Paper Joint

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/338290296
An Econometric Model for Intraday Electricity Trading
Article in Philosophical Transactions of The Royal Society A Mathematical Physical and Engineering Sciences · July 2021
DOI: 10.2139/ssrn.3489214
CITATIONS READS
11 920
3 authors:
Marcel Kremer Ruediger Kiesel

University of Duisburg-Essen University of Duisburg-Essen
13 PUBLICATIONS 72 CITATIONS 73 PUBLICATIONS 1,077 CITATIONS
SEE PROFILE SEE PROFILE
Florentina Paraschiv
Zeppelin University
43 PUBLICATIONS 796 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Understanding currency dynamics with machine learning and information theory View project
Modeling empirical dependence structures on stock markets with copulas View project
All content following this page was uploaded by Marcel Kremer on 17 December 2020.
The user has requested enhancement of the downloaded file.

An Econometric Model for Intraday Electricity
Trading
Marcel Kremer∗,a , Rüdiger Kiesela,b , and Florentina Paraschivc,d
a
Chair for Energy Trading and Finance, University of Duisburg-Essen, Universitätsstraße 12, 45141
Essen, Germany
b
Department of Mathematics, University of Oslo, PO Box 1053 Blindern, 0316 Oslo, Norway
c
NTNU Business School, Norwegian University of Science and Technology, 7491 Trondheim, Norway
d
Institute for Operations Research and Computational Finance, University of St. Gallen, Bodanstrasse
6, CH-9000 St. Gallen, Switzerland
September 3, 2020
Revised version forthcoming in Philosophical Transactions of the Royal Society A.
Abstract
This paper develops an econometric price model with fundamental impacts for in-
traday electricity markets of 15-minute contracts. A unique data set of intradaily
updated forecasts of renewable power generation is analyzed. We use a threshold
regression model to examine how 15-minute intraday trading depends on the slope of
the merit order curve. Our estimation results reveal strong evidence of mean rever-
sion in the price formation mechanism of 15-minute contracts. Additionally, prices
of neighboring contracts exhibit strong explanatory power and a positive impact on
prices of a given contract. We observe an asymmetric effect of renewable forecast
changes on intraday prices depending on the merit-order-curve slope. In general,
renewable forecasts have a higher explanatory power at noon than in the morning
and evening, but price information is the main driver of 15-minute intraday trading.
Keywords: Intraday electricity market; Econometric modeling; 15-minute con-

tracts; Renewable power forecasts; Merit order curve; Threshold regression
JEL Classification: C22; C24; C55; G10; Q20; Q21; Q40; Q41; Q42
1 Introduction
In recent years, the expansion of renewable energy sources has been forged ahead massively
across the globe with a direct impact on electricity markets. As electricity generation from
renewable energy sources cannot be predicted reliably in the long term, that is, days, weeks
or months ahead, the future of electricity trading is foreseen in short-term electricity markets.
∗
Corresponding author. Email: marcel.kremer@uni-due.de
1
Energy supply companies thus face the challenge of moving towards automatic trading, which
requires the identification of trading strategies based on local demand and supply patterns as
well as cross-border energy flows. The German market constitutes a pioneer among European
electricity markets. Not only is the German market the largest electricity market in Europe in
terms of total trading volume (European Energy Exchange AG, 2016a), but also novel develop-
ments and innovations (new contracts, reduction of lead time) are traditionally introduced on
the German market first. In Germany the most short-term electricity market is the continuous
intraday market, where hourly and 15-minute contracts can be traded until 5 minutes before
the delivery of electricity begins. Intraday electricity markets are designed according to the
needs of energy supply companies to balance forecast errors of renewable power generation.
This paper investigates four research questions: (i) Which fundamental factors drive the
intraday trading of 15-minute contracts? (ii) How do forecast changes of renewable power
generation affect the intraday trading of 15-minute contracts? (iii) Can we identify different
regimes on the intraday market where the price formation process behaves differently? (iv) How
does intraday trading depend on the time of day?
Like on any financial market, there is a natural desire for pricing models of the basic securities.
To build as realistic models as possible, we require ex-ante information on the underlying price
drivers. More market-specifically, and in light of the design of intraday electricity markets, fore-
cast errors of renewable power production have to be taken into account. Kiesel and Paraschiv
(2017) deliver the first and only work fulfilling these prerequisites. Our article builds upon
their work and develops an econometric price model with fundamental impacts for continuous
intraday markets of 15-minute contracts.
Our main modeling assumption is that the price formation process on the intraday electricity
market depends on the slope of the merit order curve. Figure 1 illustrates a merit order curve
with and without the infeed from renewable energy sources (RES). The merit order curve is a
non-linear and convex function of the marginal costs of power plants depending on the generation
capacity. Equivalently, since the marginal costs of the last running power plant needed to cover
demand set the final electricity price, the merit order curve may be interpreted as a function of
the electricity price depending on the electricity demand, too.
Let us consider the merit order curve without renewable power infeed (blue solid). When
there is low demand, only relatively cheap power generation technologies such as lignite-fired
power plants are needed, whereas if demand is high, more expensive technologies such as gas-
fired power plants might be required to satisfy demand. As exhibited in Figure 1, if demand is
low, the slope of the merit order curve is relatively small and we call the market to be in a flat
merit-order regime; however, if demand is high, the merit-order-curve slope is comparatively
large and the market is said to be in a steep merit-order regime.
Let us turn towards the merit order curve with renewable power infeed (green dash-dotted).
Since renewable energy sources have zero marginal costs, their infeed shifts the entire merit
order curve to the right. As a consequence, if demand is low (flat regime), the electricity price
decreases by a small amount ∆Pf ; however, if demand is high (steep regime), the electricity
price decreases as well but by a much larger amount ∆Ps > ∆Pf . We take into account this
asymmetric effect of renewable power infeed on electricity prices in our econometric model by
incorporating the slope of the merit order curve.
To model the merit order curve, two approaches have been proposed in the academic liter-
ature. The first approach models the merit order curve from the supply side via generation
capacities and marginal costs of power plants. While generation capacities are given by market
transparency data, one has to specify a model for the marginal costs of power plants. Pape
et al. (2016); Kallabis et al. (2016); Beran et al. (2019) model the marginal costs as a function
of fuel prices, CO2 emission allowances prices, emission intensities of fuel types, power plant
heat rates or thermal efficiencies, and other variable costs. As marginal costs depend on a wide
variety of factors, this approach is highly complex and subject to strong modeling assumptions.
2
flat steep
Electricity price
w/o RES
with RES
∆Ps
∆Pf
RES Demand
Figure 1: Merit order curve without (blue solid) and with (green dash-dotted) infeed from re-
newable energy sources (RES) indicating a flat and steep merit-order regime (red)
with electricity price changes ∆Pf and ∆Ps , respectively.
The second approach models the merit order curve from the demand side via electricity loads,
or load forecasts, and electricity prices as proposed by Burger et al. (2004); He et al. (2013).
While electricity load is an indicator of total electricity demand, electricity prices are deter-
mined by the marginal costs of the price-setting power plant (marginal power plant). As data
for both electricity loads and prices exist, this approach does not rely on modeling assumptions.
Therefore, we follow the demand-side approach.
The intraday electricity market of 15-minute contracts has only recently come into focus of
scientific research. To the best of our knowledge, only three studies exist hitherto. Kiesel
and Paraschiv (2017) provide the first econometric model for 15-minute intraday prices. They
model both deviations between day-ahead and last intraday prices as well as continuous intraday
prices in a threshold regression context. They provide evidence that 15-minute prices respond
asymmetrically to intradaily updated renewable forecast errors depending on the proportion
of expected demand covered by conventional energy sources. Märkle-Huß et al. (2018) study
how the introduction of 15-minute contracts has affected the day-ahead and intraday market
of hourly contracts. They find that prices of hourly contracts decreased and trading volumes
increased. Kath and Ziel (2018) present the first forecasting study of 15-minute prices using an
elastic net regression model. They conclude that prices in the intraday auction are much easier
to forecast than prices in the continuous trading.
This article extends the existing literature along a number of dimensions: First, we explore
a novel and unique data set of high-frequency transaction data and linked fundamental supply
and demand data. Intradaily updated forecasts of renewable power generation (solar, wind)
constitute the heart of our data collection. These are the same real-time renewable forecasts as
available to traders on the intraday market. As such, this is the most extensive data set used
in the empirical literature to study the price formation process on intraday electricity markets.
Hence, our data set allows for a more realistic model specification than proposed in previous
research.
Second, we suggest the first econometric price model for 15-minute contracts that incorporates
the slope of the merit order curve. This is a substantial improvement over Kiesel and Paraschiv
(2017) as electricity prices react asymmetrically to renewable forecast changes depending on the
merit-order-curve slope: If the merit order is steep, electricity prices change more severely in
3
the wake of renewable forecast errors than if the merit order is flat. Moreover, our econometric
model solely involves ex-ante market knowledge.
Third, and to the best of our knowledge, this is the first work studying the influence of
neighboring 15-minute contracts on the price dynamics of a given contract. This is motivated
by the fact that adjacent contracts are driven by similar market information.
This paper is organized as follows: In Section 2, we lay out our data set and perform an
empirical analysis of intraday transaction data of 15-minute contracts. In Section 3, we present
the existing econometric model by Kiesel and Paraschiv (2017) and our extended version of the
model as well as the threshold regression. In Section 4, we calibrate the econometric models to
market data and discuss our estimation results. We offer our conclusions in Section 5.
2 Stylized facts
In this section, we lay out our data set and perform an empirical analysis of the hourly season-
ality and liquidity evolution of 15-minute contracts.
2.1 Data
We investigate high-frequency trade and linked fundamental data of all 96 15-minute contracts
traded on the German continuous intraday power market at EPEX SPOT SE. Our observation
period spans from January 1 to December 31, 2015. The trade data of 15-minute contracts
involve transaction prices and trading volumes from the continuous intraday trading session with
a 1-minute time resolution and are provided by European Energy Exchange AG (2016c). This,
however, does not imply that we observe one transaction every minute, but it may take a multiple
of one minute until the subsequent transaction is observed. In case of multiple transactions
within the same trading minute, we compute the volume-weighted average price for that minute.
The continuous trading session for 15-minute contracts opens daily at 4 PM and, in 2015, ends
45 and 30 minutes before delivery begins, respectively1,2 . Furthermore, we use market clearing
prices of 15-minute contracts traded in the German 15-minute intraday auction at EPEX SPOT
SE, which takes place daily at 3 PM, provided by European Energy Exchange AG (2016b). The
auction data may also be obtained via the R package emarketcrawlR by Wagner (2018). We
shall denote a 15-minute contract by HhQq with h = 0, . . . , 23 and q = 1, . . . , 4; e.g., contract
H13Q1 refers to the delivery period 1:00–1:15 PM.
As fundamental data, we include intraday wind and solar power forecasts, expected demand,
and expected conventional capacity. The intraday renewable power forecasts involve intradaily
updated forecasts of wind and solar power production in Germany, which are the same real-
time renewable forecasts as available to traders on the intraday market. These forecasts are
updated every 15 minutes, where each update contains a forecast time series for the following
eight days in a 15-minute time resolution, provided by EWE TRADING GmbH (2016). As
such, the intraday renewable power forecasts constitute a unique data set and, to the best of
our knowledge, have solely been analyzed by Kiesel and Paraschiv (2017). As an indicator of
expected electricity demand, we use the day-ahead total load forecast for each quarter-hour on
the following day in Germany, which is published daily at 10 AM and is provided by European
Network of Transmission System Operators for Electricity Transparency Platform (2016). The
expected conventional capacity covers the expected daily average of available generation capacity
of conventional power plants on the following day in Germany, which is published daily at 10
AM and is provided by European Energy Exchange AG Transparency Platform (2016). As
1
EPEX SPOT SE reduced the lead time on the German continuous intraday power market for hourly and
15-minute contracts from 45 to 30 minutes before delivery on July 16, 2015 (EPEX SPOT SE, 2015).
2
On June 14, 2017, the lead time within the four German control zones was locally further reduced to 5 minutes
before delivery (EPEX SPOT SE, 2017).
4
conventional energy sources, we include coal, garbage, gas, lignite, oil, other, pumped-storage,
run-of-the-river, seasonal-store, uranium. A summary of the employed data may be found in
Table A.1 in Appendix A.
2.2 Hourly seasonality

2.2.1 Transaction prices
Figure 2 illustrates the volume-weighted average transaction price of 15-minute contracts during
peak hours and off-peak hours for summer and winter. We identify an hourly seasonality of
volume-weighted average prices for both peak and off-peak hours as well as for summer and
winter3 . The hourly seasonality exhibits a sawtooth-like shape: For the peak-hour contracts
H8Q1–H13Q4, the average price of the first 15-minute contract within each hour is the highest
and it declines until the last 15-minute contract within that hour, which has the lowest average
price. Conversely, for contracts H14Q1–H18Q4, the lowest average price is present for the
first 15-minute contract which increases up to the highest average price for the last 15-minute
contract in a given hour.
The hourly seasonality pattern and its change around noon may be explained by electricity
generation from solar energy: In the first half of the day, the sun rises and less electricity from
solar energy is produced during the first quarter-hour as compared to the last quarter-hour
within each hour. If a (renewable) electricity supplier sold an hourly contract on the day-ahead
market, it has to buy electricity for the first quarter-hour on the intraday market to meet its
obligation since less electricity is produced from solar energy than it has sold (buy pressure);
thus prices increase. In the last quarter-hour, however, more solar electricity is generated than
it has sold on the day-ahead market and so it wants to sell the surplus on the intraday market
to avoid entering the balancing energy market (sell pressure); hence prices decrease. In the
second half of the day, after the sun has reached its highest level (around 2 PM in Germany),
more solar power is generated during the first than in the last quarter-hour in each hour. Thus,
there is a sell pressure in the first and a buy pressure in the last quarter-hour of an hour, and
the pattern is reversed. The existence of buy and sell pressure is underpinned by the hourly
seasonality of trading volumes described in Section 2.2.2.
Similarly, the sawtooth-shaped hourly seasonality of volume-weighted average prices is found
during off-peak hours. For contracts H20Q1–H1Q4, the average price of the first and last
quarter-hourly contract within an hour is highest and lowest, respectively, while for contracts
H4Q1–H6Q4, this is reversed. The hourly seasonality at night stems from the design of conven-
tional power plants which ramp up and down.
Overall, during peak hours, average transaction prices are lower in summer than in winter
apart from a few exceptions. In the afternoon and evening hours, that is, for contracts H14Q1–
H19Q4, we find larger deviations between summer and winter average prices than in morning
and noon hours. During off-peak hours, average prices are fairly similar during both seasons
most of the time and only slightly lower in winter than in summer.
2.2.2 Trading volumes

Figure 3 shows the total trading volume of 15-minute contracts during peak hours and off-peak
hours averaged over the year. We only present the yearly average of total trading volumes as
the distinction between summer and winter does not provide additional information. Similar
to transaction prices, we observe an hourly seasonality of total trading volumes for both peak
and off-peak hours. The hourly seasonality of trading volumes has a U-shape: Larger total
trading volumes are found for the first and last 15-minute contract within each hour of the
3
The hourly seasonality preserves for unweighted average transaction prices both qualitatively and quantita-
tively: The seasonal averages of unweighted and volume-weighted average prices differ by roughly 2% only.
5
transaction prices [EUR/MWh]
transaction prices [EUR/MWh]
volume-weighted average
volume-weighted average summer
50 50 winter
Seasonal average of
Seasonal average of
40 40
30 30
20 summer 20
winter
H8Q1 H12Q1 H16Q1 H20Q1 H20Q1 H0Q1 H4Q1 H8Q1
Quarter-hourly contract Quarter-hourly contract
Figure 2: Volume-weighted average transaction price of 15-minute contracts during peak hours
(left) and off-peak hours (right) averaged over summer (red) and winter (blue).
day, while the second and third 15-minute contract in an hour always exhibit lower trading
volumes. More specifically, the last 15-minute contract entails the largest trading volume, while
the second 15-minute contract involves the lowest trading volume in an hour. The U-shaped
hourly seasonality supports our hypothesis of buy and sell pressure for the marketing of solar
power in the first and last quarter-hour during peak hours, respectively.
15-minute contracts during off-peak hours are generally associated with less trading volume
than peak-hour contracts. Total trading volumes are particularly low for contracts H0Q1–
H5Q4. However, the U-shaped hourly seasonality of trading volumes is persistent during off-
peak hours. The hourly seasonality at night results from balancing out the ramp-up and -down
of conventional power plants.
1000 1000
total trading volumes [MW]
total trading volumes [MW]
800 800
Yearly average of
Yearly average of
600 600
400 400
200 200
H8Q1 H12Q1 H16Q1 H20Q1 H20Q1 H0Q1 H4Q1 H8Q1
Quarter-hourly contract Quarter-hourly contract
Figure 3: Total trading volume of 15-minute contracts during peak hours (left) and off-peak
hours (right) averaged over the year.
2.3 Liquidity evolution

2.3.1 Gate closure
Figure 4 displays the temporal evolution of liquidity of 15-minute contracts over the trading
session towards gate closure. As measures for market liquidity, we use the number of trades and
total trading volume aggregated over all 15-minute contracts and all trading sessions. Due to
low liquidity far from gate closure, we focus on the last three trading hours prior to gate closure.
We note that EPEX SPOT SE reduced the lead time on the German intraday power market for
hourly and 15-minute contracts from 45 to 30 minutes before delivery on July 16, 2015 (EPEX
SPOT SE, 2015). Thus, to avoid effects due to the shift of lead time, we synchronize our trade
6
time series with respect to gate closure by shifting the trade time stamps of 15-minute contracts
maturing before July 16, 2015 by 15 minutes.
The number of trades increases from 952 three hours to gate closure to 3,054 one hour to
gate closure. Subsequently, the number of trades rises further and more than doubles to 7,080
half an hour to gate closure. 15 minutes to gate closure, the number of trades jumps to 16,593,
while 13 minutes before gate closure it reaches a local maximum of 28,378. The surge of trading
activity around 15 minutes to gate closure may be associated with the fact that a given contract
becomes the front 15-minute contract. The maximum value of 30,512 trades is observed one
minute to gate closure.
The total trading volume increases from 12 GW to 48 GW between three and one hour to
gate closure. Thereafter, trading volume almost triples to 135 GW half an hour to gate closure.
15 and 13 minutes before gate closure, total trading volume raises to 261 GW and 314 GW,
respectively, which coincides with becoming the front 15-minute contract. A comparison with
the number of trades at these points in time in Figure 4 reveals that indeed a vast number of
transactions is executed but with comparatively low trading volumes. The total trading volume
peaks at 960 GW one minute to gate closure.
Thus, we observe that liquidity of 15-minute contracts rises severely within the last trading
hour prior to gate closure: On average, roughly 68% of the number of transactions are executed
and roughly 74% of the total trading volume is transferred. The reason why the majority of
trading takes place close to gate closure is that forecasts of fundamentals, particularly renewable
power forecasts, become more and more precise regarding the delivery period of a given 15-
minute contract. Hence, it is desirable to trade as closest to delivery begin as possible. Glas
et al. (2020); Graf von Luckner and Kiesel (2020) document an increasing liquidity towards gate
closure for hourly contracts on the German intraday electricity market, too.
We identify a small but distinct rise in liquidity every 15 minutes. This is particularly pro-
nounced for the number of trades but present for the total trading volume, too. Consequently,
an increased amount of transactions with relatively little trading volume is conducted peri-
odically. We argue that the 15-minute periodicity in liquidity originates from newly arriving
renewable forecast updates. As described in Section 2.1, renewable forecasts are updated in 15-
minute intervals. New forecasts will not have changed much after 15 minutes and thus traders
only make minor adjustments to their positions by trading little volume. Moreover, we observe
that trading activity increases at isolated points in time and dies out immediately until the
next increase. Therefore, we conclude that renewable forecast updates are reflected in prices of
15-minute contracts within one trading minute.
1e4
3 1.0 1e6
Total trading volume [MW]
0.8
Number of trades
2 0.6
0.4
1
0.2
0 0.0
3 2 1 0 3 2 1 0
Hour to gate closure Hour to gate closure
Figure 4: Time evolution of the number of trades (left) and total trading volume (right) through
the trading session towards gate closure.
7
2.3.2 Gate opening
Figure 5 illustrates the temporal development of liquidity of 15-minute contracts after gate
opening at 4 PM. The number of trades amounts to 4,718 just after gate opening and decreases
to 705 12 minutes after gate opening. 15 minutes after gate opening, the number of trades
reaches its maximum value of 5,309, which drains within the next trading minute. Similarly,
liquidity jumps to 2,276 trades 30 minutes after gate opening and falls back to its prior level
within two trading minutes. Thereafter, liquidity approaches a fairly stable level of 185 trades,
on average, between one and two hours after gate opening.
The total trading volume peaks at 58 GW shortly after gate opening and steadily declines
to 5.6 GW 12 minutes after gate opening. 15 minutes after gate opening, total trading volume
rises to 13 GW while it amounts to 3.5 GW 30 minutes after gate opening. Comparing trading
volume and the number of trades at these points in time shows that the surge in trading activity
involves relatively little trading volume. Subsequently, the total trading volume keeps a quite
constant and low level around 1.5 GW until two hours after gate opening.
Thus, liquidity of 15-minute contracts is high close to gate opening and drops substantially
during the first 12 trading minutes. This behavior may be explained by the fact that market
participants initialize their positions. In the sequel, little volume is traded and thereby only
minor adjustments are made to the open positions. Generally, liquidity remains poor until
the last trading hour prior to gate closure as the forecasts of fundamentals are still relatively
inaccurate.
1e3 6 1e4
Total trading volume [MW]
5
Number of trades
4 4
3
2 2
1
0 0
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
Hour from gate opening Hour from gate opening
Figure 5: Time evolution of the number of trades (left) and total trading volume (right) after
gate opening.
3 Methodology
We aim at modeling the asymmetric response of 15-minute intraday electricity prices to explana-
tory variables, in particular, to renewable forecast changes. As a starting point, we reimplement
the econometric model by Kiesel and Paraschiv (2017). Moreover, we suggest an extension of
their model to overcome its weaknesses. We employ a threshold regression model to calibrate
the econometric models to market data.
3.1 Benchmark econometric model

We use the econometric model for intraday price changes of 15-minute contracts proposed
by Kiesel and Paraschiv (2017) as a benchmark. For each 15-minute contract, the model
8
specification reads
3
ατ ∆Pt−τ + α4 Vt + α5 DQ + α6 ∆wtn + α7 ∆wtp
X
∆Pt = α0 +
τ =1
√
+ α8 ∆snt + α9 ∆spt + α10 ∆t + εt , (1)
where ∆Pt = Pt − Pt−1 denotes the transaction price change between times t and t − 1,
Vt the trading volume at time t, DQ = c` the demand quota with expected demand ` and
expected conventional capacity c for a given 15-minute contract, ∆wtn = min(∆wt , 0) and
∆wtp = max(∆wt , 0) negative and positive wind power forecast changes, respectively, where
∆wt = wt − wt−1 is the last available wind power forecast change at time t, ∆snt = min(∆st , 0)
and ∆spt = max(∆st , 0) negative and positive solar power forecast changes, respectively, where
∆st = st − st−1 is the last available solar power forecast change at time t, ∆t the interarrival
iid
time between two consecutive transactions conducted at times t and t − 1, and εt ∼ N (0, σ 2 )
the error term.
The sum in Equation (1) covers three lagged price changes ∆Pt−τ , τ = 1, 2, 3, that is, au-
toregressive terms, where the number of lags has been determined by partial autocorrelation of
price changes. Moreover, we distinguish between positive and negative wind and solar power
forecast errors ∆wtp , ∆spt and ∆wtn , ∆snt , respectively, as we expect them to have opposite effects
on electricity price changes ∆Pt : Positive renewable forecast errors should decrease electricity
prices, whereas negative renewable forecast errors should increase electricity prices. We control
for the interarrival time ∆t since transactions do not take place at equidistant points in time,
but it may take one minute or several hours until the next trade is conducted.
The demand quota DQ quantifies the proportion of expected demand l which is expected
to be met by conventional power generation capacities c; or, put another way, how much the
expected conventional capacity c does cover expected demand l. We use the demand quota DQ
as threshold variable since we assume that the gap between expected demand l and expected
conventional capacity c influences the trading behavior of market participants on the intraday
market: A large gap induces a great necessity to balance electricity produced from renewable
energy sources, whereas a small gap puts less pressure to adjust renewable electricity (see Kiesel
and Paraschiv, 2017, chap. 2, for a more detailed discussion).
The demand quota DQ is the only variable remaining constant during the continuous trading
session of a 15-minute contract. However, it does depend on the specific contract since the
expected demand l is provided in quarter-hourly granularity. In particular, the value of DQ is
known daily at 10 AM and thus before continuous trading begins. From a trader’s perspective,
the demand quota is well suited as threshold variable, too, since it indicates in which regime the
market is before continuous trading begins. Thereby, the corresponding intraday price model
can be chosen ex-ante.
3.2 Extended econometric model

We refine the econometric model by Kiesel and Paraschiv (2017) along three dimensions by
incorporating supplementarily: (i) the slope of the merit order curve, (ii) price changes of
neighboring 15-minute contracts, (iii) the 15-minute intraday auction price. For a given 15-
minute contract i = 1, . . . , 96, the model specification reads
m n
(i) (i) X (i) X (i) (i+j) (i) (i) (i) (i)
∆Pt = α0 + ατ(i) ∆Pt−τ + βj ∆Pt + η1 ξ (i) + η2 P Auc,(i) + η3 Vt
τ =1 j=−n,
j6=0
p
(i) n,(i) (i) p,(i) (i) n,(i) (i) p,(i) (i) (i)
+ η4 ∆wt + η5 ∆wt + η6 ∆st + η7 ∆st + η8 ∆t(i) + εt , (2)
9
(i+j)
where ∆Pt denotes the last observed price change at time t of neighboring contract i+j, ξ (i)
the slope of the merit order curve, and P Auc,(i) the 15-minute intraday auction price of contract
i. The remaining variables correspond to the benchmark model (1) described in Section 3.1.
(i)
The first sum in Equation (2) covers m autoregressive price changes ∆Pt−τ , τ = 1, . . . , m. To
determine the number of lags, we use the partial autocorrelation of price changes and choose
(i+j)
m = 3. The second sum in Equation (2) captures the price change ∆Pt , j = −n, . . . , n, j 6=
0, at time t of n 15-minute contracts maturing before and n 15-minute contracts maturing after
contract i. For example, if n = 2 and i = H13Q1, the price change at time t of contracts H12Q3,
H12Q4, H13Q2, H13Q3 is included. The intraday auction price P Auc,(i) remains constant during
the trading session of a given contract i and can be considered as an estimate of the initial price
of contract i in the continuous trading.
We use the slope of the merit order curve ξ (i) as threshold variable instead of the demand
quota DQ compared to Kiesel and Paraschiv (2017). One weakness of the demand quota DQ
is that it does not recognize whether the market is in a flat or steep merit-order-curve regime:
Proportionally speaking, a high expected demand l and high expected conventional capacity
c lead to the same value of DQ as low demand and low capacity. Another weakness is that
the demand quota aggregates expected capacities over all conventional generation technologies.
Thus, it loses information on the price-setting power plant and the slope of the merit order
curve. We overcome these limitations in our extended econometric model (2).
We estimate ξ (i) from empirical intraday auction prices P Auc,(i) and expected demands, or
total load forecasts, `(i) in the spirit of Burger et al. (2004); He et al. (2013). This approach is
reasonable for describing the slope of the merit order curve since the level of intraday auction
prices reflects the marginal costs of power plants needed to cover expected demand. Burger
et al. (2004); He et al. (2013) construct a merit order curve for the electricity spot market as
a whole, independent of the contract, from hourly day-ahead prices and hourly load forecasts.
We, however, determine a merit order curve for each 15-minute contract i individually and thus
arrive at a more fine-grained picture. We fit a function f (`) to the price–load data and call
f (`) the empirical merit order curve4 . Then we take the derivative of the empirical merit order
curve f 0 (`) = dfd`(`) and substitute empirical

expected demands `(i) into f 0 (`) to obtain empirical
merit-order-curve slopes ξ (i) = f 0 `(i) .
The slope of the merit order curve ξ (i) remains constant during the trading session of contract
i. The value of ξ (i) can be determined daily around 3:10 PM, after the publication of the intraday
auction results, and hence before continuous trading begins. Thus, ξ (i) indicates whether the
market is in a flat or steep merit-order regime ex-ante and the appropriate intraday price model
can be chosen accordingly. This is an attractive feature of our econometric model for practical
applications.
3.3 Threshold regression

We use the threshold regression model introduced by Hansen (2000) to calibrate our economet-
ric models to market data. The threshold regression model is able to reveal asymmetries in
the impact of explanatory variables with respect to a specified threshold variable. The basic
concept of the threshold regression involves two steps: First, the entire sample is split into two
subsamples, also referred to as groups, classes, or regimes, at a certain threshold value of a
designated threshold variable; second, a linear regression model is estimated on each subsample
separately.
More formally, suppose we observe the sample {yi , xi , qi }ni=1 , where yi is the dependent vari-
able, xi ∈ Rm collects the independent variables, and qi denotes the threshold variable. The
threshold variable qi may be one of the independent variables gathered in xi and must have a
4
Burger et al. (2004) call f empirical price load curve.
10
continuous distribution. The threshold regression model reads
(
θ10 xi + εi , if qi ≤ γ
yi = ,
θ20 xi + εi , if qi > γ
= θ10 xi 1{qi ≤ γ} + θ20 xi 1{qi > γ} + εi , i = 1, . . . , n, (3)
where γ is the threshold parameter, θ1 , θ2 ∈ Rm collect the regression parameters, 1 denotes the
iid
indicator function, and εi ∼ N (0, σ 2 ) is the error term. Thus, the observed sample is split into
two subsamples along the threshold variable qi at a specific value γ. By design, the threshold
regression model (3) allows the regression parameters in θ1 , θ2 to vary between the regimes.
In order to estimate the threshold regression model, we reformulate Equation (3): Let xi (γ) =
xi 1{qi ≤ γ} and call δn = θ1 − θ2 the threshold effect. Then the threshold model (3) can be
written in the alternative form,
yi = θ 0 xi + δn0 xi (γ) + εi , i = 1, . . . , n, (4)
where θ = θ2 . Stacking the variables yi and εi results in vectors y, ε ∈ Rn , and stacking the
vectors xi0 and xi (γ)0 yields matrices X, Xγ ∈ Rn×m . Then Equation (4) can be expressed in
matrix notation,
y = X θ + Xγ δn + ε , (5)
where θ, δn , γ are the regression parameters which can be estimated by least squares. By
definition, the least squares estimates θ̂, δ̂, γ̂ jointly minimize the sum of squared errors function,
Sn (θ, δ, γ) = (y − X θ − Xγ δ)0 (y − X θ − Xγ δ) . (6)
The threshold parameter estimate γ̂ may be obtained by minimizing the concentrated sum of
squared errors function Sn (γ),
γ̂ = arg min Sn (γ) , (7)

γ
where
Sn (γ) = y0 y − y0 X∗γ (X∗0 ∗ −1 ∗0

γ Xγ ) Xγ y , (8)
and X∗γ = [X Xγ ]. Given the threshold estimate γ̂, the regression parameter estimates θ̂ = θ̂(γ̂)
and δ̂ = δ̂(γ̂) arise from ordinary least squares regression of y on X∗γ .
To test the hypothesis H0 : γ = γ0 , the likelihood ratio statistic,
Sn (γ) − Sn (γ̂)
LRn (γ) = n , (9)
Sn (γ̂)
iid
is used under the auxiliary assumption εi ∼ N (0, σ 2 ). We reject the likelihood ratio test of
H0 for large values of LRn (γ0 ). Asymptotic p-values for the likelihood ratio test can then be
computed by
1 2
2
pn = 1 − 1 − e− 2 LRn (γ0 ) . (10)
We employ the R package thrreg developed by Kremer (2020) to estimate the threshold re-
gression model.
11
4 Estimation results
4.1 Benchmark econometric model
We estimate the parameters of the benchmark econometric model (1) by Kiesel and Paraschiv
(2017) by the threshold regression described in Section 3.3 for all 96 15-minute contracts. Here, a
selection of 15-minute contracts representative for morning, noon, and evening hours is analyzed,
that is, H7, H13, H18, Q1–Q4 each. We compare our estimation results stemming from January–
December 2015 data with the results by Kiesel and Paraschiv (2017), who study January–June
2014 data. For all contracts, we use the demand quota DQ as threshold variable.
The estimation results of all 15-minute contracts in hours H13, H7, H18 are presented in
Tables 1, 2, 3, respectively. The threshold test reveals strong evidence for a threshold effect in
the demand quota DQ for all contracts in hours H13 and H7, while in hour H18, it suggests
a statistically significant threshold effect for contract Q2 only5 . The estimated threshold pa-
rameter amounts to γ̂ ≈ 1.1 for all contracts. A demand quota of DQ = 1.1 implies that the
expected total electricity demand l exceeds expected conventional capacity c by approximately
10%. Likewise this value implies that roughly 10% renewable power infeed is expected. In gen-
eral, the low demand quota regime (Regime 1) describes a market in which much conventional
power generation is planned to meet expected demand, since little renewable power production
is anticipated. Conversely, the market is in the high demand quota regime (Regime 2) if much
renewable power infeed is anticipated and consequently less conventional power generation is
planned. For all contracts, the total sample is split into subsamples of reasonable size and allows
for a sound interpretation of the estimation results. The adjusted R2 ranges between 9% and
21%.
The estimated coefficients of lagged price changes ∆Pt−1 , ∆Pt−2 are highly statistically sig-
nificant and negative in both regimes for all contracts, and even the coefficients of ∆Pt−3 are
significant and negative in most cases. For evening contracts, the higher-order autoregressive
terms lose statistical significance. The negative coefficients of lagged price changes are an in-
dicator of mean reversion in the price formation process of 15-minute contracts. As such, they
reflect the so-called “learning effect” or “participant conduct” (Karakatsani and Bunn, 2010;
Frauendorfer et al., 2018). Our results confirm previous findings of Kiesel and Paraschiv (2017)
for morning and evening contracts. For noon contracts, in contrast, they find that autoregressive
terms have less explanatory power and intraday trading is primarily driven by renewable fore-
cast changes. Our results, however, suggest that autoregressive terms and the mean reversion
pattern associated therewith play an essential role for intraday trading at noon in 2015.
The estimated coefficients of trading volume Vt are statistically significant for most of the
contracts and regimes. For contracts H13Q1 and H13Q2, they are negative, whereas for H13Q3
and H13Q4, they turn positive independent of the regime. The same sign profile is found for
contracts H18Q1 and H18Q4. This sign pattern is reasonable as it reflects the joint hourly
seasonality of transaction prices and trading volumes described in Section 2.2: In the first half
of an hour, more solar power is produced than the hourly average which induces sell pressure and
thus electricity prices decrease; in the second half of an hour, less solar power is generated than
the hourly average leading to buy pressure and electricity prices increase. In morning hours, we
observe the opposite: For contracts H7Q1 and H7Q2, the coefficients of Vt are positive, whereas
for H7Q3 and H7Q4, they turn negative independent of the regime. This sign pattern coincides
with buy and sell pressure at the beginning and end of the hour, respectively. Overall, we find
an asymmetric response of intraday price changes ∆Pt to trading volumes Vt . This confirms
the results of Kiesel and Paraschiv (2017).
In hour H13, the coefficients of negative wind forecast changes ∆wtn are statistically significant
and negative in both regimes for all contracts, whereas coefficients of positive wind forecast
5
We indicate the statistical significance level of the threshold test by superscript asterisks at the threshold
variable in the regression tables.
12
changes ∆wtp are not significant. In Kiesel and Paraschiv (2017), the coefficients of ∆wtn and
∆wtp are also negative but only significant in the high demand quota regime. In hour H7,
the coefficients of negative wind forecast changes ∆wtn are significant and negative in the high
demand quota regime for Q1–Q3, whereas coefficients of positive wind forecast changes ∆wtp are
significant and negative for Q1 and Q2. In Kiesel and Paraschiv (2017), the coefficients of ∆wtn
are not significant, while some coefficients of ∆wtp are significant but with no clear regularity. In
hour H18, the coefficients of negative and positive wind forecast changes ∆wtn , ∆wtp are rarely
significant and mostly negative. This is in line with Kiesel and Paraschiv (2017). The negative
coefficients of ∆wtn , ∆wtp are economically meaningful and reflect rising and falling electricity
prices as a result of less and more wind power infeed, respectively. Overall, no asymmetry
in the coefficients of wind forecast changes between the regimes is found. We conclude that
forecast errors of wind power generation contribute to pricing electricity for noon and morning
contracts, but they are less significant for pricing evening contracts.
In hours H13 and H7, the coefficients of negative solar forecast changes ∆snt are significant
and negative in the high demand quota regime, but never significant in the low regime for all
contracts. In hour H18, the coefficients of ∆snt are generally not significant. The coefficients’
negative sign is reasonable as electricity prices should increase if less solar power is forecasted.
In Kiesel and Paraschiv (2017), the coefficients of ∆snt are also negative in the high regimes
but only significant for some contracts. In the high demand quota regimes, the coefficients of
∆snt are at least three times larger by absolute value than in the low regimes. The high regime
implies that much renewable power infeed is expected; if solar forecast changes are negative,
they contradict our expectation of the electricity generation mix established day-ahead and thus
influence electricity prices more severely than in the low regime. In Kiesel and Paraschiv (2017),
this asymmetry is less or not at all pronounced.
The coefficients of positive solar forecast changes ∆spt are significant and negative in both
regimes for contracts H13Q1–Q3, while in hours H7 and H18, the coefficients of ∆spt are rarely
significant and mostly negative. The negative sign is intuitive and illustrates that a larger
expectation of solar power infeed decreases electricity prices. In the low demand quota regimes
in hour H13, the coefficients of ∆spt are two to four times larger than in the high regimes. The
low regime implies that little renewable power production is anticipated; it is reasonable that
positive solar forecast changes in the low regime decrease electricity prices more severely than
in the high regime where more renewable power infeed is planned anyway. Our results confirm
the findings of Kiesel and Paraschiv (2017).
Eventually, intraday price changes of 15-minute contracts are driven by both autoregressive
terms and fundamental variables at noon. In morning and evening hours, however, intraday
trading is primarily affected by lagged price changes.
4.2 Extended econometric model

We estimate the parameters of the extended econometric model (2) by the threshold regression
described in Section 3.3 for all 96 15-minute contracts. As representatives of morning, noon,
and evening hours, we exemplarily analyze 15-minute contracts H7, H13, H18, Q1–Q4 each.
For all contracts, the merit-order-curve slope ξ is used as threshold variable.
4.2.1 Merit-order-curve slope

We estimate the slope of the merit order curve ξ (i) of 15-minute contract i based on intraday
auction prices P Auc,(i) and corresponding expected demands `(i) . As both P Auc,(i) and `(i)
are provided in quarter-hourly resolution, we are able to determine a merit order curve for
each 15-minute contract i individually. Figure 6 depicts a scatter plot of empirical intraday
Auc,(i) (i)
auction prices Pd versus empirical expected demands `d observed on days d = 1, . . . , T ,
for contract i = H13Q4. We filter out a total of 13 negative auction prices. We observe a
13
H13Q1 H13Q2
Regime 1 Regime 2 Regime 1 Regime 2
DQ∗∗∗ ≤ 1.100 DQ∗∗∗ > 1.100 DQ∗∗∗ ≤ 1.065 DQ∗∗∗ > 1.065
Variable Estimate Std error Estimate Std error Variable Estimate Std error Estimate Std error
Const −1.337 (0.990) −0.391 (1.725) Const −1.403 (1.534) −1.765 (1.455)
DQ 1.466 (0.977) 0.361 (1.464) DQ 1.600 (1.574) 1.442 (1.239)
∆Pt−1 −0.318∗∗∗ (0.026) −0.269∗∗∗ (0.024) ∆Pt−1 −0.281∗∗∗ (0.028) −0.304∗∗∗ (0.029)
∆Pt−2 −0.151∗∗∗ (0.021) −0.072∗∗∗ (0.018) ∆Pt−2 −0.142∗∗∗ (0.024) −0.093∗∗∗ (0.019)
∆Pt−3 −0.093∗∗∗ (0.025) −0.059∗∗ (0.019) ∆Pt−3 −0.047∗ (0.025) 0.014 (0.030)
Vt −0.022∗∗∗ (0.004) −0.014∗∗∗ (0.004) Vt −0.022∗∗∗ (0.007) −0.005 (0.008)
∆wtn −1.137∗∗∗ (0.322) −1.180∗∗∗ (0.260) ∆wtn −1.300∗∗∗ (0.351) −1.175∗∗∗ (0.245)
∆wtp −0.796∗ (0.332) −0.587 (0.303) ∆wtp −0.606 (0.378) −0.220 (0.208)
∆snt 0.111 (0.428) −1.730∗∗∗ (0.534) ∆snt −0.390 (0.552) −1.738∗∗∗ (0.385)
∆spt −3.545∗∗∗ (0.492) −1.374∗∗∗ (0.428) ∆spt −3.442∗∗∗ (0.432) −1.329∗∗ (0.438)
∆t 0.092∗∗∗ (0.020) 0.061∗∗∗ (0.021) ∆t 0.053 (0.026) 0.035 (0.019)
#Obs 6499 6554 #Obs 4279 7764
2
Radj 0.158 0.143 2
Radj 0.121 0.136
H13Q3 H13Q4
DQ∗ ≤ 1.019 DQ∗ > 1.019 DQ∗∗∗ ≤ 1.148 DQ∗∗∗ > 1.148
Const 0.574 (1.825) −2.007 (1.175) Const −2.969 (0.712) −4.635 (2.765)
DQ −0.877 (1.950) 1.507 (1.023) DQ 2.609 (0.671) 3.765 (2.299)
∆Pt−1 −0.270∗∗∗ (0.028) −0.269∗∗∗ (0.021) ∆Pt−1 −0.292∗∗∗ (0.021) −0.219∗∗∗ (0.023)
∆Pt−2 −0.116∗∗∗ (0.030) −0.138 (0.019) ∆Pt−2 −0.128∗∗∗ (0.015) −0.102∗∗∗ (0.021)
∆Pt−3 −0.027 (0.023) −0.073 (0.016) ∆Pt−3 −0.049∗∗∗ (0.014) −0.094∗∗∗ (0.020)
Vt 0.014∗ (0.008) 0.012 (0.004) Vt 0.014∗∗ (0.003) 0.007 (0.003)
∆wtn −0.892∗ (0.479) −1.125 (0.254) ∆wtn −1.527∗∗∗ (0.298) −1.203∗∗∗ (0.313)
∆wtp −0.134 (0.318) −0.522 (0.219) ∆wtp −0.355 (0.220) −0.299 (0.443)
∆snt 0.204 (0.449) −2.063∗∗∗ (0.446) ∆snt −0.811 (0.333) −3.001∗∗∗ (0.684)
∆spt −2.292∗∗∗ (0.518) −1.267 (0.461) ∆spt −0.450 (0.498) −1.779 (0.696)
∆t −0.023 (0.028) 0.026 (0.019) ∆t −0.036 (0.020) −0.025 (0.027)
#Obs 3538 9278 #Obs 10 372 4782
2
Radj 0.085 0.125 2
Radj 0.110 0.148
∗p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01
Table 1: Estimation results of the benchmark econometric model (1) for intraday price changes
∆Pt of 15-minute contracts H13Q1–4.
14
H7Q1 H7Q2
DQ∗∗∗ ≤ 1.022 DQ∗∗∗ > 1.022 DQ∗∗∗ ≤ 1.042 DQ∗∗∗ > 1.042
Const 1.277 (1.613) −0.575 (2.142) Const 0.332 (1.557) 0.668 (2.484)
DQ −2.147 (1.833) −0.379 (1.896) DQ −0.748 (1.743) −1.099 (2.180)
∆Pt−1 −0.332∗∗∗ (0.031) −0.375∗∗∗ (0.025) ∆Pt−1 −0.307∗∗∗ (0.026) −0.303∗∗∗ (0.023)
∆Pt−2 −0.148∗∗∗ (0.031) −0.169∗∗∗ (0.021) ∆Pt−2 −0.180∗∗∗ (0.025) −0.128∗∗∗ (0.018)
∆Pt−3 −0.087∗∗∗ (0.028) −0.063∗∗∗ (0.018) ∆Pt−3 −0.100∗∗∗ (0.021) −0.083∗∗∗ (0.016)
Vt 0.043∗∗∗ (0.006) 0.038∗∗∗ (0.003) Vt 0.059∗∗∗ (0.012) 0.021∗∗ (0.008)
∆wtn −1.271 (0.742) −0.808∗∗ (0.353) ∆wtn −0.376 (0.600) −0.686∗∗ (0.264)
∆wtp −1.331∗∗∗ (0.383) −0.577∗∗ (0.265) ∆wtp −1.471∗∗∗ (0.430) −0.622∗ (0.322)
∆snt 8.318 (4.777) −3.889 (4.843) ∆snt 2.863 (2.657) −8.559∗ (4.051)
∆spt −6.672 (5.434) 8.814∗ (4.036) ∆spt −2.338 (4.187) 1.674 (2.340)
∆t −0.094 (0.050) 0.093∗∗ (0.034) ∆t −0.145∗∗ (0.046) 0.113∗∗∗ (0.033)
#Obs 3092 6953 #Obs 2813 6180
2
Radj 0.155 0.159 2
Radj 0.159 0.110
H7Q3 H7Q4
DQ∗∗∗ ≤ 1.053 DQ∗∗∗ > 1.053 DQ∗∗∗ ≤ 1.156 DQ∗∗∗ > 1.156
Const −1.927 (1.514) −2.104 (2.772) Const −2.178 (0.781) −1.899 (5.697)
DQ 2.532 (1.646) 1.670 (2.429) DQ 2.115 (0.743) 1.423 (4.773)
∆Pt−1 −0.322∗∗∗ (0.026) −0.321∗∗∗ (0.029) ∆Pt−1 −0.332∗∗∗ (0.019) −0.228∗∗∗ (0.034)
∆Pt−2 −0.139∗∗∗ (0.024) −0.148∗∗∗ (0.022) ∆Pt−2 −0.154∗∗∗ (0.016) −0.101∗∗∗ (0.029)
∆Pt−3 −0.061∗∗ (0.022) −0.093∗∗∗ (0.021) ∆Pt−3 −0.061 (0.015) −0.039 (0.033)
Vt −0.026∗∗ (0.010) −0.020∗∗∗ (0.006) Vt −0.014∗ (0.003) −0.014∗∗ (0.005)
∆wtn −0.111 (0.568) −0.781∗ (0.356) ∆wtn −0.888 (0.400) −0.187 (0.647)
∆wtp −0.636 (0.335) −0.281 (0.402) ∆wtp −1.075∗∗ (0.307) 0.715 (0.472)
∆snt 2.767 (2.280) −10.365∗ (4.773) ∆snt 1.202 (2.256) −8.426∗∗ (3.314)
∆spt −3.519 (3.005) −1.861 (1.975) ∆spt −6.350∗∗ (2.088) −3.801 (2.609)
∆t −0.085 (0.043) 0.131∗∗∗ (0.035) ∆t 0.064 (0.033) 0.164∗∗ (0.053)
#Obs 3154 6399 #Obs 8308 3173
2
Radj 0.111 0.139 2
Radj 0.126 0.120
∗p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01
15
H18Q1 H18Q2
DQ ≤ 1.135 DQ > 1.135 DQ∗ ≤ 1.118 DQ∗ > 1.118
Const 0.151 (0.984) 9.257 (7.735) Const −0.145 (1.045) −1.549 (5.186)
DQ 0.075 (0.916) −7.531 (6.525) DQ 0.180 (0.986) 1.364 (4.503)
∆Pt−1 −0.278∗∗ (0.039) −0.431∗∗ (0.174) ∆Pt−1 −0.322∗∗∗ (0.019) −0.277∗∗∗ (0.029)
∆Pt−2 −0.115 (0.022) −0.176 (0.166) ∆Pt−2 −0.112∗∗ (0.016) −0.138∗∗∗ (0.023)
∆Pt−3 −0.018 (0.020) 0.077 (0.128) ∆Pt−3 −0.063 (0.014) −0.010 (0.020)
Vt −0.003 (0.003) −0.012 (0.005) Vt 0.010 (0.006) −0.006 (0.007)
∆wtn −1.831 (0.243) −0.609 (0.729) ∆wtn 0.101 (0.618) −3.040 (1.238)
∆wtp −0.539 (0.606) −2.141 (0.546) ∆wtp −1.736∗∗ (0.371) 0.652 (0.668)
∆snt −1.823 (1.332) −2.994 (3.769) ∆snt −0.848 (1.465) 1.752 (3.483)
∆spt −4.223 (1.224) −2.563 (2.356) ∆spt −1.716 (1.551) 0.654 (1.895)
∆t −0.052 (0.015) −0.036 (0.042) ∆t −0.016 (0.016) −0.015 (0.029)
#Obs 9790 2403 #Obs 7672 2973
2
Radj 0.105 0.191 2
Radj 0.121 0.103
H18Q3 H18Q4
DQ ≤ 1.144 DQ > 1.144 DQ ≤ 1.171 DQ > 1.171
Const −0.443 (0.893) 5.784 (9.044) Const −0.378 (0.730) 21.412 (39.252)
DQ 0.236 (0.838) −4.823 (7.749) DQ 0.279 (0.693) −17.492 (32.591)
∆Pt−1 −0.308∗∗∗ (0.016) −0.291∗∗∗ (0.044) ∆Pt−1 −0.312∗∗∗ (0.021) −0.386∗ ( 0.212)
∆Pt−2 −0.150∗∗∗ (0.016) −0.045 (0.032) ∆Pt−2 −0.126∗∗∗ (0.016) −0.118 ( 0.148)
∆Pt−3 −0.044∗∗∗ (0.012) −0.009 (0.026) ∆Pt−3 −0.030 (0.013) −0.266 ( 0.101)
Vt 0.006 (0.007) −0.008 (0.011) Vt 0.001 (0.002) 0.051 ( 0.027)
∆wtn −0.423 (0.326) −3.505∗∗∗ (0.822) ∆wtn −1.897 (0.977) 2.568 ( 5.729)
∆wtp −1.480∗∗ (0.673) 2.335 (2.162) ∆wtp −0.180 (0.297) −5.008 ( 3.700)
∆snt −4.480∗∗ (1.830) 0.073 (2.401) ∆snt 1.935 (1.927) 7.501 ( 6.666)
∆spt −0.267 (1.432) 16.354∗∗ (7.125) ∆spt −4.024∗ (1.563) 4.548 ( 6.819)
∆t 0.018 (0.014) −0.066 (0.055) ∆t 0.008 (0.016) −0.130 ( 0.127)
#Obs 9745 1249 #Obs 12 779 326
2
Radj 0.112 0.149 2
Radj 0.114 0.208
∗p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01
16
Auc,(i)
positive relationship between intraday auction prices and expected demands: Pd increases
(i)
as `d increases. For high levels of demand, more expensive generation technologies are in
use which puts upward pressure on prices. In particular, we may identify two clusters: One
cluster encompasses expected demands ` < 58 GW, and the other cluster comprises ` ≥ 58 GW.
These clusters reflect low and high electricity demand on weekends and weekdays, respectively.
Overall, the data points exhibit a fairly wide range of variation which stems from the strongly
varying expectation of renewable power infeed.
For the empirical merit order curve f (`), we use the exponential function f (`) = ea `+b
following He et al. (2013). Of course other choices for f (`) are possible, but we want to keep
our model as simple as possible. We fit the exponential function f (`) to the empirical intraday
auction price P Auc,(i) as a function of expected demand `(i) . Technically, we minimize the sum of
squared errors between the logarithm of the function f Φ (`) implied by the choice of a parameter

Auc,(i) (i)
set Φ = {a, b}, and the logarithm of the empirical intraday auction price Pd `d ,
T 2
X Auc,(i) (i)
min log Pd `d − log f Φ (`) . (11)
Φ
d=1
The least squares fit for contract i = H13Q4 is shown in Figure 6. The parameter estimates of
a, b for contracts H7, H13, H18, Q1–Q4 each, are reported in Table 4.
We observe an hourly seasonality of the parameter estimates â, b̂. For contracts H4Q1–H14Q4,
the estimates â, b̂ increase and decrease from the first to the last 15-minute contract within an
hour, respectively, whereas for contracts H15Q1–H2Q4, the estimates â, b̂ decrease and increase,
respectively. Thus, the curvature of the empirical merit order curve f (`) grows from Q1 to Q4 in
each hour during morning and noon hours, whereas it declines from Q1 to Q4 during afternoon
and evening hours. This hourly seasonality may be associated with rising and falling demand
in the first and second half of the day, respectively, following human activity (Paraschiv, 2013).
Consequently, in the first half of the day, more expensive power plants are needed to cover
demand in the last than in the first quarter-hour in each hour, whereas in the second half of
the day, more expensive power plants are operated in the first than in the last quarter-hour per
hour. Φ
Eventually, we take the derivative of the fitted merit-order-curve function f 0 Φ (`) = dfd`(`)
0Φ (i)
expected demands `d into f (`) to obtain empirical merit-order-curve
and substitute empirical
(i) (i)
slopes ξd = f 0 Φ `d on days d = 1, . . . , T , for 15-minute contract i.
Auction price P Auc [EUR/MWh]
●
●
●
data ●
● ●●
fit ●● ●
●●
●●
● ● ●●
●●
●
● ● ●● ● ● ●●
●● ● ●
●
40 ● ●
● ●
●
● ● ●●
● ●● ●● ●●
● ●● ● ●
●● ●●● ●●●
● ●
●
●
●●● ●●●●
● ● ●● ●● ●● ● ●●●●● ● ● ●
● ●● ● ●● ●
● ● ●● ●
● ● ● ●● ● ●● ● ●●● ● ●
● ●●● ● ● ●● ● ●
● ●●
● ●●●●
●
● ●● ● ● ● ● ●● ● ●●● ●● ●●●● ● ●
●
● ●
● ●● ● ●● ● ● ● ●● ● ●●
● ● ● ●
● ● ●● ●
● ●●●●●
● ●● ●●
● ● ● ●●● ● ●
● ● ●
● ● ●●
● ● ● ● ●● ● ● ● ● ● ●
●● ● ● ●● ● ●
20 ● ●
●
● ●● ● ●
● ●
● ●●
●●
●
●
●
●
●● ● ● ●
●●
●● ● ●
●● ● ●
● ● ● ●
●
● ●● ● ● ● ●●
● ● ●
● ●●
●●●
● ● ●
● ●
●
●
●● ●● ●
● ● ● ● ●●
●● ●
● ●●● ●● ● ●● ● ● ●●
● ●
●
● ●
●
● ● ●● ●
● ●
●
●●
0 ● ●● ● ● ● ●
40 50 60 70
Expected demand l [GW]
Figure 6: Empirical merit-order-curve function f (`) = ea `+b (red) fitted to the empirical intra-
Auc,(i) (i)
day auction price Pd as a function of expected demand `d (blue) observed on
days d = 1, . . . , T , for 15-minute contract i = H13Q4.
17
Parameter
a b
Contract Estimate Std error Estimate Std error
H7Q1 0.029 (0.004) 1.632 (0.268)
H7Q2 0.036 (0.002) 1.412 (0.134)
H7Q3 0.045 (0.002) 0.961 (0.125)
H7Q4 0.045 (0.002) 0.940 (0.130)
H13Q1 0.022 (0.002) 2.194 (0.141)
H13Q2 0.037 (0.003) 1.214 (0.201)
H13Q3 0.045 (0.004) 0.548 (0.217)
H13Q4 0.065 (0.006) -0.856 (0.389)
H18Q1 0.061 (0.005) -0.372 (0.313)
H18Q2 0.039 (0.003) 1.246 (0.160)
H18Q3 0.021 (0.002) 2.546 (0.109)
H18Q4 0.013 (0.002) 3.074 (0.140)
Table 4: Parameter estimates of a, b of the empirical merit-order-curve function f (`) = ea `+b

for 15-minute contracts H7, H13, H18, Q1–Q4 each.
4.2.2 Model calibration

The estimation results of all 15-minute contracts in hours H13, H7, H18 are presented in Ta-
bles 5, 6, 7, respectively. The threshold test shows strong evidence for a threshold effect in the
slope of the merit order curve ξ for all contracts in hours H13 and H7, while in hour H18, it
indicates a statistically significant threshold effect for contract Q2 and Q3. A merit-order-curve
slope of ξ = 1 EUR/MWh GW implies that a change of 1 GW in expected demand causes a change
of 1 EUR/MWh in the intraday auction price. Regime 1 corresponds to a market in which the
merit order curve is flat, whereas Regime 2 reflects a market in which the merit order curve is
steep. For all contracts, the total sample is split into subsamples of appropriate size and enables
us to interpret the estimation results adequately.
The adjusted R2 ranges between 11% and 22%. Thus, overall, our extended econometric
model exhibits higher explanatory power in explaining intraday price changes ∆Pt than the
benchmark model by Kiesel and Paraschiv (2017). In particular, our extended model outnum-
bers the benchmark model’s adjusted R2 for each contract and regime in hours H13 and H7. In
hour H18, the extended model only improves the explanation of intraday price changes ∆Pt in
either regime compared to the benchmark model.
The estimated coefficients of lagged price changes ∆Pt−1 , ∆Pt−2 , ∆Pt−3 are highly statisti-
cally significant and negative in both regimes for all contracts in hours H13 and H7. In hour
H18, the higher-order autoregressive terms show less statistical significance. The negative coeffi-
cients of lagged price changes suggest mean reversion in the price formation process of 15-minute
contracts. We recall that Kiesel and Paraschiv (2017) find less explanatory power of autoregres-
sive terms during noon hours when renewable forecast changes primarily dominate 15-minute
intraday trading. We, however, provide evidence of a significant mean reversion effect for noon
contracts in 2015.
Overall, in hour H13, the estimated coefficients of price changes of neighboring contracts
(i−2) (i+2)
∆Pt , . . . , ∆Pt are highly statistically significant and positive in both regimes for all
(i−2) (i+2)
contracts. In hour H7, the estimated coefficients of ∆Pt , . . . , ∆Pt are highly statistically
significant and positive in the steep merit-order regime for all contracts as well as in the flat
regime for contracts Q3 and Q4. Thus, price changes of neighboring contracts have strong
18
explanatory power and a positive effect on one another. Moreover, we observe that price changes
of the nearest neighbors i ± 1 have a stronger impact on price changes of contract i than price
changes of next-nearest neighbors i±2 6 . In particular, for the first 15-minute contract Q1 in an
(i+1)
hour, price changes ∆Pt of the following contract Q2 have the greatest influence, whereas for
(i−1)
the last contract Q4 in that hour, price changes ∆Pt of the preceding contract Q3 exhibit the
largest impact. Hence, 15-minute contracts within the same hour are the most important price
(i−2) (i+2)
drivers. In hour H18, the coefficients of ∆Pt , . . . , ∆Pt are rarely statistically significant.
The coefficients of the intraday auction price P Auc are not significant for all contracts and
regimes. This is not surprising since P Auc can merely be considered as an estimate of the initial
price of a 15-minute contract in the continuous trading. Thus, it is not expected that P Auc
affects continuous trading beyond the first price.
In hour H13, the coefficients of trading volume Vt are significant in the flat merit-order regime
for all contracts and in the steep regime for contracts Q1 and Q4. Independent of the regime,
they are negative for Q1 and Q2 and positive for Q3 and Q4. This sign profile illustrates the
joint hourly seasonality of intraday prices and volumes presented in Section 2.2: As more solar
power is generated than the hourly average in Q1 and Q2, there is sell pressure at the beginning
of the hour and intraday prices decrease. Conversely, buy pressure increases intraday prices at
the end of the hour (Q3 and Q4) when below-average solar power is produced. In hour H7,
the coefficients of Vt are significant for all contracts and regimes and we observe the opposite
sign pattern: For contracts Q1 and Q2, the coefficients are positive, whereas for Q3 and Q4,
they turn negative independent of the regime. Thereby, the positive and negative coefficients
reflect buy and sell pressure at the beginning and end of the hour, respectively. In hour H18,
the coefficients of Vt are not significant. In general, we do not find an asymmetric adjustment
of intraday price changes ∆Pt to trading volumes Vt .
Overall, the coefficients of negative wind forecast changes ∆wtn are significant and negative in
the steep merit-order regime for all contracts. The negative sign is economically meaningful as
electricity prices should increase if less wind power is forecasted. In hour H13, the coefficients
of positive wind forecast changes ∆wtp are only significant and negative in the steep regime for
Q1 and Q3. In hour H7, the coefficients of ∆wtp are significant and negative in the flat regime
for all contracts as well as in the steep regime for Q1 and Q2, while in hour H18, they are
significant and negative in both regimes for Q3. The negative coefficients imply that electricity
prices decline in the wake of rising wind power infeed, which is consistent with our intuition.
No asymmetry in the coefficients of wind forecast changes between the regimes is observed.
We conclude that generally wind forecast errors have explanatory power to explaining intraday
price changes ∆Pt .
The coefficients of negative solar forecast changes ∆snt are significant and negative in the
steep merit-order regime, but not significant in the flat regime for contracts in hours H13 and
H7. In hour H18, the coefficients of ∆snt are significant in both regimes for Q3. In the steep
regimes, the coefficients of ∆snt are at least two times larger by absolute value than in the
flat regimes. This reflects the fact that renewable forecast changes affect electricity prices more
severely in the steep than in the flat merit-order regime. In hour H13, the coefficients of positive
solar forecast changes ∆spt are significant and negative in both regimes for contracts Q1 and Q2.
In hour H7, the coefficients of ∆spt are significant and negative in both regimes for Q4, while
in hour H18, they are generally not significant. The negative coefficients of both negative and
positive forecast errors of solar power production are reasonable: A lower expectation of solar
power infeed should increase electricity prices, whereas a larger solar power prediction decreases
electricity prices.
Generally, renewable forecast changes are more significant in the steep than in the flat merit-
order regime. When the market is in the steep merit-order regime, market participants rely
6
The terms nearest and next-nearest neighbors are motivated from physics where a nearest- and next-nearest-
neighbor interaction exists.
19
on the use of more expensive power generation technologies, which sets additional pressure on
them to balance the volatile renewable energies on the intraday market.
Eventually, intraday trading of 15-minute contracts is driven by both price-related and funda-
mental variables at noon, while the price-related variables have a slight surplus importance. For
morning and evening contracts, however, intraday price changes are predominantly influenced
by past, idiosyncratic price information and the price information of neighboring contracts.
5 Conclusion
This paper develops an econometric price model with fundamental impacts for intraday electric-
ity markets of 15-minute contracts. We analyze a novel and unique data set of high-frequency
transaction data, fundamental supply and demand data, and intradaily updated forecasts of
wind and solar power generation. The nature of our data set allows the model specification to
solely include ex-ante market information.
We perform an empirical exploration of transaction prices and trading volumes of 15-minute
contracts. Empirical evidence suggests that, on average, transaction prices of 15-minute con-
tracts exhibit a sawtooth-shaped and trading volumes a U-shaped hourly seasonality. Moreover,
liquidity increases sharply within the last trading hour before gate closure. Our empirical anal-
ysis also indicates that renewable forecast updates are reflected in intraday prices within one
trading minute.
We refine the econometric model by Kiesel and Paraschiv (2017) along three dimensions by
incorporating: (i) the slope of the merit order curve, (ii) price changes of neighboring 15-minute
contracts, (iii) the 15-minute intraday auction price. We calibrate our econometric model to
market data for a selection of morning, noon, and evening contracts. A threshold regression
model is used to examine how 15-minute intraday trading depends on the slope of the merit
order curve.
Our estimation results reveal that autoregressive price changes up to the third order are highly
statistically significant and negative, independent of the time of day. This behavior provides
clear evidence of mean reversion in the price formation mechanism of 15-minute contracts.
Additionally, price changes of neighboring contracts exhibit strong explanatory power and a
positive impact on price changes of a given 15-minute contract. We observe an asymmetric
effect of positive and negative renewable forecast changes on intraday prices depending on
the merit-order-curve slope: Renewable forecasts affect electricity prices more severely in the
steep than in the flat merit-order regime. In general, renewable forecast changes have a higher
explanatory power for pricing noon than morning and evening contracts, but price information
is the key driver of 15-minute intraday trading. Overall, we conclude that the importance of
influencing factors on the intraday electricity market has changed from fundamental towards
trading-related factors.
As our econometric model exclusively involves ex-ante market knowledge, it allows to develop
trading strategies for intraday electricity markets, tailor-made for each contract. Furthermore,
it helps to design forecasting models for single intraday transaction prices in the continuous
trading – to our knowledge, an unexplored territory of scientific research hitherto. Moreover,
our article provides a valuable step towards the optimization of the bidding behavior on intraday
markets. Eventually, our insights should prove useful to energy companies within the process
of automating intraday electricity trading.
20
H13Q1 H13Q2
ξ ∗∗ ≤ 0.689 ξ ∗∗ > 0.689 ξ ∗∗∗ ≤ 0.912 ξ ∗∗∗ > 0.912
Const −1.504 (1.643) −0.867 (0.911) Const −0.201 (1.225) 0.304 (0.734)
ξ 2.610 (2.672) 1.021 (1.054) ξ 0.931 (1.622) −0.371 (0.548)
∆Pt−1 −0.363∗∗∗ (0.040) −0.311∗∗∗ (0.018) ∆Pt−1 −0.332∗∗∗ (0.034) −0.322∗∗∗ (0.023)
∆Pt−2 −0.188∗∗∗ (0.030) −0.131∗∗∗ (0.015) ∆Pt−2 −0.172∗∗∗ (0.028) −0.132∗∗∗ (0.016)
∆Pt−3 −0.120∗∗∗ (0.035) −0.088∗∗∗ (0.015) ∆Pt−3 −0.058∗ (0.032) −0.021 (0.020)
(i−2) (i−2)
∆Pt 0.064∗∗ (0.025) 0.020 (0.010) ∆Pt 0.093∗∗∗ (0.027) 0.053∗∗∗ (0.014)
(i−1) (i−1)
∆Pt 0.063∗∗ (0.025) 0.051∗∗∗ (0.012) ∆Pt 0.212∗∗∗ (0.034) 0.168∗∗∗ (0.021)
(i+1) (i+1)
∆Pt 0.144∗∗∗ (0.034) 0.136∗∗∗ (0.016) ∆Pt 0.109∗∗∗ (0.025) 0.140∗∗∗ (0.024)
(i+2) (i+2)
∆Pt 0.049∗∗ (0.024) 0.036∗∗∗ (0.011) ∆Pt 0.021 (0.030) 0.042 (0.016)
P Auc 0.000 (0.015) 0.002 (0.006) P Auc −0.011 (0.021) 0.006 (0.007)
Vt −0.025∗∗∗ (0.006) −0.017∗∗∗ (0.003) Vt −0.022∗ (0.010) −0.007 (0.007)
∆wtn −0.528 (0.378) −1.135∗∗∗ (0.208) ∆wtn −0.564 (0.280) −1.179∗∗∗ (0.202)
∆wtp −0.554 (0.421) −0.688∗∗ (0.247) ∆wtp −0.390 (0.346) −0.228 (0.188)
∆snt 0.852 (0.362) −1.604∗∗∗ (0.417) ∆snt 0.181 (0.344) −1.750∗∗∗ (0.374)
∆spt −2.842∗∗∗ (0.513) −1.860∗∗∗ (0.403) ∆spt −2.629∗∗∗ (0.398) −1.485∗∗∗ (0.376)
∆t 0.098∗∗∗ (0.029) 0.058∗∗∗ (0.016) ∆t 0.029 (0.029) 0.027 (0.016)
#Obs 3131 9922 #Obs 2759 9284
2
Radj 0.201 0.190 2
Radj 0.188 0.198
H13Q3 H13Q4
ξ ∗∗ ≤ 1.324 ξ ∗∗ > 1.324 ξ ∗∗ ≤ 0.995 ξ ∗∗ > 0.995
Const −0.687 (0.390) −1.215 (0.986) Const −0.396 (0.541) −0.183 (0.377)
ξ 1.016 (0.497) 0.707 (0.649) ξ −0.111 (0.800) −0.196 (0.204)
∆Pt−1 −0.287∗∗∗ (0.024) −0.306∗∗∗ (0.023) ∆Pt−1 −0.331∗∗∗ (0.027) −0.298∗∗∗ (0.019)
∆Pt−2 −0.143∗∗∗ (0.024) −0.165∗∗∗ (0.021) ∆Pt−2 −0.176∗∗∗ (0.025) −0.135∗∗∗ (0.014)
∆Pt−3 −0.058∗ (0.018) −0.080∗∗∗ (0.017) ∆Pt−3 −0.095∗∗∗ (0.023) −0.071∗∗∗ (0.013)
(i−2) (i−2)
∆Pt 0.061∗∗∗ (0.017) 0.044 (0.017) ∆Pt 0.038 (0.018) 0.043∗∗∗ (0.013)
(i−1) (i−1)
∆Pt 0.059∗∗∗ (0.016) 0.147∗∗∗ (0.031) ∆Pt 0.114∗∗∗ (0.021) 0.168∗∗∗ (0.022)
(i+1) (i+1)
∆Pt 0.147∗∗∗ (0.021) 0.230∗∗∗ (0.029) ∆Pt 0.069∗∗∗ (0.021) 0.062∗∗∗ (0.013)
(i+2) (i+2)
∆Pt 0.075∗∗ (0.019) 0.044 (0.014) ∆Pt 0.045∗∗ (0.018) 0.061∗∗∗ (0.014)
P Auc −0.016 (0.017) −0.004 (0.008) P Auc 0.011 (0.014) 0.010∗ (0.006)
Vt 0.012∗ (0.006) 0.014 (0.004) Vt 0.012∗ (0.005) 0.013∗∗∗ (0.003)
∆wtn −1.029 (0.333) −0.753∗∗∗ (0.205) ∆wtn −1.473∗∗∗ (0.396) −1.188∗∗∗ (0.231)
∆wtp −0.094 (0.184) −0.594∗∗ (0.261) ∆wtp −0.366 (0.329) −0.407 (0.222)
∆snt 0.133 (0.355) −2.276∗∗∗ (0.437) ∆snt −0.211 (0.377) −2.403∗∗∗ (0.406)
∆spt −1.967∗∗∗ (0.319) −0.943 (0.474) ∆spt −1.232 (0.481) −0.259 (0.496)
∆t −0.011 (0.021) 0.027 (0.021) ∆t −0.068∗ (0.028) −0.031 (0.019)
#Obs 5626 7190 #Obs 3893 11 261
2
Radj 0.129 0.214 2
Radj 0.144 0.182
∗p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01
Table 5: Estimation results of the extended econometric model (2) for intraday price changes
21
H7Q1 H7Q2
ξ ∗∗∗ ≤ 0.722 ξ ∗∗∗ > 0.722 ξ ∗∗∗ ≤ 1.050 ξ ∗∗∗ > 1.050
Const 2.057 (1.198) −0.224 (1.234) Const 0.233 (0.995) −0.429 (0.954)
ξ −5.064∗∗ (1.958) −0.541 (1.179) ξ −0.567 (1.232) −0.084 (0.605)
∆Pt−1 −0.354∗∗∗ (0.032) −0.390∗∗∗ (0.023) ∆Pt−1 −0.341∗∗∗ (0.028) −0.322∗∗∗ (0.021)
∆Pt−2 −0.179∗∗∗ (0.032) −0.184∗∗∗ (0.020) ∆Pt−2 −0.193∗∗∗ (0.026) −0.149∗∗∗ (0.017)
∆Pt−3 −0.114∗∗∗ (0.029) −0.064∗∗∗ (0.018) ∆Pt−3 −0.116∗∗∗ (0.022) −0.094∗∗∗ (0.015)
(i−2) (i−2)
∆Pt 0.024 (0.022) 0.032∗∗ (0.012) ∆Pt 0.024 (0.021) 0.014 (0.014)
(i−1) (i−1)
∆Pt −0.001 (0.026) 0.012 (0.014) ∆Pt 0.201∗∗∗ (0.029) 0.104∗∗∗ (0.017)
(i+1) (i+1)
∆Pt 0.155∗∗∗ (0.025) 0.101∗∗∗ (0.015) ∆Pt 0.065∗∗∗ (0.024) 0.137∗∗∗ (0.019)
(i+2) (i+2)
∆Pt −0.011 (0.022) 0.060∗∗∗ (0.016) ∆Pt 0.050 (0.030) 0.047∗∗∗ (0.017)
P Auc 0.016 (0.016) −0.009 (0.010) P Auc 0.002 (0.018) −0.002 (0.013)
Vt 0.040∗∗∗ (0.006) 0.037∗∗∗ (0.003) Vt 0.054∗∗∗ (0.012) 0.023∗∗∗ (0.008)
∆wtn 0.202 (0.828) −1.234∗∗∗ (0.277) ∆wtn 0.384 (0.448) −0.855∗∗∗ (0.217)
∆wtp −1.388∗∗∗ (0.386) −0.661∗∗∗ (0.242) ∆wtp −1.367∗∗∗ (0.437) −0.658∗∗ (0.285)
∆snt 6.741 (4.483) −5.254 (3.889) ∆snt 0.411 (2.432) −9.320∗∗∗ (3.522)
∆spt 0.415 (5.261) 6.708∗ (4.008) ∆spt 0.384 (3.563) 1.298 (2.208)
∆t −0.089∗ (0.049) 0.085∗∗ (0.032) ∆t −0.154∗∗∗ (0.044) 0.101∗∗∗ (0.030)
#Obs 2923 7122 #Obs 2573 6420
2
Radj 0.186 0.188 2
Radj 0.218 0.157
H7Q3 H7Q4
ξ ∗∗∗ ≤ 1.403 ξ ∗∗∗ > 1.403 ξ ∗∗∗ ≤ 1.459 ξ ∗∗∗ > 1.459
Const −0.498 (0.770) 0.217 (0.750) Const −0.855 (0.684) 0.421 (0.645)
ξ 0.995 (0.881) −0.219 (0.376) ξ 1.118 (0.736) −0.456 (0.320)
∆Pt−1 −0.364∗∗∗ (0.025) −0.342∗∗∗ (0.027) ∆Pt−1 −0.363∗∗∗ (0.031) −0.319∗∗∗ (0.019)
∆Pt−2 −0.169∗∗∗ (0.025) −0.172∗∗∗ (0.020) ∆Pt−2 −0.193∗∗∗ (0.028) −0.148∗∗∗ (0.017)
∆Pt−3 −0.061∗∗ (0.022) −0.108∗∗∗ (0.019) ∆Pt−3 −0.063∗∗∗ (0.023) −0.069∗∗∗ (0.019)
(i−2) (i−2)
∆Pt 0.068∗∗ (0.024) 0.052∗∗∗ (0.015) ∆Pt 0.049∗ (0.019) 0.077∗∗∗ (0.012)
(i−1) (i−1)
∆Pt 0.063∗∗ (0.020) 0.102∗∗∗ (0.015) ∆Pt 0.150∗∗∗ (0.024) 0.160∗∗∗ (0.019)
(i+1) (i+1)
∆Pt 0.187∗∗∗ (0.027) 0.147∗∗∗ (0.019) ∆Pt 0.036 (0.029) 0.034∗ (0.018)
(i+2) (i+2)
∆Pt 0.087∗∗ (0.029) 0.070∗∗∗ (0.026) ∆Pt 0.052∗ (0.028) 0.065∗∗∗ (0.015)
P Auc −0.001 (0.019) 0.000 (0.012) P Auc 0.002 (0.015) 0.006 (0.009)
Vt −0.029∗∗ (0.010) −0.018∗∗∗ (0.006) Vt −0.017∗∗ (0.005) −0.013∗∗∗ (0.003)
∆wtn 0.479 (0.429) −0.769∗∗ (0.342) ∆wtn 1.038 (0.649) −0.894∗∗ (0.352)
∆wtp −0.686∗ (0.350) −0.190 (0.347) ∆wtp −1.058∗∗ (0.364) −0.285 (0.363)
∆snt 0.724 (1.919) −10.354∗∗ (4.116) ∆snt 0.848 (2.429) −8.062∗∗∗ (2.585)
∆spt −1.730 (2.524) −3.032 (1.680) ∆spt −5.688∗∗ (2.541) −3.543∗ (1.621)
∆t −0.112∗∗ (0.040) 0.127∗∗∗ (0.030) ∆t −0.050 (0.044) 0.144∗∗∗ (0.029)
#Obs 2694 6859 #Obs 3407 8074
2
Radj 0.189 0.189 2
Radj 0.190 0.177
∗p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01
22
H18Q1 H18Q2
ξ ≤ 2.546 ξ > 2.546 ξ ∗∗ ≤ 1.906 ξ ∗∗ > 1.906
Const 0.162 (0.252) 4.401 (3.252) Const 0.031 (0.325) 1.318 (12.363)
ξ 0.051 (0.140) −1.542 (1.226) ξ 0.133 (0.251) −0.782 ( 6.228)
∆Pt−1 −0.287∗∗∗ (0.036) −0.434∗∗ (0.191) ∆Pt−1 −0.323∗∗∗ (0.016) −0.163∗∗∗ ( 0.049)
∆Pt−2 −0.117∗∗∗ (0.021) −0.197 (0.181) ∆Pt−2 −0.127∗∗∗ (0.014) −0.026 ( 0.043)
∆Pt−3 −0.013 (0.019) 0.052 (0.137) ∆Pt−3 −0.055 (0.012) 0.019 ( 0.042)
(i−2) (i−2)
∆Pt −0.004 (0.009) 0.017 (0.030) ∆Pt 0.004 (0.006) −0.013 ( 0.021)
(i−1) (i−1)
∆Pt −0.002 (0.007) −0.004 (0.017) ∆Pt 0.017∗∗ (0.007) 0.018 ( 0.036)
(i+1) (i+1)
∆Pt 0.030∗∗∗ (0.008) −0.046 (0.037) ∆Pt 0.007 (0.010) −0.055 ( 0.029)
(i+2) (i+2)
∆Pt 0.004 (0.009) −0.037 (0.031) ∆Pt 0.003 (0.009) 0.011 ( 0.023)
P Auc 0.003 (0.005) 0.004 (0.017) P Auc −0.005 (0.007) 0.007 ( 0.029)
Vt −0.005∗ (0.003) −0.001 (0.006) Vt 0.006 (0.005) −0.004 ( 0.021)
∆wtn −1.743∗∗ (0.367) −0.808 (0.949) ∆wtn −0.120 (0.489) −5.117 ( 0.573)
∆wtp −0.622 (0.588) −1.936∗∗ (0.865) ∆wtp −1.352∗∗∗ (0.506) 1.639 ( 1.776)
∆snt −2.191 (1.313) 2.391 (2.441) ∆snt −1.676 (1.299) 7.783 ( 1.173)
∆spt −4.131∗∗∗ (1.094) 1.285 (1.796) ∆spt −1.059 (1.268) 0.206 ( 1.767)
∆t −0.053∗∗∗ (0.016) −0.038 (0.038) ∆t −0.010 (0.014) −0.101 ( 0.040)
#Obs 10 142 2051 #Obs 10 072 573
2
Radj 0.113 0.187 2
Radj 0.117 0.146
H18Q3 H18Q4
ξ ∗ ≤ 1.037 ξ ∗ > 1.037 ξ ≤ 0.632 ξ > 0.632
Const −1.037 (0.605) −18.093 (14.430) Const −0.864 (1.401) −2.063 (2.029)
ξ 0.725 (0.745) 17.247 (13.490) ξ 1.307 (2.499) 1.617 (3.013)
∆Pt−1 −0.318∗∗∗ (0.017) −0.264∗∗∗ ( 0.035) ∆Pt−1 −0.303∗∗∗ (0.023) −0.333∗∗∗ (0.045)
∆Pt−2 −0.144∗∗∗ (0.016) −0.112∗∗ ( 0.029) ∆Pt−2 −0.152∗∗∗ (0.020) −0.107 (0.032)
∆Pt−3 −0.043∗∗∗ (0.012) −0.031 ( 0.026) ∆Pt−3 −0.012 (0.017) −0.081 (0.030)
(i−2) (i−2)
∆Pt 0.006 (0.007) −0.021 ( 0.027) ∆Pt −0.001 (0.010) −0.014 (0.011)
(i−1) (i−1)
∆Pt 0.007 (0.008) 0.003 ( 0.016) ∆Pt 0.009 (0.010) 0.044 (0.012)
(i+1) (i+1)
∆Pt 0.005 (0.009) 0.064 ( 0.021) ∆Pt −0.002 (0.008) 0.009 (0.004)
(i+2) (i+2)
∆Pt 0.002 (0.004) 0.030 ( 0.017) ∆Pt −0.001 (0.008) 0.010 (0.010)
P Auc 0.005 (0.008) −0.003 ( 0.017) P Auc 0.004 (0.009) 0.013 (0.007)
Vt 0.008 (0.006) −0.021 ( 0.010) Vt 0.000 (0.003) 0.003 (0.003)
∆wtn −0.413 (0.389) −2.264∗∗∗ ( 0.759) ∆wtn −1.041 (0.699) −2.165 (1.214)
∆wtp −1.225∗ (0.619) −3.952∗∗ ( 1.315) ∆wtp −0.461 (1.040) −0.225 (0.349)
∆snt −4.195∗∗ (1.641) 7.669∗∗∗ ( 0.978) ∆snt 0.633 (2.239) 0.440 (3.082)
∆spt 3.485 (3.236) −12.870∗∗∗ ( 1.538) ∆spt −2.177 (1.714) −5.394 (2.870)
∆t 0.010 (0.017) 0.045 ( 0.032) ∆t −0.023 (0.020) 0.035 (0.021)
#Obs 9597 1397 #Obs 5835 7270
2
Radj 0.113 0.126 2
Radj 0.111 0.133
∗p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01
23
Acknowledgements
Florentina Paraschiv thanks the funding from Adolf Øiens Donasjonsfond Energizing New Com-
putational Frontiers [grant number L10079] and the Isaac Newton Institute for Mathematical
Sciences for its hospitality during the programme “The mathematics of energy systems” which
was supported by the Engineering and Physical Sciences Research Council (EPSRC) [grant
number EP/R014604/1]. This work has been performed within the +CityxChange (Positive
City ExChange, https://cityxchange.eu/) project under the Smart Cities and Communities
topic which was supported by the European Union’s Horizon 2020 Research and Innovation
programme [grant number 824260]. Florentina Paraschiv’s research was financially supported
by Innosuisse as part of the activities within SCCER CREST.
Appendices
A Data
We provide a summary of explanatory variables, granularities, detailed descriptions, and data
sources in Table A.1.
References
Beran, P., Pape, C., and Weber, C. (2019). Modelling German electricity wholesale spot prices
with a parsimonious fundamental model – Validation & application. Utilities Policy, 58:27–39.
Burger, M., Klar, B., Müller, A., and Schindlmayr, G. (2004). A spot market model for pricing
derivatives in electricity markets. Quantitative Finance, 4(1):109–122.
EPEX SPOT SE (2015). Press release: EPEX SPOT and ECC successfully reduce lead time
on all intraday markets. https://www.epexspot.com/sites/default/files/download_
center_files/2015-07-16_EPEX%20SPOT_Lead%20time%20reduction.pdf. (Accessed on
Jan 16, 2020).
EPEX SPOT SE (2017). Press release: Exchange Council approves the introduction of 15-
minute contracts on the Belgian and Dutch market – Trading until delivery to be launched on
the German market on 14 June 2017. https://www.epexspot.com/sites/default/files/
download_center_files/170612_EPEX%20SPOT_Exchange%20Council.pdf. (Accessed on
Jan 16, 2020).
European Energy Exchange AG (2016a). EEX annual report 2015. https://www.eex.com/

blob/16500/ff13c25a4813320ff521bf67ac397e73/eex-gb-2015-en-data.pdf. (Accessed
on Mar 1, 2019).
European Energy Exchange AG (2016b). EPEX SPOT DE 15-minute intraday auction market
data. https://www.eex.com.
European Energy Exchange AG (2016c). EPEX SPOT DE/AT 15-minute continuous intraday
market transaction data. https://www.eex.com.
European Energy Exchange AG Transparency Platform (2016). Available generation capacity

data. https://www.eex-transparency.com.
European Network of Transmission System Operators for Electricity Transparency Platform

(2016). Day-ahead total load forecast data. https://transparency.entsoe.eu.
24
Variable [Unit] Description Source
(Granularity)
Transaction price Pt Transaction price of 15-minute con- European Energy Ex-
[EUR/MWh] tracts traded on the German continu- change AG (2016c)
(1-minute) ous intraday power market at EPEX
SPOT SE
Trading volume Vt Trading volume of 15-minute contracts European Energy Ex-
[MW] traded on the German continuous in- change AG (2016c)
(1-minute) traday power market at EPEX SPOT
SE
Auction price P Auc Market clearing price of 15-minute con- European Energy Ex-
[EUR/MWh] tracts traded in the German 15-minute change AG (2016b)
(quarter-hourly) intraday auction at EPEX SPOT SE
(published: daily after 3:10 PM)
Wind power forecast wt Intradaily updated forecast of wind EWE TRADING
[GW] power generation for each quarter-hour GmbH (2016)
(quarter-hourly) on the delivery day in Germany
Solar power forecast st Intradaily updated forecast of solar EWE TRADING
[GW] power generation for each quarter-hour GmbH (2016)
(quarter-hourly) on the delivery day in Germany
Expected demand l Day-ahead total load forecast for each European Network of
[GW] quarter-hour on the delivery day in Transmission System
(quarter-hourly) Germany (published: daily at 10 AM) Operators for Elec-
tricity Transparency
Platform (2016)
Expected conventional Expected daily average of available European Energy
capacity c generation capacity of conventional Exchange AG Trans-
[GW] power plants on the delivery day in parency Platform
(daily) Germany (published: daily at 10 AM) (2016)
Table A.1: Summary of explanatory variables, granularities, detailed descriptions, and data
sources. Variables indexed by t vary over the continuous intraday trading session of
a 15-minute contract.
25
EWE TRADING GmbH (2016). Intradaily updated forecast data of wind and solar power
generation. https://www.ewe.com.
Frauendorfer, K., Paraschiv, F., and Schürle, M. (2018). Cross-border effects on Swiss electricity
prices in the light of the energy transition. Energies, 11(9):2188.
Glas, S., Kiesel, R., Kolkmann, S., Kremer, M., Graf von Luckner, N., Ostmeier, L., Urban,
K., and Weber, C. (2020). Intraday renewable electricity trading: Advanced modeling and
numerical optimal control. Journal of Mathematics in Industry, 10(1):1–17.
Graf von Luckner, N. and Kiesel, R. (2020). Modeling market order arrivals on the intraday
market for electricity deliveries in Germany with the Hawkes process. https://ssrn.com/
abstract=3526795. (Accessed on Jan 31, 2020).
Hansen, B. E. (2000). Sample splitting and threshold estimation. Econometrica, 68(3):575–603.
He, Y., Hildmann, M., Herzog, F., and Andersson, G. (2013). Modeling the merit order curve
of the European energy exchange power market in Germany. IEEE Transactions on Power
Systems, 28(3):3155–3164.
Kallabis, T., Pape, C., and Weber, C. (2016). The plunge in German electricity futures prices
– Analysis using a parsimonious fundamental model. Energy Policy, 95:280–290.
Karakatsani, N. V. and Bunn, D. W. (2010). Fundamental and behavioural drivers of electricity

price volatility. Studies in Nonlinear Dynamics & Econometrics, 14(4).
Kath, C. and Ziel, F. (2018). The value of forecasts: Quantifying the economic gains of accurate
quarter-hourly electricity price forecasts. Energy Economics, 76:411–423.
Kiesel, R. and Paraschiv, F. (2017). Econometric analysis of 15-minute intraday electricity

prices. Energy Economics, 64:77–90.
Kremer, M. (2020). thrreg: Threshold regression model. https://github.com/mlkremer/

thrreg. R package version 0.1.0.
Märkle-Huß, J., Feuerriegel, S., and Neumann, D. (2018). Contract durations in the electricity
market: Causal impact of 15 min trading on the EPEX SPOT market. Energy Economics,
69:367–378.
Pape, C., Hagemann, S., and Weber, C. (2016). Are fundamentals enough? Explaining price
variations in the German day-ahead and intraday power market. Energy Economics, 54:376–
387.
Paraschiv, F. (2013). Price dynamics in electricity markets. In Kovacevic, R., Pflug, G. C., and
Vespucci, M. T., editors, Handbook of Risk Management in Energy Production and Trading,
volume 199 of International Series in Operations Research & Management Science, pages
47–69. Springer Boston MA.
Wagner, T. (2018). emarketcrawlr: Crawling energy market data at EPEX SPOT (intra-
day trading, intraday auction, day-ahead auction). https://github.com/wagnertimo/
emarketcrawlR. R package version 0.1.0.
26
View publication stats

Paper Joint

Uploaded by

Copyright:

Available Formats

Paper Joint

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Paper Joint

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

An Econometric Model for Intraday Electricity Trading

Marcel Kremer Ruediger Kiesel

SEE PROFILE SEE PROFILE

The user has requested enhancement of the downloaded file.

Revised version forthcoming in Philosophical Transactions of the Royal Society A.

Keywords: Intraday electricity market; Econometric modeling; 15-minute con-

2.2 Hourly seasonality

2.2.2 Trading volumes

2.3 Liquidity evolution

3.1 Benchmark econometric model

3.2 Extended econometric model

3.3 Threshold regression

yi = θ 0 xi + δn0 xi (γ) + εi , i = 1, . . . , n, (4)

Sn (θ, δ, γ) = (y − X θ − Xγ δ)0 (y − X θ − Xγ δ) . (6)

γ̂ = arg min Sn (γ) , (7)

Sn (γ) = y0 y − y0 X∗γ (X∗0 ∗ −1 ∗0

4.2 Extended econometric model

4.2.1 Merit-order-curve slope

Table 4: Parameter estimates of a, b of the empirical merit-order-curve function f (`) = ea `+b

4.2.2 Model calibration

European Energy Exchange AG (2016a). EEX annual report 2015. https://www.eex.com/

European Energy Exchange AG Transparency Platform (2016). Available generation capacity

European Network of Transmission System Operators for Electricity Transparency Platform

Hansen, B. E. (2000). Sample splitting and threshold estimation. Econometrica, 68(3):575–603.

Karakatsani, N. V. and Bunn, D. W. (2010). Fundamental and behavioural drivers of electricity

Kiesel, R. and Paraschiv, F. (2017). Econometric analysis of 15-minute intraday electricity

Kremer, M. (2020). thrreg: Threshold regression model. https://github.com/mlkremer/

You might also like