You are on page 1of 7

Electrical Power and Energy Systems 61 (2014) 2026

Contents lists available at ScienceDirect

Electrical Power and Energy Systems


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

A time series spot price forecast model for the Nord Pool market
Tarjei Kristiansen
9000 lborg, Denmark

a r t i c l e

i n f o

Article history:
Received 11 January 2013
Received in revised form 16 February 2014
Accepted 7 March 2014
Available online 2 April 2014
Keywords:
Nord pool
Spot price forecasting
Regression model
Electricity market

a b s t r a c t
We present three relatively simple spot price forecast models for the Nord Pool market based on historic
spot and futures prices including data for inow and reservoir levels. The models achieve a relatively
accurate forecast of the weekly spot prices. The composite regression model achieves a mean absolute
percentage error (MAPE) of around 7.5% and under-forecasts the actual spot price by some 1.4 NOK/
MW h in the sample period. Out of sample testing achieves a MAPE of around 7.4% including a match
of the actual spot price. A myopic model using the previous weeks spot price as a predictor for the next
weeks spot price achieves a MAPE of 7.5% and under-forecasts the actual spot price by some 0.9 EUR/
MW h. A futures model using the futures price for next week as a predictor for next weeks spot price
achieves a MAPE of 5.3% and over-forecast the actual spot price by some 4.3 EUR/MW h.
2014 Elsevier Ltd. All rights reserved.

Introduction
Short-term price forecasting presents a crucial activity for the
protability of generation companies, end-user companies and
trading companies. It enables them to optimize their trading strategies. The planning horizon for generation and consumer companies typically spans weeks and/or months. Thus there is a need
for models that can forecast the average weekly spot prices. Time
series models represent an important category of models [1]. These
models typically utilize some external variables related to supply
and/or demand to explain pricing behavior. Other approaches include multiple support vector machine [2], extreme learning machine [3], auto-regressive fractionally integrated moving average
[4] and Grey models [5].
The most commonly applied model for medium term spot price
forecasting in the Nord Pool market is the EMPS model (or Samkjringsmodellen). It is a fundamental stochastic dynamic programming model for hydro-thermal scheduling. The results from
the model are proprietary and differ for each market player making
it difcult to facilitate a comparison. Likewise the power market
analytics service provider Point Carbon [6] utilizes a stochastic
dual dynamic programming (SDDP) model which they claim yields
improved accuracy compared to the EMPS model. However they
provide no numerical results to back up their statements. In sum,
these models may appear as black box approaches to outsiders
which do not have access to the inputs. Therefore they are unable
to adjust these if necessary.
Tel.: +45 201144160.
E-mail address: tarjeikr@yahoo.com
http://dx.doi.org/10.1016/j.ijepes.2014.03.007
0142-0615/ 2014 Elsevier Ltd. All rights reserved.

In this paper we develop three relatively simple spot price models that utilize previous weeks spot prices, futures prices including
data for inow and reservoir levels. The models provide relative
high precision spot price forecasts for the Nord Pool market.
The structure of the paper is as follows. Section Literature
review provides a literature review. Section The Nordic power
market describes the Nordic power market and its fundamental
parameters. Section Spot forecasting models develops three spot
price forecasting models. Section Model performance measures
the model performances. Section Conclusion concludes.
Literature review
Weron [7] analyzes forecasting of day-ahead electricity prices
but nds that only some approaches are suitable. Time series models are among the models with strongest forecasting ability. Weron
[7] also identies that single hour model specications achieve
better forecasting accuracy than multi-hour model specications.
Time series models include:
 Auto-regressive (AR) models.
 Auto-regressive moving average (ARMA).
 Auto-regressive integrated moving average (ARIMA) and seasonal
ARIMA models [8,9].
 Auto regressions with heteroskedastic [10].
 Heavy-tailed innovations [11].
 AR models with exogenous (fundamental) variables.
 Dynamic regression (or ARX) and transfer function (or ARMAX)
models [12].
 Vector auto regressions with exogenous effects [13].

T. Kristiansen / Electrical Power and Energy Systems 61 (2014) 2026

 Threshold AR and ARX models [14].


 Regime-switching regressions with fundamental variables [15].
 Mean-reverting jump diffusions [16].
Kristiansen [17] developed an auto-regressive ARX model for
Nord Pool day ahead prices. The model is based on Weron and
Misiorek [11] but reduced in terms of estimation parameters
(from 24 sets to 1) and modied to include Nordic demand and
Danish wind power as exogenous variables. Prices are modeled
across all hours in the analysis period rather than across each single 24 h. By applying three model variants on Nord Pool data, a
weekly mean absolute percentage error (WMAE) of around 6
7% and an hourly mean absolute percentage error (MAPE) ranging
from 8% to 11% were achieved. Out of sample results yielded a
WMAE and an hourly MAPE of around 5%. Kristiansen [18] also
developed a fundamental stack model for the German electricity
market.
Weron and Misiorek [11] achieved a weekly-weighted mean
absolute error of 4.66% for four ve-week periods (i.e. 20 weeks)
from 1998 to 1999 and 3.37% for four ve-week periods (i.e.
20 weeks) from 2003 to 2004. Kristiansens [17] weekly-weighted
mean absolute errors are somewhat larger but he has formulated
a different model and included a larger and more recent data period so results are not directly comparable.

21

Torro [1] developed an ARIMAX model to forecast weekly futures prices in the Nord Pool market for the period 19972003.
The time-series model contains external variables such as temperature, precipitation, reservoir levels and the basis (futures price
minus the spot price) which, overall, reects the typical seasonal
patterns in the weekly spot price. Torro [1] achieved an average error of 0.78 NOK/MW h for one week ahead futures but provided
no model for spot price forecasting. Torro [1] also proposed two
alternative and simpler forecasting methods:
 A myopic method that utilizes the present spot price as a forecast for next week. Torro [1] explains that a myopic method can
be considered as the minimum accuracy deemed from any forecasting method.
 A futures method that utilizes the present futures price as a
proxy for next weeks spot price.
We utilize these models as benchmarks in this paper.

The Nordic power market


The Nordic power exchange, Nord Pool, was established in 1993
and consisted of Norway, subsequently joined by Sweden in 1996.

Fig. 1. Average, minimum and maximum weekly spot prices in the Nord Pool market from 1999 to 2010.

Fig. 2. Average, minimum and maximum reservoir levels in percentage in Norway for the period 19992010.

22

T. Kristiansen / Electrical Power and Energy Systems 61 (2014) 2026

Fig. 3. Average, minimum and maximum weekly inow levels in GWh for Norway and Sweden for the period 19992010.

Finland in September 1998, Western Denmark in January 1999,


and eastern Denmark in October 2000. Nord Pool operates the spot
market [19] while NASDAQ OMX operates the nancial market.
The daily spot market is cleared the day before operation based
on the market players offers to sell or bids to buy. The system price
is the price for Nordic market ignoring transmission constraints. It
is used as a settlement price for the forward and futures markets.
Transmission congestion causes area prices to differ within and between the Nordic countries.
Approximately 50% of the Nordic electricity is generated from
hydropower plants. Norway has close to 100% hydro while Sweden
and Finland have around 50% and 20% share, respectively. The
majority of the hydropower is hydro storage with potential for
multi-year storage in Norway. Hydro storage and demand inuence both the short- and long-term prices. Hydro storage levels
peak in SeptemberNovember and reach their lowest levels in
AprilMay. Conversely, demand levels peak in the winter (DecemberMarch) due to demand for electric heating. Electricity prices
thus have some seasonality and reach their lowest levels in the
period MayAugust. Healthy hydro supplies facilitate fast ramping
and exibility features. During demandsupply shocks prices are
less volatile compared to a pure thermal power system. However
during drought and ooding situations volatility may increase.
The historic system spot price and hydro storage levels in Norway exhibited a seasonal pattern during the period 19962001 as
shown in Fig. 1. However, the seasonal pattern weakened after
2001.
There is a relationship between hydro storage and the spot price
levels. Norway has the dominant share of storage capacity
(84.3 TWh) in the Nordic power system. The major price increases
in the fall/winter 2002/2003 and summer 2006 reected low reservoir levels (see Fig. 2). A similar low reservoir level in 1996/97 also
led to an increase in price. The period 20022006 was on average

drier than the period 19962001 (i.e. lower reservoir levels). The
period 20072010 included high reservoir levels for 2007 and
2009 while 2010 had low reservoir levels.
The relationship between reservoir levels and electricity futures
prices at Nord Pool has been studied by Gjolberg and Johnsen [20]
and Botterud et al. [21]. The authors consistently found that the
reservoir level and its seasonality dependence largely explain the
price formation for spot and future prices. Storage theory states
that seasonality generates seasonal patterns in spot and futures
electricity prices [1]. Reservoir levels impact the sensitivity of demand and supply shocks. A shock occurring at a high reservoir level is uncritical. Conversely a shock at a low reservoir level is
harder to match and may cause rising prices. Full reservoirs may
cause spill and reduced prots. Thus a negative convenience yield
(i.e. high reservoir levels) may cause producers to sell at lower
prices rather than risking spill [1]. The result would be lower spot
prices than futures prices. Conversely a positive convenience yield
(i.e. low reservoir level), may cause higher spot prices than futures
prices. Likewise, precipitation expectations inuence the power
price formation.
The inow levels to reservoirs from snow melting peak between
weeks 18 and 27 as shown in Fig. 3. Conversely during the winter,
inows are at their lowest levels for the weeks 117.

Spot forecasting models


In the following we propose three models to forecast average
weekly spot price in the Nord Pool market. The rst model assumes
that the average weekly spot price in the current week depends on
the average weekly spot price in the previous week (auto-correlation), the inow level in Norway and Sweden in the previous week
in GWh and the difference in reservoir level in Norway in the two

Table 1
Correlation between the natural logarithm of the de-meaned weekly spot prices in week t and week t  1, the natural logarithm of inow levels in week t  1 in Norway and
Sweden in GWh and the natural logarithm of change in reservoir level in week t  1 minus week t  2 in Norway in percent of maximum reservoir level.

ln average spot week t


ln average spot week t  1
ln inow (NO + SWE) (GWh) week
t1
ln change in reservoir levels % week
t1

ln average spot
week t

ln average spot week


t1

ln inow (NO + SWE) (GWh) week


t-1

1
0.97
0.32

1
0.32

0.21

0.22

0.91

ln change in reservoir levels % week


t1

23

T. Kristiansen / Electrical Power and Energy Systems 61 (2014) 2026

preceding weeks as percentage of maximum reservoir level. We


have calculated the natural logarithm of all variables. Furthermore,
we have de-meaned the price time series by removing the mean
of log return from the log returns such that the average is zero. We
have performed this operation to facilitate that the process under
consideration is a stationary stochastic process. We will test for
this issue later in this paper but the mean of the Nordic electricity
price is not constant across time but exhibits seasonality.
These variables exhibit the correlation shown in Table 1. There
is a high correlation between the spot price in the current week
and the previous week. There is a weaker negative correlation between spot prices and inow levels. This reects that a higher inow level improves the supply situation and causes prices to
decline. Conversely a lower inow level deteriorates the supply situation and causes price to incline. Likewise there is a weak negative correlation between the spot price and the change in the
Norwegian reservoir levels in the two preceding weeks. A deteriorating reservoir level causes the price to increase. Conversely an
improving reservoir level causes the price to decrease.
To test for multicollinearity we used the Variance Ination Factor (VIF) that measures multicollinearity in the model. VIF measures the magnitude of the change of the estimated coefcients
over the no correlation case among the other variables. If no variables are correlated the VIFs will be 1. If the VIF is around 4 or
greater for one variable, then one independent variable is highly
correlated with the other variables.
We regressed the variables the natural logarithm of the spot
price in week t and week t  1 (ln S t and ln S t  1), the natural logarithm of the inow in week t  1 (ln inf t  1) and the natural logarithm of the change in reservoir level from week t  2 to week
t  1 (ln delta res t  1) against each other and calculated the
VIF = 1/(1R2). The VIF associated with the regression of ln delta
res t  1 and ln S t  1 against ln inf t  1 was 3.55, the VIF associated with the regression of ln inf t  1 and ln S t  1 against ln delta
res t  1 was 3.36 while the VIF associated with the regression of ln
delta res t  1 and ln inf t  1 against ln S t  1 was 1.02. Thus we
conclude there is no substantial muliticolinarity .
The model takes the following functional form:

st c a  st1 b  it1 d  rt1 et


where c is the intercept of the regression equation, st = ln(St) is the
natural logarithm of the weekly spot price in the current week,
st1 = ln(St1) is the natural logarithm of the weekly spot price in
the previous week, it1 = ln(It1) is the natural logarithm of inow
level in the previous week in GWh in Norway and Sweden,
r t1 ln RRt1
is the ratio between the reservoir levels in Norway
t2

Time series

Ln de-meaned spot price 19992010

Dickey Fuller test statistic


p-Value
Lag order

5.4914
0.01
15

in the previous week and the preceding week in percent of maximum reservoir level, a, b, d are regression coefcients and et is assumed to be independent and identically distributed with zero
mean and nite variance.
To estimate the parameters of the suggested model we have
obtained historical data for weekly spot prices, inow levels in
Norway and Sweden and reservoir levels in Norway for the period
from week 44 in 1999 to week 52 in 2010. The results of the regression are shown Tables 2 and 3. The R square equals 0.97 and is
relatively high. In regression analysis, the t-stat, coupled with its
p-value, indicates the statistical signicance of the relationship between the independent and dependent variable. All p-values are
below 0.05 and indicates that the selected variables provide a suitable description of the spot price in the current week.
The characteristics for a stationary time series are a constant
mean and variance. Therefore a time series with trend is nonstationary. Table 4 shows the augmented dickey fuller (ADF) test
for stationary1 conditions for the time series from 1999 to 2010.
The critical values for the ADF test based on F-Statistic are 3.98
for p = 0.01 and 3.68 for p = 0.025 when the number of observations is equal to 500. The ADF test gives p = 0.01 and a Dickey Fuller
test statistic of 5.4914.
We calculated the Hurst exponent to 0,9775 which indicates
that the time series have persistence or long term positive autocorrelation.
In addition we utilize the models proposed by Torro [1] as
benchmarks:
 A myopic method that utilizes the present spot price as a forecast for next week. Torro [1] explains that a myopic method can
be considered as the minimum accuracy required from any
forecasting method
 A futures method that utilizes the closing price of futures for the
last workday in the preceding week as a proxy for next weeks
spot price.
Model performance
To measure the performance of the models we have used the
mean absolute percentage error (MAPE) dened as the average
absolute difference between the actual value and the forecast value
divided by the actual value:

Table 2
Regression statistics for the auto-regressive model.
Regression statistics
Multiple R
R Square
Adjusted R square
Standard error
Observations

0.9746
0.9498
0.9495
0.1056
581

Table 3
Regression coefcients and associated statistics for the auto-regressive model for the
19992010 data period.

Intercept
ln S t  1
ln inf t  1
ln delta res t  1

Table 4
Augmented Dickey Fuller (ADF) test for the period 19992010.

Coefcients

Standard error

t Start

p-Value

0.3285
0.9684
0.0412
0.5858

0.1122
0.0101
0.0142
0.2027

2.9279
96.3215
2.9038
2.8899

0.0035
0.0000
0.0038
0.0040


n 
1X
At  F t 

MAPE
n t1  At 
where A is actual value and F is forecast value.
The performance and accuracy of the price forecasting models
are summarized in Table 5. Likewise we have charted the forecasting models against the actual spot price in Fig. 4.
1
Stationarity, is dened as a quality of a process in which the statistical parameters
(mean and standard deviation) of the process do not change with time. The most
important property of a stationary process is that the auto-correlation function (ACF)
depends on lag alone and does not change with the time at which the function was
calculated. A weakly stationary process has a constant mean and ACF (and therefore
variance). A truly stationary (or strongly stationary) process has all higher-order
moments constant including the variance and mean.

24

T. Kristiansen / Electrical Power and Energy Systems 61 (2014) 2026

Table 5
Weekly MAPE and absolute price deviation (NOK/MW h) for the three proposed
models.
Model

Regression
model
Myopic
model
Futures
model

Deviation forecast
and spot price
(NOK/MW h)

Average absolute deviation


forecast and spot price
(NOK/MW h)

7.5%

1.4

19.6

7.5%

0.9

19.9

5.3%

4.3

14.2

MAPE

Table 6
Regression coefcients and associated statistics for the 19992006 data period which
is used for the out sample model testing.

Intercept
ln S t  1
ln inf t  1
ln delta res t  1

Coefcients

Standard error

t Stat

p-Value

0.3888
0.9589
0.0497
0.6289

0.1380
0.0128
0.0175
0.2483

2.8179
74.8375
2.8340
2.5332

0.0051
0.0000
0.0048
0.0117

Table 7
Weekly MAPE and absolute price deviation (NOK/MW h) for the out of sample period.

The regression model achieves a mean absolute percentage error of 7.5% and under-forecasts the spot price by 1.4 NOK/MW h.
This result is comparable with what Kristiansen [17] achieved with
a weekly mean absolute percentage error (WMAE) of around 67%.
However the price deviation was somewhat higher and in the
range 1.72.6 NOK/MW h.
The myopic model achieves a MAPE 7.5% and over-forecasts the
spot price by some 0.9 NOK/MW h. The futures model achieves a
MAPE 5.3% and over-forecasts the spot price by some 4.3 NOK/
MW h.

Model

MAPE

Deviation forecast and spot price (NOK/MW h)

7.4%

0.0

The futures model achieves the best performance in terms of


MAPE equal to 5.3% while the regression and myopic models have
the highest MAPEs of 7.5%. The deviation is lowest for the myopic
and regression models. The absolute deviation is lowest (14.2 NOK/
MW h) for the futures model and highest for the myopic model
(19.9 NOK/MW h).

Fig. 4. Comparison of price forecasting models and actual spot price.

Fig. 5. Autocorrelation function for the residual (condence interval in red). (For interpretation of the references to colours in this gure legend, the reader is referred to the
web version of this paper.)

T. Kristiansen / Electrical Power and Energy Systems 61 (2014) 2026

25

Fig. 6. Partial autocorrelation function for the residual (condence interval in red). (For interpretation of the references to colours in this gure legend, the reader is referred
to the web version of this paper.)

Fig. 7. Autocorrelation function for the squared residual (condence interval in red). (For interpretation of the references to colours in this gure legend, the reader is referred
to the web version of this paper.)

Fig. 8. Partial autocorrelation function for the squared residual (condence interval in red). (For interpretation of the references to colours in this gure legend, the reader is
referred to the web version of this paper.)

Thus we can conclude that a pure myopic model which utilizes


the previous weeks spot price as a forecast for the current week
has a similar performance as the regression model. However the
futures model has the best overall performance in terms of MAPE
and absolute price deviation but it over-forecasts the spot price.
It is worth noting that the futures model has some large deviations
for the end of 2002 which was a period with drought. Also for the

spot and the regression models there were some larger deviations
for the end of 2002.
To further test the regression model we performed out of sample test by estimating the regression coefcients in Table 6 from
the period week 44 in 1999 to week 52 in 2006. The R square
was 0.97. We subsequently applied the regression parameters to
data for the period from 2007 to 2010.

26

T. Kristiansen / Electrical Power and Energy Systems 61 (2014) 2026

The performance of the out of sample test is shown in Table 7.


The performance of the MAPE is similar to the previous test. However the deviation between the forecast and the spot price is somewhat lower.
For a given stochastic process one is often interested in the connection between two random variables of a process at different
points in time. One way to measure a linear relationship is with
the auto-correlation function (ACF), which measures the correlation between these two variables. We have applied the ACF with
various time lags on the residuals and the squared residuals in
Figs. 58. We note the alternating and tapering patterns for the
residual and the squared residuals. For the residuals both charts
exhibit a similar pattern. The ACF and PACF functions show a definite pattern with decreasing lags and thus a trend in the data. The
magnitude of the ACF and PACFs are relatively small but the lags
are larger at orders 1, 2, 4, 5, 10 and 16 for the ACF and at orders
1, 2, 3, 5 and 16 for the PACF. Also the squared residuals ACF and
PACF functions exhibit a similar pattern with decreasing lags.
The ACF tails off around 6 lags while the PACF tails off after 2 lags
but also have one spike at lag 5.
Conclusion
This paper has presented three relatively simple spot price
forecasting models for the Nord Pool market. The regression model
requires historic spot prices, inow and reservoir levels for the
estimation of the regression coefcients. The model achieves a
R-square of around 0.97. The mean absolute percentage error is 7.5%
and it under-forecasts the actual spot price with 1.4 NOK/MW h.
Out of sample achieves a mean absolute percentage error of 7.4%
and matches the actual spot price.
The myopic model based on historic spot prices by extrapolating the current weeks spot price as a predictor for next weeks
price. Its MAPE is 7.5% and the price deviation is 0.9 NOK/
MW h. The futures price model which takes the last futures closing
price for the week as a predictor for next weeks spot price, has a
MAPE of 5.3% and deviation of 4.3 NOK/MW h.
We conclude that the futures model outperforms the other
models with a lower MAPE and average absolute price deviation.
The regression and myopic models have a similar performance.
The last futures closing price for the week ahead contract should
discount all available information to the market. Thus it could be
considered as a price forecast for the following week. The myopic

model is a nave model by assuming an extrapolation of the current


trend. The regression model has a similar feature but also utilizes
fundamental information such as reservoir and inow levels.
References
[1] Torro H. Electricity futures prices: some evidence on forecast power at Nord
Pool. J Energy Markets 2009;2(3):326.
[2] Yan X, Chowdhury NA. Mid-term electricity market clearing price forecasting:
a multiple SVM approach. Int J Electr Power Energy Syst 2014;58:20614.
[3] Shrivastava NA, Panigrahi BK. A hybrid wavelet-ELM based short term price
forecasting for electricity markets. Int J Electr Power Energy Syst
2014;55:4150.
[4] Chabane N. A hybrid ARFIMA and neural network model for electricity price
prediction. Int J Electr Power Energy Syst 2014;55:18794.
[5] Lei M, Feng Z. A proposed grey model for short-term electricity price
forecasting in competitive power markets. Int J Electr Power Energy Syst
2012;43(1):5318.
[6] Thomson Reuters Point Carbon. <http://www.pointcarbon.com/>.
[7] Weron R. Modeling and forecasting electricity loads and prices: a statistical
approach. Chichester: Wiley; 2006.
[8] Contreras J, Espnola R, Nogales FJ, Conejo AJ. ARIMA models to predict nextday electricity prices. IEEE Trans Power Syst 2003;18(3):101420.
[9] Zhou M, Yan Z, Ni Y, Li G, Nie. Electricity price forecasting with condenceinterval estimation through an extended ARIMA approach. IEE Proc Gener
Trans Distrib 2006;153(2):2338.
[10] Garcia RC, Contreras J, van Akkeren M, Garcia JBC. A GARCH forecasting model
to predict day-ahead electricity prices. IEEE Trans Power Syst
2005;20(2):86774.
[11] Weron R, Misiorek A. Forecasting spot electricity prices: a comparison of
parametric and semiparametric time series models. Int J Forecast
2008;24:74463.
[12] Conejo AJ, Contreras J, Espnola R, Plazas MA. Forecasting electricity prices for a
day-ahead pool-based electric energy market. Int J Forecast 2005;21(3):
43562.
[13] Panagiotelis A, Smith M. Bayesian forecasting of intraday electricity prices
using multivariate skew-elliptical distributions. Int J Forecast 2008;24(4):
71027.
[14] Misiorek A, Truck S, Weron R. Point and interval forecasting of spot electricity
prices: linear vs. non-linear time series models. Nonlinear Dyn Econom
2006;10(3):2.
[15] Karakatsani N, Bunn D. Forecasting electricity prices: the impact of
fundamentals and time-varying coefcients. Int J Forecast 2008;24(4):76485.
[16] Knittel CR, Roberts MR. An empirical examination of restructured electricity
prices. Energy Econ 2005;27(5):791817.
[17] Kristiansen T. Forecasting Nord Pool day-ahead prices with an auto-regressive
model. Energy Policy 2012;49:32832.
[18] Kristiansen T. Power trading analytics and forecasting Germany. Electr J
2011;24(8):4155.
[19] Nord Pool. 2012. <http://www.nordpool.no.>.
[20] Gjolberg O, Johnsen T. Electricity futures: inventories and price relationships
at Nord Pool. Discussion Paper D-16/2001, IS. 2001.
[21] Botterud A, Kristiansen T, Ilic M. The relationship between spot and futures
prices in the Nord Pool electricity market. Energy Econ 2010;32(5):96778.

You might also like