification accuracy as compared to individual
methods [39–41]3. For the outlier filtering problem, the main ad-
vantage of this method is mitigation of risk of incorrectly a priori de-
fined threshold parameter. While there are many different combination
schemes [46], in this study we use quite an intuitive scheme of voting in
order to identify the final class of the observation.
Thus, the application of the proposed combined filter on prices
consists of the following steps:
1. Each time-series observation is classified using all the K = 5 above-
(footnote continued)
time-series is shorter than M, this part can either be merged with the previous
one, or considered separately. But in our study, for the purpose of convenience,
M was chosen in such a way that the calibrating window splits into an integer
number of filtering windows.
3
In the context of time-series forecasting, an analogue (in some sense) to the
committee machine approach is the forecast combinations [42,43] which often
allows to achieve higher accuracy as compared to individual models and, as of
recently, is actively used in the energy literature (see, for example, [44,45]).
D.O. Afanasyev, E.A. Fedorova
mentioned individual filtering methods (TFP, SFP, RFP, MFP, PFP),
and subsets of outliers Xto, k , k = 1, …, K are identified.
2. The individual results are combined using a voting scheme in order
to get the final subset of price spikes Xto :
K tion. We split the original time-series of the electricity price (and,
1 correspondingly, energy consumption prediction) into two sub-periods.
Xto = Xt : Xt Xto, k Q
K k=1
where (Xt Xto, k ) is the indicator function reflecting attribution of
observation Xt to subset of outliers Xto, k ;Q is a fraction of ”votes”
(filters) which must be exceeded to finally classify observation Xt as
an outlier.
When choosing a specific value of Q, there are two mutually ex-
cluding options: either to recognize as outlier only those observations
that all the filters recognize as such (i.e., Q = 1); or those observations
which are said to be outliers by at least one filter (i.e., 0 < Q < 1). In
this paper we adopt a more balanced strategy, specifically, the majority
vote rule, i.e. set Q = 0.5 (in our case, three out of five filters).
2.2.3. Outlier replacement
At the next step, outliers Xto identified in stochastic component Xt
by the above-considered methods are to be replaced by ”normal” values
of the price. To this end, Borovkova and Permana [30] propose to use
the upper and lower price thresholds as set by the filtering procedure.
Weron [47] calculate the average of the previous and the following (as
related to the outlier) prices. Bierbrauer et al. [48] substitute spikes for
the median of the prices in the same day of the week and the same
month of the year. Nowotarski et al. [45] apply a damping scheme: all
the prices above threshold X are substituted for
Xt = X + X log10 (Xt /X ) . Finally, Janczura et al. [19] use sample
average of seasonally adjusted price. We adopt the same approach here.
Since the mean of stochastic component Xt is zero, such scheme vir-
tually implies that in the original price Pt the spike is replaced with a
sum of long-term trend-seasonal and short-term seasonal components,
i.e. the replacement is done for a typical (for the current time of the
year, weekday, and day hour) value of the price.
2.3. Forecasting with outlier pre-filtering
Summarizing the ideas on the above-proposed forecasting metho-
dology for SCARXm-XXX based on the SCARXm model and outlier fil-
tering routine XXX, below we give the principal steps of application of
this model (see also Fig. 1 for a flowchart):
1. Decompose the original electricity log-price time-series Pt into long-
term Lt and short-term St components using a wavelet smoothing
with scale parameter m (see Section 2.1).
2. Eliminate short-term weekly and daily seasonality Dt from St using
the median-based approach (see Section 2.2.1) and obtain stochastic
component Xt .
3. Apply one of the outlier filtering routines XXX to Xt as described in
Section 2.2.2 (where XXX is one of the filters – TFP, SFP, RFP, MFP,
PFP, CFP) and replace the detected extreme values with a sample
average of seasonally adjusted prices (see Section 2.2.3) to obtain
filtered stochastic component Xt .
4. Restore outlier electricity price time-series Pt as a sum of Lt , Dt and
Xt .
5. Apply the SCARXm model to the pre-processed time-series Pt and
calculate a day-ahead forecast of electricity price (see Section 2.1)4.
4
Let us note that in general it is possible to use very wide class of models at
this step, not only the SCARX model.
Applied Energy 236 (2019) 196–210
2.4. Experiment design and models’ performance evaluation
To compare performance of different outlier filters we conduct the
numeric forecasting experiment based on out-of-sample cross-valida-
(7) The first one is calibrating, 52 weeks (or 8736 h) long, where we esti-
mate the models. The second is test sub-period, 104 weeks (or 17472 h)
long. The parameters of the model were estimated on the initial cali-
brating sub-period, and after that the forecast for the first 24 h of the
test sub-period was calculated using the estimated model. Then, the
initial calibrating window was moved forward by 24 h, the model was
re-estimated, and the forecast for the next 24 h was calculated. The
procedure was repeated until the end of the test sub-period. The ex-
periment was conducted both for the original time-series, and for the
time-series pre-filtered for outliers using the 6 methods described in
Section 2.2.2.
For the forecasting performance evaluation of the models, we use
two independent approaches, which, in our opinion, yields to more
objective conclusions. Our first approach is based on a weekly-weighted
mean absolute error WMAE which, as compared to the traditional
MAE , is more robust to close-to-zero prices that occur in the market
from time to time [16,15]:
1
168
Ph Ph
WMAE =
168 h=1
P168 (8)
168
where P168 = Ph is the weekly average electricity price. Note that
h=1
though WMAE is called absolute error, actually it is a ratio of mean
absolute error MAE to the weekly average price. This is why hereinafter
it is expressed in percentage points.
Our second approach is a formal statistical methodology of models’
predictive ability comparison based on the model confidence set (MCS)
proposed in Hansen et al. [49,50] and used in Nan [51], Samuels and
Sekkel [52], Garcia et al. [53]. MCS is a models subset 1 of a set
0 for which the null hypothesis of equal predictive ability (EPA) is
not rejected for a chosen significance level . One of the advantages of
the MCS over other statistical approaches (see, for example [54,55]) is
that its doesn’t require to select one reference model to compare other
models with, but implies pair-wise comparison of all available models.
At the first step, setting = 0
, the procedure of 1 identification
can be specified as the following algorithm:
1. For set an EPA-hypothesis is formulated:
H0 : (dij, t ) = 0, i, j (9)
where i , j are indices of models from ;t = 1, …, is the test sub-
period; dij, t is the loss differential which is equal to the difference of
loss functions dij, t = Li, t Lj, t and is subject to the common condi-
tions of weak stationarity; (dij, t ) is finite and time-independent
expected value of dij, t .
2. For all pairs of models ij a number of intermediate t -statistics is
calculated:
dij
t ij =
var (dij, t ) (10)
where dij = d is the sample average loss of model i re-
1
t = 1 ij, t
lative to model j;var (dij, t ) is the estimate of variance dij, t .
3. Test statistic TR = max i, j t ij is calculated.
4. For the obtained TR the corresponding p-value of the MCS proce-
dure is found. Since the asymptotic distribution of TR is non-tra-
ditional [50], the stationary bootstrap [56] is used to find the p-
value. When forming bootstrapped samples, this ensures both the
identity of their structures to the structure of the original time-series
200
D.O. Afanasyev, E.A. Fedorova Applied Energy 236 (2019) 196–210

Fig. 1. Flowchart of the day-ahead (h = 1, …, 24 ) electricity price forecasting methodology for SCARXm-XXX based on the SCARXm model and outlier filtering routine
XXX (where m is the scale parameter of wavelet smoothing and XXX is one of the methods – TFP, SFP, RFP, MFP, PFP and CFP).

dij, t , and preserving their stationarity5. will naturally allow to rank them according to their MCS p-values. This
5. If the p-value is smaller than the chosen level of significance , then peculiarity of the MCS will later be used by us when comparing the
the null hypothesis H0 is rejected, with model indexed by models with outlier filters and without them. Notice that in this study
eR = argmax i supj t ij being excluded from , after which the we use the absolute loss function Lt = pt pt due for the first models
procedure is repeated again for the obtained smaller model subset. comparison approach we also use absolute error (weighted by the
Otherwise, the current subset is recognized the wanted 1 . weekly average price) as an error measure.

The above-described steps can be repeated until in subset there is

only one model left6. Then the obtained sequence of models exclusion
5 Section 2.4 was run on four historical datasets of day-ahead market
We used 2000 generations of bootstrapped sub-samples with the average
width of the block of 24 h. The latter value was chosen in light of the fact that it
corresponds directly to the seasonal component with the smallest period in the
structure of the price time-series.
6
Technically speaking, it can be achieved by setting = 1 , where is an
arbitrarily small value.
3. Data
The empirical forecasting experiment which design was described in
prices and consumption forecasts: the Europe-Ural price area (EU) and
the Siberia price area (SI) of the Russian ATS power market; Nord Pool
(NP), the leading European power market; Pennsylvania-New Jersey-
Maryland (PJM), the US power market. In Fig. 2 the time-series used in
this research are depicted.

201
D.O. Afanasyev, E.A. Fedorova Applied Energy 236 (2019) 196–210

Fig. 2. Time-series of hourly electricity prices and consumption forecasts for the Europe-Ural price area (EU) and the Siberia price area (SI) of the Russian ATS power
market, the Nord Pool (NP) power market in Europe, and the Pennsylvania-New Jersey-Maryland (PJM) power market in the USA. Vertical dashed line in the graphs
separates two principal areas: (1) area to the left is the initial calibrating period of 52 weeks; (2) area to the right is the test period of 104 weeks.

For the Russian price areas we used data for the period from the 1st of
January, 2015, to the 27th of December, 2017. Historical electricity prices
were freely obtained from the official web-site of joint-stock company
”System Operator of the Unified Energy System” (JSC ”SO UES”)7. In the
data for the EU price area there were 94 missing values that we substituted
for the prices of the corresponding hour of the previous day. In the same
manner, we dealt with the 53 missing values for the SI price area. Taking
into account that the fractions of these values are 0.36% and 0.2%, cor-
respondingly, in the total number of observations, such substitutions, in our
opinion, couldn’t have substantially distorted the results of the experiment.
For the historical values of electricity consumption we used predicted
hourly load values, also freely available from the SO UES web-site8.
For the European Nord Pool power market we used data from the
1st of January, 2013, to the 28th of December, 2015. The historical
values of hourly system prices and consumption predictions were ob-
tained from the official web-site of the power market9. Finally, for the
Pennsylvania-New Jersey-Maryland power market we used data from
the 1st of January, 2015, to the 27th of December, 2017. As electricity
prices we took day-ahead locational marginal prices (LMPs) for the
aggregated price area PJM-RTO10. As for electricity consumption, we
7
Indicator ”Day-ahead price”, the ”Price Indicators” report, http://br.so-ups.
ru/Public/MainPageData/BR/IndicatorBR.aspx.
8
Indicator ”Planned consumption”, the ”Production and consumption” re-
port, http://br.so-ups.ru/Public/MainPageData/BR/GenConsum.aspx.
9
Indicators ”SYS” and ”Consumption prognosis” (sum for Norway, Sweden,
Finland, and Denmark), the ”Historical Market Data” report, http://
nordpoolspot.com/historical-market-data/.
10
Indicator ”System Energy Prices Day Ahead”, the ”Day-Ahead Hourly
LMPs” report, https://dataminer2.pjm.com/feed/da_hrl_lmps
used predicted hourly values of the aggregated load as calculated by the
trade operator as of 11:45AM the day before11. In these data on the
predicted load there were 74 missing values (or 0.28% of the total
number of observations) that, similarly to the price missing values for
the EU and SI price areas, were substituted for the values of corre-
sponding hours of the previous days. Again, the small number of such
substitutions, in our opinion, couldn’t have substantially spoilt the ex-
periment results.
4. Results
4.1. The mechanics of outliers’ impact on the accuracy of forecasting
Before we switch to the analysis of results of the empirical experi-
ment described in Section 2.4, let’s consider two examples of the me-
chanics of outliers’ influence on the deterministic and stochastic com-
ponents in the price of electricity, as well as on the accuracy of the
obtained forecasts. Fig. 3 shows a little bit different situations in two
markets (the Siberia price area of the Russian ATS power exchange and
the European Nord Pool power exchange), which allow to explicitly
demonstrate such influence (see the description of the graphs in the
footnote).
In the first case (left-hand side of Fig. 3) the major part of outliers in
the stochastic component Xt is concentrated in the middle section of the
calibrating period (see the top-left graph). As a consequence, the long-
term trend-seasonal component Lt estimated for the original time-series
11
The ”Historical Load Forecasts” report, https://dataminer2.pjm.com/feed/
load_frcstd_hist.

202
D.O. Afanasyev, E.A. Fedorova Applied Energy 236 (2019) 196–210

Fig. 3. Mechanics of outliers’ influence on electricity price forecasting - the cases of the Siberia price area (SI, left-hand side) and the Nord Pool power exchange (NP,
right-hand side). Top panels: stochastic component Xt in the calibrating period and outliers Xto as identified by the CFP filter (the ”cross” markers). Middle panels:
logarithmic prices Pt , estimates of LTSC for the original (Lt , dashed line) and filtered (Lt , solid line) time-series. Bottom panels: day-ahead price forecasts (the
SCARX10-model) for the original (dashed line) and filtered (solid line) time-series, and actual prices (dot-dashed line).

(the middle-left graph, dashed line), is shifted in this section down-
wards relative to the deterministic component Lt estimated for the CFP-
filtered time-series Pt (the middle-left graph, solid line). While there is
no such difference between Lt and Lt in the concluding part of the time-
series. In this case, the parameters of Eq. (2) for the corresponding
short-term seasonally stochastic components St and St appear to be
different, which leads to different forecasts calculated on their basis. As
it can be seen in the bottom-left graph, price forecasts for the filtered
time-series (solid line) appears (on the average) to be closer to the
actual values of the price (dot-dashed line) as compared to the price
forecasts for the original time-series (dashed line). This is also con-
firmed by the daily mean absolute error DMAE which is calculated
similarly to WMAE (see Eq. (8)), except the averaging is run over 24 h.
For the original time-series this error is 6.40%, while for the filtered one
it is 2.34% (approximately 2.7 times smaller)12.
The second case (right-hand side of Fig. 3) is somewhat different
from the first one and is characterized by the fact that a certain portion
of most significant outliers of the stochastic component Xt is con-
centrated in the concluding part of the calibrating period, while the rest
of the outliers are not that prominent. As a consequence, the estimates
of the long-term trend-seasonal components of the original and the
filtered time-series are different basically in the concluding section of
the calibrating window. In this case, the parameters of the model for the
corresponding seasonally stochastic components St and St are almost
12
Note that the provided values of DMAE are calculated for the prices in
levels, although in Fig. 3 the price forecasts are given in the logarithmic scale
for compatibility with other graphs.
equal (as well as the produced corresponding predictions), while the
persistent forecasts of Lt and Lt are different. As it can be seen in the
bottom-right graph, these differences are very prominent in the final
electricity price forecasts for the first four hours of the day, while for
the rest of the trade periods the forecasts almost coincide. As a result,
daily error in case of the original time-series is 3.95%, while in the case
of the filtered time-series it is equal to 2.37%. In this second case, the
relative difference in the errors is smaller (as compared to the first
case), but is still quite considerable and equal to 1.7 times.
Thus, based on the above-considered two examples, it can be seen
that, generally speaking, outlier filters do impact the forecasts both by
influencing the forecast values of the short-term seasonally stochastic
component St (through bias of Eq. (2) parameters estimation), and by
influencing the forecasts of the long-term trend-seasonal component Lt
(through changing its values in the concluding part of the calibrating
window). It goes without saying that these two examples are just
standalone cases, while a more rigorous and comprehensive analysis
requires accuracy comparisons on a long test period, including con-
sideration of different filters for price spikes.
4.2. Comparison of outlier filters performance
Let us now consider the analysis of influence of outlier filtering on
forecasting accuracy, as well as comparing the performance of different
outlier identification procedures. Table 1 gives the values of WMAE
averaged over 104 weeks of the test period, and its absolute deviation
from the model for the original (unfiltered) data.
For the ATS EU power market, the best results were produced by

203
D.O. Afanasyev, E.A. Fedorova Applied Energy 236 (2019) 196–210

Table 1
Comparison of forecasting performance (in the WMAE sense) of SCARX models (with smoothing parameters m = 8, …, 13) with and without outlier filtering for the
four power markets: ATS Europe-Ural (EU), ATS Siberia (SI), Nord Pool, Pennsylvania-New Jersey-Maryland.
Models Outlier filters

SRC TFP SFP RFP MFP PFP CFP

ATS Europe-Ural (EU)

ARX 4.72 – – – – – –
SCARX8 4.82 4.53 (−0.29) 4.57 (−0.25) 4.95 (0.13) 4.66 (−0.16) 5.13 (0.31) 4.58 (−0.24)
SCARX9 4.73 4.46 (−0.27) 4.48 (−0.25) 4.76 (0.03) 4.60 (−0.13) 5.00 (0.27) 4.49 (−0.24)
SCARX10 4.67 4.44 (−0.23) 4.46 (−0.21) 4.64 (−0.03) 4.56 (−0.11) 4.90 (0.23) 4.47 (−0.20)
SCARX11 4.66 4.43 (−0.23) 4.44 (−0.22) 4.58 (−0.08) 4.60 (−0.06) 4.93 (0.27) 4.48 (−0.18)
SCARX12 4.65 4.44 (−0.21) 4.45 (−0.20) 4.55 (−0.10) 4.63 (−0.02) 4.90 (0.25) 4.48 (−0.17)
SCARX13 4.62 4.46 (−0.16) 4.50 (−0.12) 4.76 (0.14) 4.68 (0.06) 4.89 (0.27) 4.56 (−0.06)
Summary
– −0.23 −0.21 0.01 −0.07 0.27 −0.18
# < SRC – 6 (100%) 6 (100%) 3 (50%) 5 (83%) 0 (0%) 6 (100%)

ATS Siberia (SI)

ARX 8.84 – – – – – –
SCARX8 8.93 7.75 (−1.18) 7.72 (−1.21) 8.02 (−0.91) 8.09 (−0.84) 7.82 (−1.11) 7.75 (−1.18)
SCARX9 9.07 7.58 (−1.49) 7.64 (−1.43) 7.74 (−1.33) 7.94 (−1.13) 7.86 (−1.21) 7.58 (−1.49)
SCARX10 8.77 7.47 (−1.30) 7.53 (−1.24) 7.41 (−1.36) 7.73 (−1.04) 7.51 (−1.26) 7.46 (−1.31)
SCARX11 8.83 7.42 (−1.41) 7.52 (−1.31) 7.34 (−1.49) 7.74 (−1.09) 7.52 (−1.31) 7.41 (−1.42)
SCARX12 9.35 7.53 (−1.82) 7.69 (−1.66) 7.45 (−1.90) 8.03 (−1.32) 7.89 (−1.46) 7.50 (−1.85)
SCARX13 8.72 7.96 (−0.76) 7.45 (−1.27) 7.36 (−1.36) 7.71 (−1.01) 7.63 (−1.09) 7.35 (−1.37)
Summary
– −1.33 −1.35 −1.39 −1.07 −1.24 −1.44
# < SRC – 6 (100%) 6 (100%) 6 (100%) 6 (100%) 6 (100%) 6 (100%)

Nord Pool (NP)

ARX 8.58 – – – – – –
SCARX8 8.58 8.41 (−0.17) 8.50 (−0.08) 9.17 (0.59) 8.53 (−0.05) 8.98 (0.40) 8.53 (−0.05)
SCARX9 8.43 8.31 (−0.12) 8.35 (−0.08) 8.88 (0.45) 8.49 (0.06) 8.75 (0.32) 8.39 (−0.04)
SCARX10 8.38 8.29 (−0.09) 8.30 (−0.08) 8.83 (0.45) 8.46 (0.08) 8.63 (0.25) 8.31 (−0.07)
SCARX11 8.45 8.68 (0.23) 8.42 (−0.03) 9.10 (0.65) 8.66 (0.21) 9.15 (0.70) 8.69 (0.24)
SCARX12 8.53 9.05 (0.52) 8.56 (0.03) 9.07 (0.54) 8.84 (0.31) 9.59 (1.06) 8.80 (0.27)
SCARX13 8.55 9.95 (1.40) 9.30 (0.75) 12.81 (4.26) 8.93 (0.38) 10.31 (1.76) 9.64 (1.09)
Summary
– 0.29 0.09 1.16 0.17 0.75 0.24
# < SRC – 3 (50%) 4 (67%) 0 (0%) 1 (17%) 0 (0%) 3 (50%)

Pennsylvania-New Jersey-Maryland (PJM)

ARX 10.59 – – – – – –
SCARX8 11.00 11.46 (0.46) 11.19 (0.19) 11.50 (0.50) 11.93 (0.93) 11.84 (0.84) 11.46 (0.46)
SCARX9 10.79 11.35 (0.56) 10.99 (0.20) 11.33 (0.54) 11.84 (1.05) 11.53 (0.74) 11.30 (0.51)
SCARX10 10.67 11.22 (0.55) 10.82 (0.15) 11.16 (0.49) 11.73 (1.06) 11.25 (0.58) 11.12 (0.45)
SCARX11 10.61 11.30 (0.69) 10.82 (0.21) 11.10 (0.49) 11.98 (1.37) 11.19 (0.58) 11.15 (0.54)
SCARX12 10.60 11.43 (0.83) 10.87 (0.27) 11.15 (0.55) 12.15 (1.55) 11.27 (0.67) 11.21 (0.61)
SCARX13 10.72 11.37 (0.65) 10.88 (0.16) 11.09 (0.37) 12.32 (1.60) 11.75 (1.03) 11.21 (0.49)
Summary
– 0.62 0.20 0.49 1.26 0.74 0.51
# < SRC – 0 (0%) 0 (0%) 0 (0%) 0 (0%) 0 (0%) 0 (0%)

Summary for all markets

# < SRC – 15 (63%) 16 (67%) 9 (38%) 12 (50%) 6 (25%) 15 (63%)

Comments: WMAE values (in percents) averaged over the test period are given in rows with models names. Absolute deviations of errors of the model with outlier
filtering from errors of the model without outlier filtering are presented in parentheses. Minimum values of WMAE for specific filter and power market (which also
are smaller than the error of ARX-model) are given in bold. Minimum values of WMAE for all filters within a specific power market are underlined. is the average
change of error within one specific filter. ”# < SRC” is the number of SCARX models with a filter, which produced errors smaller than the errors of the same models
on the original data.

filters TFP, SFP, and CFP. The application of these filters with the
SCARX models with all the smoothing parameters (i.e., in 6 out of the 6
considered cases) results in accuracy gain. Average decrease in WMAE
over all the models is 0.23% for TFP, 0.21% for SFP, and 0.18% for CFP.
Comparatively less inspiring results are demonstrated by the MFP filter
for which in 5 out of the 6 cases (i.e., in 83% of cases) the error de-
crease, with the corresponding average accuracy gain being only
0.07%. Filters RFP and PFP (as compared to other filters) show quite a
poor performance: the former demonstrates a decrease in the forecast
error only in 3 out of the 6 cases, while the latter gives no advantages of
its usage in either case.
The most substantial decrease in forecast error for the
corresponding models (i.e., for the same value of smoothing para-
meter), as compared to the forecast on the original data, is 0.29% and is
found for the SCARX8-TFP model. The best accuracy gain among all the
scale parameters and filters is obtained by the SCARX11-TFP model,
while for the original data the best performance was demonstrated by
the SCARX13-model. While accuracy gain for the latter model over the
ARX-model (this model does not take account of the long-term dy-
namic) is 0.1%, accuracy gain of the SCARX11-TFP model over SCARX13
is another 0.19%. In total, this allows for additional accuracy gain of
0.29% over the reference ARX-model both due to separate forecasting
of the long-term and short-term components of the electricity price, and
due to the application of the filtering procedures.

204
D.O. Afanasyev, E.A. Fedorova Applied Energy 236 (2019) 196–210

Fig. 4. Comparison of forecasting performance (in the sense of MCS) of the SCARX models (for the smoothing parameters m = 8, …, 13) with and without outlier pre-
filtering for the four power markets: ATS Europe-Ural (EU), ATS Siberia (SI), Nord Pool (NP), and Pennsylvania-New Jersey-Maryland (PJM). The figure shows the
MCS p-value differences p of SCARX models calibrated on a filtered and original time-series for all considered filters (TFP, SFP, RFP, MFP, PFP, CFP). The cross
denotes the case of an increase of the probability for the model to be included in MCS, while the dot denotes the cases when there is no such increase. The total
numbers of the cases when p > 0 for given market and filter are presented in parentheses.

It is worth noting that the application of pre-filtering for the ATS EU
power market results in the fact that the biggest accuracy gain is found
for the scale parameter values that are different from the ones applied
for the original electricity price time-series. But, as it can be seen from
Table 1, for filters TFP, SFP, and CFP these parameters are not that
different and are equal, correspondingly, 10, 10, and 11. The same can
be said about the forecast errors, because they are approximately equal
to each other (4.43%, 4.44%, and 4.47%, correspondingly).
For the ATS SI power market, the application of any of the filters with
any of the SCARX models results in a decrease of the forecast error, i.e.
in 100% of the cases we witness the advantages of pre-filtering for
outliers, regardless of the applied filter (of the ones considered in
Section 2.2.2). Here, the biggest accuracy gain of the averaged 1.44% is
yielded by the CFP filter, while the second and the third best filters here
are, correspondingly, RFP (forecast error decrease of 1.39%) and SFP
(forecast error decrease of 1.35%).
The most prominent decrease in WMAE for the above-mentioned
models (relative to SRC) for the ATS SI power market is 1.90% which is
yielded by the SCARX12-RFP model. A little smaller value of 1.85% is
demonstrated by the SCARX12-CFP model that uses the combined filter.
The biggest accuracy gain (0.12% over the ARX-model) for different
scale parameters on the original data is obtained with the SCARX13-
model. For all the scale parameter values and filtering procedures for
the ATS SI power market, the SCARX11-RFP model has the highest
precision, whose forecast error is only 7.34% which is 1.38% smaller
than the one for the SCARX13-model. Overall, by using the SCARX
models along with outlier filtering, the biggest accuracy gain is 1.5%.
With the average weekly error of the ARX model of 8.58% for the ATS
SI electricity market, such a decrease is very substantial and potentially
will let the market participants to get a significant financial efficiency.
Note that the SCARX13-CFP model that uses the proposed combined
filter (see Section 2.2.2) demonstrates the weekly error of 7.35% which
is almost equal to the error of the above-mentioned best model.
For the Nord Pool power market the situation is less optimistic as
compared to the above-described markets. In this case, though the best
results are again shown by filters TFP, SFP, and CFP, but SFP gives
accuracy gain only for 4 of the 6 models, while the two other methods –
only for 3 of the 6 models. The average decrease of WMAE (in those
cases where we have an accuracy gain) for TFP is 0.13%, for SFP is
0.07%, and for CFP is 0.05%. Thus, this indicator is smaller than the
one for the ATS EU market, and is quite considerably less than one for
the ATS SI market. Filters RFP and PFP never show any advantages of
their application, while MFP allows to decrease the error for one model
only (the SCARX8-MFP model), though not substantially (only by
0.05%).
Nevertheless, the maximum possible accuracy gain that we are able
to obtain (relative to the simple ARX-model) due to both the application
of the SCARX models, and pre-filtering of price spikes, is 0.29%. Of this
figure, 0.20% are yielded by the SCARX10-model, and the rest 0.09% –
by the TFP filter. The SCARX10-SFP model gives almost same results,
and the SCARX10-CFP model provides a little smaller change in WMAE
(specifically, 0.22%). These results look more significant as related to

205
D.O. Afanasyev, E.A. Fedorova
the average accuracy gain described above.
Finally, for the PJM power market outlier pre-filtering does not
produce any accuracy gain for either of the filters. It is important to
note that for this market application of the SCARX model does not seem
to be justified at all, because for any of the considered scale parameters
the average weekly error is bigger than the one for the simple ARX-
model. With other test samples (markets) there having situations when
the SCARX model for this or that m demonstrates lower accuracy than
the ARX model, and the application of filtering procedures allowing to
correct it (see, for example, the SCARX8 model on the original data and
with filters TFP, SFP, and CFP for the ATS EU power market or the
SCARX12-model with any of the filters for the ATS SI power market),
this does not happen to the PJM power market. Taking this into ac-
count, we can conclude that in the case of the PJM market it is ne-
cessary to look more critically at the inspiring conclusions of
Nowotarski and Weron [9] who say that the SCARX model, with
properly chosen filter and smoothing parameter, yields accuracy gain
relative to the ARX-model.
Let us now consider the results of comparison of the filters’ per-
formance in the sense of a formal statistical inference based on iden-
tification of model confidence set (MCS). Recall that the MCS naturally
ranks the models by the probability of their inclusion in this set, i.e. by
the MCS p-value (see Section 2.4). Thus, if an outlier filter’s application
for a specific model (in our case, for a given value of m) results in an
increase of its MCS p-value (relative to the model without pre-filtering),
then we say that such filter increases predictive ability of the model,
since it now has a higher probability to be included into the MCS.
For each of the filters and all smoothing parameters m = 8, …, 13 we
build sets of forecasts obtained on the full test period for original and
outlier filtered time-series (twelve models per set). After that, for each
of the sets (i.e., for each filter) we ranked models by MCS p-values and
found out the number of cases when pre-filtering led to an increase of
the probability for a model to be included in MCS, as well as the
number of cases when this does not happen (i.e., the probability is ei-
ther unchanged or decreases). The obtained final results are given in
Fig. 4.
Analysis of Fig. 4 shows that the obtained results of estimation of
outlier filtering influence on the predictive ability (in the sense of MCS)
support the findings obtained with WMAE . But we should note the
peculiarity that for the ATS EU and ATS SI power markets the com-
parison of performance by MCS, on the average, somewhat rarer de-
monstrates accuracy gain with the use of outlier filters than the com-
parison by WMAE . For example, for the ATS SI power market a
decrease in WMAE is found for all the filters in a 100% of cases (see
Table 1). At the same time, as it can be seen in Fig. 4 (top-right panel),
increases in MCS p-value occur, at most, in 5 cases out of 6 (filters TFP,
RFP, PFP, and CFP), and at least – in 4 cases out of 6 (filters SFP and
MFP). In turn, for the NP and PJM power markets comparisons both by
MCS, and by WMAE , lead to identical conclusions.
Summarizing the above-described results, we make the following
conclusions. Firstly, outlier filtering often allows to get an accuracy gain
for electricity price prediction, with the forecast error being sub-
stantially smaller in a number of cases. Secondly, the selection of a
specific price outlier filter that will give a decrease of forecasting error
is generally unobvious. Nevertheless, the studied TFP and SFP filters
have a very good performance: TFP yields accuracy gain (in the WMAE
sense) in 63% of cases and in 58% of cases increases the probability of a
model to be included in MCS, while SFP has 67% and 54%, corre-
spondingly (see the summary for all markets in Table 1 and Fig. 4).
Comparatively worse results are produced by the MFP filter which has
advantages by WMAE only in 50% of cases. Although the decrease of
forecast error for MFP is substantially smaller than for TFP and SFP.
With the use of MCS, the performance of MFP is even less inspiring:
only in 38% of the cases MCS p-values increase due to the application of
MFP.
Thirdly, the proposed combined CFP filter is a well competing
206
Applied Energy 236 (2019) 196–210
alternative to TFP and SFP. Indeed, in 63% of the considered cases it
actually improves the accuracy of electricity price forecasting, while its
forecast error is insubstantially greater than the ones of TFP and SFP, if
any. Usage of a formal statistical inference on the basis of model con-
fidence set confirms the improvement of prediction ability in 54% of
cases. Two main advantages of CFP should be emphasized: (1) this filter
is less subjective as compared to the individual filtering procedures
because it takes account of several independent methods simulta-
neously; (2) even when within CFP several less efficient (in the sense of
higher forecast error) individual filters are used, the final result is still
comparable with the most efficient individual filters (TFP and SFP).
Thus, the application of CFP, on the one hand, partly mitigates the risks
of an a priori choice of a filter, and on the other hand, at least, performs
not worse than the best individual filters.
Finally, fourthly, though outlier filtering with different filters often
leads to the accuracy gain, practically speaking, there may be situations
when both such pre-processing, and the approach of separate modeling
of the long-term and short-term dynamics itself result only in worsening
the accuracy of forecasting. This is why the decision on outlier pre-
filtering of the electricity price time-series to be predicted, should be
taken, at least, after testing of this or that filter for a specific power
market over the closest historical periods.
4.3. Equivalence of the outlier filters
As it was shown above, not all of the considered filters lead to in-
spiring results in the sense of forecast accuracy. At the same time, all
the filters implement a similar principle: detect a price threshold and
split the observations into two classes (as related to this threshold).
Thus, there arises a question of what are the causes of such differences
among the principally similar methods?.
As it was mentioned in the beginning of the paper, outlier sub-
stitution, on the one hand, may lead to accuracy gain, but on the other
hand, a fanatic outlier filtering on the original data may result in sub-
stantial distortion of the properties of the data which is expected to
increase forecasting error. Evidently, the aggressiveness of outlier fil-
tering is reflected by the percent of removed outliers, which is directly
linked to the parameter of this or that filter. Taking this into account, it
can be assumed that the causes of the obtained above non-satisfactory
results for some of the filters lie in non-optimality of their parameter
values. To test this hypothesis, we run the following experiment.
For each of the markets, we select one SCARX model that yields the
smallest forecast error in the original (SRC) data, specifically: SCARX13
for ATS EU and ATS SI, SCARX10 for NP, and SCARX12 for PJM (see
Table 1). Further, we consider four filters, two (TFP, SFP) out of which
earlier (see Section 4.2) demonstrated accuracy gain in the majority of
cases, and the other two (RFP, MFP) basically resulted in an increase in
forecast error. For each of the markets we choose a range of values for
the main parameter of the filter: for TFP – the threshold for logarithmic
price; for SFP and RFP – the number of standard deviations; for MFP –
the predictive interval13. Note that for the filters under consideration
increasing value of their main parameters is inversely related to the
percentage of identified outliers. Further, on this grid of parameters we
repeat the previously conducted empirical experiment and estimate
empirical dependence between WMAE , averaged over the test period,
and the percentage of removed outliers (also averaged over all the ca-
libration windows). Fig. 5 shows the obtained results of parameter
optimization. Analysis of the graphs of the empirical dependencies
leads to the following interesting conclusions.
Firstly, our hypothesis on the existence of dependence of the forecast
13
For MFP the choice of the predictive interval essentially automatically
determines the number of standard deviations, exceeding which will classify the
observation as an outlier. But this happens indirectly under the assumption of
normality of distribution of the logarithmic prices.
D.O. Afanasyev, E.A. Fedorova Applied Energy 236 (2019) 196–210

Fig. 5. Dependence of WMAE , averaged over the test period, on the average percent of removed outliers for filters TFP, SFP, RFP, and MFP for power markets ATS
Europe-Ural (model SCARX13), ATS Siberia (model SCARX13), Nord Pool (model SCARX10), and Pennsylvania-New Jersey-Maryland (model SCARX12). Horizontal
dot-dashed lines in the graphs represent the error of the corresponding model on the original data (SRC); horizontal dashed lines – the error of the model with a filter
and an a priori chosen parameter (see Table 1). The location of the dashed line over the dot-dashed line corresponds to the situation when the error of the model with
the a priori filter parameter exceeds the error of the model on the original data. As it can be seen from the graphs, in many cases the choice of a proper value of a
filter’s parameter allows to improve the situation.

accuracy on aggressiveness of outlier removal finds empirical support. Table 2

It can be seen in the majority of the graphs that as the fraction of re- Minimum averaged WMAE (or marginal performance) for the most precise
moved outliers grows, the weekly error, at first, decreases from the SCARX models, as found by the out-of-sample optimization of parameter values
value on the original data (i.e., from 0% of removed outliers; the dot- for filters TFP, SFP, RFP, and MFP on power markets ATS Europe-Ural
dashed line) to some minimum level in inherent saddle point, after (SCARX13), ATS Siberia (SCARX13), Nord Pool (SCARX10), and Pennsylvania-
New Jersey-Maryland (SCARX12).
which it starts growing. The founded point is a point of filter marginal
performance that can not be exceeded (for a given forecasting model). Market Outlier filters
Secondly, the proper choice of the parameter value for this or that
SRC TFP SFP RFP MFP CFP∗
filter is crucially important to receive accuracy gain. Indeed, as it can be
seen in the majority of graphs in Fig. 5, on the one hand, the lower is EU 4.62 4.44 (0.75/ 4.45 (6.5/ 4.44 (9.0/ 4.46 (1.000/ 4.56
the fraction of removed outliers, the less prominent is the influence of 0.3%) 0.2%) 0.2%) 0.3%)
outlier filtering on the forecast error. On the other hand, increasing this SI 8.72 7.43 (0.75/ 7.36 (2.0/ 7.33 (2.0/ 7.71 (0.950/ 7.35
1.9%) 1.8%) 2.2%) 4.2%)
fraction (through lowering the filter parameter value) bears the fol-
NP 8.38 8.23 (0.75/ 8.23 (4.5/ 8.25 (6.0/ 8.23 (0.999/ 8.31
lowing quite probable risk: even insubstantial deviation from the op- 0.5%) 0.4%) 0.3%) 0.9%)
timal parameter value (1–2 steps on the grid) towards increasing of the PJM 10.60 10.57 10.58 (6.5/ 10.58 (6.0/ 10.60 (1.000/ 11.21
fraction of outliers results in a prominent increase in the forecasting (1.50/ 0.1%) 0.1%) 0.1%)
error which ultimately becomes even higher than for the original data. 0.1%)

Finally, thirdly, though a number of filters with a priori defined

parameters were found as not giving accuracy gain (specifically, filters
RFP and MFP; see Section 4.2), still, with a proper selection of the
parameter value, they also prove to be quite efficient tools for de-
creasing the average forecast error. As it can be seen in Table 2 which
gives minimum values of WMAE (or marginal performance) and the
corresponding values of filter parameter, for three of the markets there
is a decrease in forecasting error due to the application of all four filters
(relative to the model on the original data). And within a market, the
minimum values of WMAE appear to be quite similar for each of the
Comments: the optimal parameter values and the corresponding average frac-
tion of removed outliers for the specified WMAE are given in parentheses. SRC
denotes the results for the original, unfiltered data. CFP∗ is the combined filter
based on the individual filters with unoptimized parameters, which was used
earlier in Section 4.2.
filters. This allows to conclude that selection of a specific individual
filter is intrinsically not a principal problem – it is much more im-
portant to set the filter parameter appropriately. Thus, the reasons why
filters RFP and MFP did not show earlier a positive impact on the

207
D.O. Afanasyev, E.A. Fedorova
forecast accuracy lie, primarily, in inadequate values of their price
threshold parameters.
It also should be noted that although for the PJM power market
even filter parameter optimization did not allow to substantially de-
crease the forecast error, still, at least, we were able to make the
forecasting accuracy (after filtering with the optimal parameter value)
not fall relative to the model for the original data. But this is achieved
only by filtering out no more than 0.1% of outliers, with further low-
ering of the filter threshold resulting in a linear forecasting error in-
crease.
4.4. Practical recommendations
Summarizing the above-given results of numerical experiments, we
propose the following strategy for practical application of outlier fil-
tering aimed at electricity price forecasting error reduction for a con-
crete day-ahead power market.
First step is basic empirical validation of the outlier filtering effi-
ciency in the sense of forecasting accuracy gain. To this end, we re-
commend to use historical data on the market under study, and apply
the ”rolling forecasting origin” principle (with calibration window not
narrower than one year). TFP with the threshold of 0.5 and SFP with 3
standard deviations can be applied as reference filters. This is motivated
by the fact that on principally different markets (considered in this
research) these methods most often lead to a forecasting error decrease
(even with a priori defined cut-off thresholds). This simple step allows
to sort out situations when the use of filters is not justified, as it could
be observed above for the PJM market. But if the result is successful
then the practitioner may decide to apply outlier filtering and move to
the next step.
Second step implies two alternatives. First (obvious) option is to
grid-optimize the outlier filter cut-off parameter as it was done in
Section 4.3 while the choice of the filtration method itself is of no great
importance (as it was demonstrated above). However, there are some
complications with this approach: seeking for the optimal parameter
value, on the one hand, is quite time-consuming14 that may be crucial
for real applications, but on the other hand, practically speaking,
cannot be performed by the above-used scheme since it implies avail-
ability of the future information15.
Taking this into account, we recommend employing the proposed
combined filter on prices (CFP) as an efficient second alternative. This
recommendation comes from comparison of the minimum forecast er-
rors of grid-optimized individual filters with the error of CFP which is
shown in Table 2. It can be concluded that these errors are almost equal
for a specific market: they either coincide, or CFP demonstrates slightly
lower accuracy (though, in a number of cases, CFP outperforms other
filters). Thus, in practice the CFP filter based on the committee machine
is a strong competitive alternative to the time-consuming optimization
of a specific filter parameter.
The proposed CFP method presumably could be used not only for
forecasting purposes, but also for in-sample modeling of electricity
price since it eliminates the risk of insufficiency of an a priori defined
threshold parameter. The outlier pre-filtering based on CFP allows to
prevent a bias in the models’ coefficient estimates which is caused by
spikes, and to obtain a more accurate value of the residuals’ distribution
variance. This will lead to more correct estimation of the value-based
risk measures such as Value-at-Risk and Earning-at-Risk, which are
14
Indeed, high density of points near the WMAE-axis in the graphs of Fig. 5
speaks in favor of slow change in the fraction of removed outliers as the main
filter parameter decreases (with a constant step of the grid). This makes con-
secutive selection of the value inefficient and resource-consuming.
15
Undoubtedly, the filter parameters can be optimized on historical data. But
the applicability of the values thus obtained to the unavailable future data is
doubtful.
208
Applied Energy 236 (2019) 196–210
wide-spread in the risk-management practice while being highly sen-
sitive to distortions in the corresponding models’ residuals distribu-
tions.
Let us note that while this research has tested different outlier filters
in the framework of application of concrete SCARX forecasting model,
the obtained finding are adequate for a wide class of time-series models.
This statement is based on the fact that filtering is conducted in-
dependently from the modeling process as it can be seen in the meth-
odology flowchart (see Fig. 1).
5. Conclusion
In this research we rise the question of outlier pre-filtering impact
on the accuracy of electricity price forecasting. While direct modeling
seems to be a more natural way to take account of the price spikes,
preliminary identification and replacement of extreme values can be
considered as a simpler and less time-consuming alternative. At the
same time, outlier filtering is a quite common approach for in-sample
price modeling while its relevance to problem of forecasting has not
been sufficiently addressed in the existing literature.
Using the seasonal component autoregressive model we compare
performance of the following filters: threshold filter on prices (TFP),
standard deviation filter on prices (SFP), percentage filter on prices
(PFP), recursive (RFP) and moving (MFP) filters on prices. We also
propose a combined filter on prices (CFP) based on a committee ma-
chine with majority vote rule. For the empirical experiment, we use
four historical datasets of day-ahead electricity markets: the Europe-
Ural and Siberia areas of the Russian ATS market; the biggest European
market Nord Pool, as well as the US power market Pennsylvania-New
Jersey-Maryland. The performance evaluation is based on two in-
dependent approaches – weekly weighted mean absolute error and a
formal statistical procedure of model confidence set (MCS) identifica-
tion. The obtained results allowed us to make the following principal
conclusions.
Firstly, outlier pre-filtering of the original data often lead to a sub-
stantial forecasting accuracy gain. For example, in the Siberia area of
the Russian ATS market the obtained absolute decrease of forecasting
error achieved 1.8–1.9% of the average weekly price. This improve-
ment of accuracy potentially may bring significant financial saving to
the market participants.
Secondly, among the studied filters with an a priori set cut-off
threshold parameter, the most stable positive results were demon-
strated by TFP (with the threshold for the log-price being equal to 0.5),
SFP (with the threshold being equal to 3 standard deviations), and CFP
(using the majority vote rule). TFP provided accuracy gain in 63% of
the considered cases, SFP – in 67%, and CFP – in 63%. In the context of
MCS-based ranking, increase of probability for a model to get into the
MCS occurred for these filters in 58%, 54%, and 54% cases, corre-
spondingly. Despite the a priori defined cut-off parameter, the TFP and
SFP can be applied to different power markets as efficient and fast tools
to achieve forecasting accuracy gain.
Thirdly, the main cause of the filters’ performance discrepancy lies
not in the difference of their algorithms, but in the improper a priori
choice of the cut-off threshold parameter. If a grid-search to find fore-
cast error-minimizing value of the filter parameter is run, then the
difference between the prediction ability of filters is be negligible. In
particular, RFP and MFP, after finding the optimal values of their
parameters, allowed to obtain forecast errors only slightly different
from the ones of TFP and SFP for each of the markets. In this sense, the
all considered filters have equivalent marginal forecasting performance.
Fourthly, the proposed combined filter based on the committee
machine is a quite competitive alternative to the individual methods
both in the case of a priori set parameters, and in the case of grid-search
optimized values of these parameters. In the former case, CFP plays a
role of a more objective method which mitigates the risk of in-
sufficiency of an a priori defined, standalone filter parameter, as it takes
D.O. Afanasyev, E.A. Fedorova
into account the results of each of the independent filters. In the latter
case, it should be kept in mind that grid-search procedures are time-
[14]
consuming and cannot be applied to the future, practically unavailable
data, while CFP uses un-optimized filters as basis algorithms and is not
resource demanding, but at the same time results in forecasting accu- [15]
racy which is comparable to the best of individual filters after optimi- [16]
zation of their threshold parameters.
Finally, fifthly, although the above-given conclusions seem in-
[17]
spiring, still, in a rare case outlier filtering may worsen the results of
forecasting. The example of such a case, among the considered power [18]
markets, is the PJM. This is why the decision on application of pre-
filtering should be taken with care. Validation on historical data with
one of the reference filters (TFP or SFP) may be useful here. [19]
Although the results of this study shed some light on certain aspects
of the short-term electricity price forecasting, still there arise a number
[20]
of new questions. Specifically, a combination of forecasts is a promising
technique of increasing forecasting accuracy [42,51]. But, as it was [21]
shown in a number of previous papers, not all of the combination
[22]
methods allow to obtain good results [44,45]. From this perspective, it
would be challenging to understand whether pre-filtering helps in sol- [23]
ving of existing here problems. Another challenging question is which is
[24]
a more efficient tool for obtaining accuracy gain – combination of filters
or combination of forecasts on filtered data? We leave these and other [25]
questions for further research.
[26]
Acknowledgments
[27]
The authors are grateful to Gilenko E.V. (Ph.D., St. Petersburg State
[28]
University, Russia) for his valuable and stimulating comments on the
paper, as well as assist for translation. This paper also greatly benefited [29]
from the helpful comments and suggestions of three anonymous re-
[30]
viewers and Prof. J. Yan (Editor-in-chief). This work was supported by
the Russian Foundation for Basic Research (RFBR) [Grant No. 16-06-
00237 A]. The given in this study point of view of Dmitriy O. Afanasyev [31]
is not the official position of his employer JSC Greenatom and may not [32]
coincide with it.
[33]
Appendix A. Supplementary material
[34]
Supplementary data associated with this article can be found, in the
online version, at https://doi.org/10.1016/j.apenergy.2018.11.076.
[35]
References
[36]
[37]
[1] Casazza J, Delea F. Understanding electric power systems: an overview of the
technology and the marketplace. Hoboken, New Jersey: Wiley; 2003. [38]
[2] Eydeland A, Wolyniec K. Energy and power risk management. 2nd ed.. Hoboken,
New Jersey: Wiley; 2012. [39]
[3] Carmon R, Coulon M. A survey of commodity markets and structural models for
electricity prices. Quantitative Energy Finance. New York: Springer; 2014. p. 41–83. [40]
[4] Guirguis H, Felder F. Further advances in forecasting day-ahead electricity prices
using time series models. KIEE Int Trans Power Eng 2004;4:159–66. [41]
[5] Karakatsani N, Bunn D. Forecasting electricity prices: the impact of fundamentals
and time-varying coefficients. Int J Forecast 2008;24:764–85. [42]
[6] He K, Yu L, Tang L. Electricity price forecasting with a BED (Bivariate EMD
Denoising) methodology. Energy 2015;91:601–9. [43]
[7] Aggarwal SK, Saini LM, Kumar A. Electricity price forecasting in deregulated mar-
kets: a review and evaluation. Int J Electr Power Energy Syst 2009;31(1):13–22.
[8] Weron R. Electricity price forecasting: a review of the state-of-the-art with a look [44]
into the future. Int J Forecast 2014;30:1030–81.
[9] Nowotarski J, Weron R. On the importance of the long-term seasonal component in [45]
day-ahead electricity price forecasting. Energy Econ 2016;57:228–35.
[10] Keles D, Scelle J, Paraschiv F, Fichtner W. Extended forecast methods for day-ahead
electricity spot prices applying artificial neural networks. Appl Energy [46]
2016;162:218–30.
[11] Marcjasz G, Uniejewski B, Weron R. On the importance of the long-term seasonal [47]
component in day-ahead electricity price forecasting with NARX neural networks.
Int J Forecasting 2018. https://doi.org/10.1016/j.ijforecast.2017.11.009. [48]
[12] Tan Z, Zhang J, Wang J, Xu J. Day-ahead electricity price forecasting using wavelet
transform combined with ARIMA and GARCH models. Appl Energy [49]
2010;87(11):3606–10.
[13] Kurbatsky V, Tomin N. Forecasting prices in the liberalized electricity market using [50]
209
Applied Energy 236 (2019) 196–210
the hybrid models. In: Energy conference and exhibition (EnergyCon), 2010 IEEE
international; 2010. p. 363–68.
Misiorek A, Trück S, Weron R. Point and interval forecasting of spot electricity
prices: linear vs. non-linear time series models. Stud Nonlinear Dyn Econ
2006;10(3):Article 2.
Weron R, Misiorek A. Forecasting spot electricity prices: a comparison of parametric
and semiparametric time series models. Int J Forecast 2008;24:744–63.
Conejo A, Plazas M, Espínola R, Molina A. Day-ahead electricity price forecasting
using the wavelet transform and ARIMA models. IEEE Trans Power Syst
2005;20:1035–42.
Weron R. Modeling and forecasting electricity loads and prices: a statistical ap-
proach. Chichester: Wiley; 2006.
Trück S, Weron R, Wolff R. Outlier treatment and robust approaches for modeling
electricity spot prices. MPRA Paper No. 4711. Hugo Steinhaus Center, Wroclaw
University of Technology; 2007. URL <http://mpra.ub.uni-muenchen.de/4711/1/
MPRA_paper_4711.pdf>.
Janczura J, Trück S, Weron R, Wolff R. Identifying spikes and seasonal components
in electricity spot price data: a guide to robust modelings. Energy Econ
2013;38:96–100.
Cuaresma JC, Hlouskova J, Kossmeier S, Obersteiner M. Forecasting electricity spot-
prices using linear univariate time-series models. Appl Energy 2004;77(1):87–106.
Cartea A, Figueroa M. Pricing in electricity markets: a mean reverting jump diffu-
sion model with seasonality. Appl Math Finance 2005;12(4):313–35.
De Jong C. The nature of power spikes: a regime-switching approach. Stud
Nonlinear Dyn Econ 2006;10(3). [Article 3].
Janczura J, Weron R. An empirical comparison of alternate regime-switching
models for electricity spot prices. Energy Econ 2010;32:1059–73.
Amjady N, Keynia F. Electricity market price spike analysis by a hybrid data model
and feature selection technique. Elect Power Syst Res 2010;80(3):318–27.
Zhao J, Dong Z, Li X. Electricity market price spike forecasting and decision making.
IET Generation Trans Distribution 2007;1(4):647–54.
Voronin S, Partanen J. Price forecasting in the day-ahead energy market by an
iterative method with separate normal price and price spike frameworks. Energies
2013;6:5897–920.
Lapuerta C, Moselle B. Recommendations for the Dutch Electricity Market. London:
The Brattle Group Report; 2001.
Fanone E, Gamba A, Prokopczuk M. The case of negative day-ahead electricity
prices. Energy Econ 2013;35:22–34.
Clewlow L, Strickland C. Energy derivatives - pricing and risk management. Lacima
Publications; 2000.
Borovkova S, Permana FJ. Modelling electricity prices by the potential jump-dif-
fusion. In: Shiryaev AN, Grossinho MR, Oliveira PE, Esquível ML, editors. Stochastic
Finance. US, Boston, MA: Springer; 2006. p. 239–63.
Nowotarski J, Tomczyk J, Weron R. Robust estimation and forecasting of the long-
term seasonal component of electricity spot prices. Energy Econ 2013;39:13–27.
Lisi F, Nan F. Component estimation for electricity prices: procedures and com-
parisons. Energy Econ 2014;44:143–59.
Afanasyev D, Fedorova E. The long-term trends on the electricity markets: com-
parison of empirical mode and wavelet decompositions. Energy Econ
2016;56:432–42.
Xiao L, Shao W, Yu M, Ma J, Jin C. Research and application of a hybrid wavelet
neural network model with the improved cuckoo search algorithm for electrical
power system forecasting. Appl Energy 2017;198:203–22.
Yang Z, Ce L, Lian L. Electricity price forecasting by a hybrid model, combining
wavelet transform, ARMA and kernel-based extreme learning machine methods.
Appl Energy 2017;190:291–305.
Mitchell T. Machine learning. McGraw-Hill Science/Engineering/Math; 1997.
Hastie T, Tibshirani R, Friedman J. The elements of statistical learning. Springer;
2001.
Haykin S. Neural networks: a comprehensive foundation. 2nd ed.. Prentice-Hall;
1998.
Ablow C, Kaylor D. A committee solution of the pattern recognition problem. IEEE
Trans Inf Theory 1965;11:453–5.
Mazurov V. Committee of inequality systems and the recognition problem.
Kibernetika 1971;3:140–6. [in Russia].
Tresp V. Committee machines. In: Hu Y, Hwang J-N, editors. Handbook for Neural
Network Signal Processing. CRC Press; 2001. p. 135–52.
Bates J, Granger C. The combination of forecasts. J Operational Res Soc
1969;20(4):451–68.
Timmermann A. Chapter 4 forecast combinations. Elliott G, Granger C,
Timmermann A, editors. Handbook of Economic Forecasting, vol. 1. Elsevier; 2006.
p. 135–96.
Bordignon S, Bunn D, Lisi F, Nan F. Combining day-ahead forecasts for British
electricity prices. Energy Econ 2013;35:88–103.
Nowotarski J, Raviv E, Trück S, Weron R. An empirical comparison of alternate
schemes for combining electricity spot price forecasts. Energy Econ
2014;46:395–412.
Kittler J, Hatef M, Duin R, Matas J. On combining classifiers. IEEE Trans Pattern
Anal Mach Intell 1998;20(3):226–39.
Weron R. Market price of risk implied by Asian-style electricity options and futures.
Energy Econ 2008;30:1098–115.
Bierbrauer M, Menn C, Rachev S, Trück S. Spot and derivative pricing in the EEX
power market. J Banking Finance 2007;31:3462–85.
Hansen P, Lunde A, Nason J. Model confidence sets for forecasting models. Working
Paper 2005-7, Federal Reserve Bank of Atlanta; 2005.
Hansen P, Lunde A, Nason J. The model confidence set. Econometrica
D.O. Afanasyev, E.A. Fedorova Applied Energy 236 (2019) 196–210

2011;79:453–97. high-dimensional models: the case of Brazil. Int J Forecast 2017;33(3):679–93.

[51] Nan F. Forecasting next-day electricity prices: from different models to combination
[Ph.D Thesis]. Universita degli Studi di Padova; 2009. URL <http://paduaresearch.
cab.unipd.it/2147/>.
[52] Samuels JD, Sekkel RM. Model confidence sets and forecast combination. Int J
Forecast 2017;33(1):48–60.
[53] Garcia MG, Medeiros MC, Vasconcelos GF. Real-time inflation forecasting with
[54] Diebold F, Mariano R. Comparing predictive accuracy. J Business Econ Statistics
1995;13:253–63.
[55] White H. A reality check for data snooping. Econometrica 2000;68:1097–126.
[56] Politis DN, Romano JP. The stationary bootstrap. J Am Stat Assoc
1994;89:1303–13.

210