Ro h, Alexande ; Schmid , Felix
A icle — Accep ed Manusc ip (Pos p in )
No only a mild win e : Ge man consume s change hei
beha io o sa e na u al gas
Joule
P o ided in Coope a ion wi h:
Ge man Ins i u e o Economic Resea ch (DIW Be lin)
Sugges ed Ci a ion: Ro h, Alexande ; Schmid , Felix (2023) : No only a mild win e : Ge man
consume s change hei beha io o sa e na u al gas, Joule, ISSN 2542-4351, Else ie , Ams e dam,
Vol. 7, Iss. 6, pp. 1081-1086,
h ps://doi.o g/10.1016/j.joule.2023.05.001
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/296667
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h ps://c ea i ecommons.o g/licenses/by-nc-nd/4.0/
No only a mild win e : Ge man consume s change hei
beha iou o sa e na u al gas
Alexande Ro ha,∗, Felix Schmid a
aGe man Ins i u e o Economic Resea ch (DIW Be lin), Moh ens aße 58, 10117
Be lin, Ge many
Keywo ds: na u al gas consump ion, ene gy sa ings, beha iou al changes
1. In oduc ion
By he s a o he 2022/2023 hea ing season, Ge many and many o he Eu opean
coun ies ound hemsel es acing a po en ial gas supply sho age in he wake o Rus-
sia’s in asion o Uk aine. In sea ch o a esponse, au ho i ies called on esiden ial
and comme cial sec o s o sa e na u al gas. Exploi ing win e 2022/23 as a “na u al
expe imen ”, we shed ligh on he magni ude o beha iou al gas sa ings using open
da a and a machine lea ning me hod. Despi e being exposed o incomple e p ice
signals, we ind signi ican beha iou al gas sa ings by Ge man households and busi-
nesses, con ibu ing o closing he supply gap. We unco e empe a u e-dependen
sa ing dynamics and discuss he po en ial oles o di e en d i e s o his change.
Finally, we highligh he pi o al ole o a imely and con inuous p o ision o openly
accessible da a and analysis o in o m he gene al public as well as policymake s.
∗Co esponding au ho
Email add ess: [email p o ec ed] (Alexande Ro h)
Manusc ip (PDF Ve sion) Click he e o iew linked Re e ences
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
This is he pos p in o he a icle No Only a Mild Win e : Ge man Consume s Change Thei
Beha io o Sa e Na u al Gas : Commen a y. Published in Joule 7 (2023), 6, S. 1081-1086,
a ailable online a : h ps://doi.o g/10.1016/j.joule.2023.05.001
© <2023>. This manusc ip e sion is made a ailable unde he CC-BY-NC-ND 4.0 license:
h ps://c ea i ecommons.o g/licenses/by-nc-nd/4.0/
2. Con ex
The Russian in asion o Uk aine in Feb ua y 2022 has c ea ed an unp eceden ed
supply c unch in Eu opean na u al gas ma ke s. Up un il Feb ua y 2022, Russia had
been Eu ope’s la ges supplie o na u al gas, expanding i s posi ion in p io yea s.
Doub ing he eliabili y o Russia’s gas supplies, he ques ion o whe he enough
gas would ha e been supplied o he Eu opean ma ke led o spi alling wholesale gas
p ices. A he end o Augus 2022, p ices peaked a o e 300 Eu o pe megawa hou
(MWh) a he benchma k hub TTF a e Russia s opped deli e ing gas h ough i s
No d S eam 1 pipeline.[1] Slowly ising in he mon hs p io o he in asion, p ices had
been luc ua ing a ound 20 Eu o pe MWh in ecen yea s .[1] Following he closu e
o No d S eam 1, he secu i y o supply was called in o ques ion wi h espec o he
upcoming win e o 2022/23.[2]
Wi hin a yea , (Cen al) Eu ope’s gas supply s uc u e changed adically. While
his o ically, a ound 40% o all gas impo ed o Ge many had been coming h ough
Russian pipelines, his numbe d opped o almos 0% by he end o 2022.[3] Much
o he Russian supply was subs i u ed by addi ional pipeline impo s om No way
and lique ied na u al gas (LNG) shipmen s om o he coun ies. The emaining
po en ial sho all ga e ise o a discussion on how much gas could and would be
sa ed by whom.
Wi h espec o gas consump ion, he e a e h ee p incipal g oups: gas- i ed
powe plan s, la ge indus ial consume s, and he esiden ial and comme cial sec o s,
which comp ise households and small- and medium-sized businesses. Gas- i ed powe
plan s consume gas o elec ici y p oduc ion, ye some also supply hea o dis ic
2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
hea ing ne wo ks. La ge indus ial consume s use gas ei he as eeds ock o sou ce
o p ocess hea . The esiden ial and comme cial sec o s need gas p edominan ly o
sa is y hea demand.
These consume g oups a e di e en in e ms o he p ice signals hey ecei e,
as well as he po en ial o and consequences o gas demand educ ions o en o ced
cu ailmen . Gas- i ed powe plan s usually buy gas sho - e m o se e peak elec-
ici y demand and hus eac immedia ely o p ice signals in bo h elec ici y and
gas ma ke s. P o ided he e is su icien al e na i e elec ici y supply, e.g. om
coal- i ed powe plan s, gas demand om he powe sec o is a he lexible. La ge
indus ial consume s, unless p o ec ed by long- e m gas supply con ac s o com-
p ehensi e hedging, a e simila ly exposed o p ice changes in he spo ma ke and
he e o e ha e an incen i e o educe gas consump ion in case o a supply c unch.
A leas in he sho un, he indus y can educe i s gas consump ion by cu bing
p oduc ion, subs i u ing he ene gy ca ie , o buying al e na i e ups eam p oduc s.
Mos ly supplied unde ixed-p ice con ac s, esiden ial and comme cial consume s
do no bea he consequences o ising p ices in he spo ma ke un il a con ac has
o be enewed. E en in he case o an acu e gas sho age, i is no clea whe he
a con olled gas cu ailmen o supply o esiden ial and comme cial sec o s in he
dis ibu ion g ids would ha e been possible, as i would ha e been challenging o
implemen o a ious echnical[4,5] and poli ical easons.
In he ace o a looming gas sho age, he public deba e ini ially concen a ed
on indus y hal ing p oduc ion, leading o a s ong economic down u n, he size o
which was deba ed con o e sially among economis s.[6,7]. To a oid di e economic
3
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
consequences o p oduc ion cu backs o indus ial consume s and because o limi ed
means o he go e nmen o impose a ioning, olun a y sa ings by esiden ial and
comme cial sec o s e en ually gained impo ance in closing he gas supply gap.
3. Gas sa ings om changes in beha iou
Since he beginning o he gas supply c unch, Ge many has been he ocus o dis-
cussion due o i s la ge economy and ela i ely high dependence on Russian gas
impo s. In Sep embe 2022, he Ge man Fede al Ne wo k Agency Bundesne zagen-
u announced ha a 20% educ ion in gas consump ion (compa ed o he a e age
consump ion o he p eceding ou yea s) would ha e been necessa y o a oid an
acu e gas sho age.[8]
In he ollowing, we aim o shed ligh on he e o s by esiden ial and comme -
cial sec o s o sa e gas. The s ong dependency o esiden ial and comme cial gas
demand on wea he condi ions implies ha ela i ely wa me o colde wea he has
a la ge e ec on whe he he a ge is ac ually achie able o no . Building on a ich
li e a u e on he ela ionship be ween hea demand, gas demand, empe a u es and
p ices[9–12], we use a e y lexible machine lea ning me hod o isola e hose gas de-
mand d i e s ha a e no go e ned by wea he a ia ions. We subsume hese d i e s
as he beha iou al componen .
The me hod used in his commen a y o es ima e sa ings is a causal o es , which
has wo impo an ea u es: (1) I is ully non-pa ame ic and da a-d i en, and (2)
i allows isola ing sa ings e ec s di e en ia ed by empe a u e. Causal o es s[13]
4
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
ex end a classic machine lea ning algo i hm, andom o es s[14]. The gene al idea o
andom o es s is o pa i ion he da a se based on alues o explana o y a iables
and i local models wi hin hese pa i ions, which a e oge he capable o ep e-
sen ing non-linea ela ionships wi hou ha ing o speci y a unc ional o m. Causal
o es s ex end his concep by using he same logic as a ool o iden i y local sa ing
e ec s. We p o ide ex ensi e explana ions, de ails, and obus ness checks o ou
model in he Supplemen al In o ma ion sec ions SI.2-SI.4. The causal o es model
enables us o p edic daily beha iou al sa ings depending on he wea he condi ions
o he day. In o de o con ol o wea he condi ions, we include mean, minimum
and maximum empe a u es o a gi en day as well as se e al lags o con ol o
he mal ine ia. I adia ion e ec s a e p oxied by sunshine du a ion, and we include
mon h and weekend/holiday indica o s o accoun o beha iou al a ia ions.
Ou model allows us o eco e wo al e na i e scena ios o es ima ed consump-
ion. The i s scena io is he es ima ed ac ual consump ion, including beha iou al
sa ings. The second scena io is he es ima ed coun e ac ual consump ion, which
would be expec ed in he absence o he sa ings. By design, he di e ence be ween
hese wo scena ios yields ou es ima e o beha iou al sa ings. By ocusing on es-
ima ed coun e ac ual consump ion and es ima ed ac ual consump ion (ins ead o
obse ed consump ion), we ensu e a like- o -like compa ison and ha ou sa ings
a e no d i en by andom e o . This assumes implici ly ha he model e o s,
gi en by he di e ence be ween he es ima ed ac ual consump ion and he obse ed
consump ion, would ha e been he same in he absence o beha iou al sa ings.
In he uppe panel o Figu e 1, he es ima ed ac ual consump ion is depic ed
5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
as a solid black line, while a solid ed line ep esen s he es ima ed coun e ac ual
consump ion (in he absence o sa ings). The dashed black line gi es he obse ed
consump ion. We s a measu ing he sa ings e ec as o Sep embe 2022, when he
isk o a supply sho age became p essing wi h he s a o he hea ing pe iod and
he end o No d S eam deli e ies. None heless, ou model allows o he possibili y
o beha iou al sa ings om he beginning o he Russian in asion o Uk aine on
24 Feb ua y 2022. We discuss he implica ions o his assump ion in de ail in he
Supplemen al In o ma ion sec ion SI.4.
Gas consump ion has been going up as expec ed wi h colde empe a u es (Figu e
1). Wi h he beginning o he hea ing season in Sep embe , we see ha Ge man es-
iden ial and comme cial sec o s ha e consis en ly sa ed be ween 66 and 285 GWh o
gas pe day. As e ealed in he lowe panel o Figu e 1, es ima ed sa ings a e s a is-
ically signi ican o all days in he Sep embe o Decembe pe iod. Decembe 2022
was excep ionally cold, also e lec ed by spiking gas demands. A ound he Ch is mas
pe iod, sa ings e o s diminished. Cumula i ely, we es ima e ha households and
comme cial sec o s ha e sa ed ca. 23 TWh [95% CI: 18.7; 27.3] by changing hei
beha iou om he beginning o Sep embe un il he end o Decembe 2022.
Relying on he esul s abo e, we can a ibu e he di e ences in gas consump ion
be ween 2022 and he a e age o he pe iod 2018-2021 o di e en e ec s (Figu e
2). The wea he e ec (g ey) is compu ed as he di e ence be ween he es ima ed
2018-2021 a e age consump ion and he es ima ed coun e ac ual consump ion in
2022. Beha iou al sa ings ( ed) esul om he di e ence be ween es ima ed ac ual
and coun e ac ual consump ion. The sum o wea he and beha iou al sa ings does
6
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
no add up o he o al di e ence in consump ion, ep esen ed by he solid line, due
o he unobse ed e o componen discussed abo e. The 20% sa ings a ge de ined
by Ge man Fede al Ne wo k Agency is e lec ed by he dashed line.
To al sa ings compa ed o he a e age o 2018-2021 a ied subs an ially be ween
di e en weeks (Figu e 2). This a ia ion is mos ly d i en by he wea he componen .
Meanwhile, he beha iou al componen emains ela i ely s able, sligh ly inc easing
o e ime. Compa ed o 2018-2021, we obse e wo cold spells: one in Sep embe (as
o calenda week 36) and one in mid-Decembe (as o calenda week 50), in which
he wea he componen d o e up gas consump ion. E en in hese colde pe iods,
es ima ed beha iou al sa ings did no change much. In he las wo weeks o he yea ,
sa ings dec eased sligh ly compa ed o he p e ious weeks. This may be explained
ei he by he Ch is mas pe iod o by a educed u gency, as i became inc easingly
e iden by Decembe ha a gas sho age in he win e o 2022/23 would be a he
unlikely. Gas s o age le els emained well abo e he ange o p e ious yea s.
On agg ega e, we ind ha he wea he e ec alone did no play a signi ican ole
when compa ing he Sep embe o Decembe 2022 gas consump ion wi h p e ious
yea s ( igh panel o Figu e 2). A leas o he i s hal o he win e , his is possibly
a odds wi h o he analyses asse ing ha a compa ably mild win e induced mos
sa ings.[15] Consis en beha iou al sa ings con as highly a iable wea he - ela ed
sa ings. Especially he cold spell in Decembe o se mos o he sa ings by wea he
due o milde empe a u es in he weeks be o e. Howe e , he wea he may ha e had
an indi ec e ec , as a colde win e would ha e made i e en ha de o households
o sa e gas in he same way.
7
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
The win e mon hs o 2022 also shed ligh on he sa ings dynamics o he esiden-
ial and comme cial sec o s ela i e o empe a u es. We ind a nega i e ela ionship
be ween ela i e gas sa ings, de ined as absolu e gas sa ings di ided by es ima ed
coun e ac ual consump ion, and empe a u e (lowe panel, Figu e 3). The esiden-
ial and comme cial sec o s seem o ela i ely easily supp ess hei hea ing demand
when empe a u es a e a he mild. These le els o ela i e sa ings canno be ca -
ied o e o lowe empe a u es. I ou side empe a u es a e a ound 12°C, dec easing
hea ing e o s by a ce ain amoun will ha e a much lowe e ec on oom empe a-
u es compa ed o a si ua ion when ou side empe a u es ange a ound 0°C.
Rega ding he ele ance o a e ing a gas sho age, ela i e sa ings a e, howe e ,
only o mino impo ance. The e o e, we highligh he subs an ial and consis en
absolu e sa ings du ing cold empe a u e days (uppe panel o Figu e 3). Al hough
hey ell sho o he a ge ed 20% goal by he ede al egula o , hey added mo e o
ad e ing a gas sho age han he highe ela i e sa ings in au umn.
4. Conclusions and ou look
Win e 2022/23 happened o be a “na u al expe imen ” o Eu ope and Ge many on
how he economy would eac o a gas supply c unch o e en a looming sho age. I
es ed he capaci y and willingness o households and comme cial consume s o cu
gas demand mainly used o hea ing. Using a da a-d i en causal o es model, we can
show ha esiden ial and comme cial sec o s ha e educed hei gas consump ion.
In con as , he wea he had e en an inc easing e ec .
8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
1000
2000
3000
Daily gas consump ion, GWh
Exp. ac ual consump ion Exp. coun e ac ual consump ion Obse ed consump ion
(A)
-300
-200
-100
0
35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 01
Calenda week 2022
Beha iou al sa ings, GWh
CI lowe CI uppe P edic ions
(B)
Figu e 1 main body Click he e o access/download;Figu e;Figu e 1.pd
Desi ed sa ings
Obse ed sa ings
-6
-3
0
3
6
35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52
Calenda week 2022
Change in gas consump ion, TWh
(A)
Wea he :
2 TWh
Wea he :
2 TWh
Beha iou :
-23 TWh
Beha iou :
-23 TWh
-60
-30
0
30
60
To al
Beha iou
Wea he
(B)
Figu e 2 main body Click he e o access/download;Figu e;Figu e 2.pd
(B) Rela i e (%)
(A) Absolu e (GWh)
0 10 20
-250
-200
-150
-100
-25
-20
-15
-10
Mean empe a u e, °C
Beha iou al sa ings
Sep
Oc
No
Dec
Figu e 3 main body Click he e o access/download;Figu e;Figu e 3.pd
SI. Supplemen al In o ma ion
This sec ion p o ides de ails on he deployed me hodology and p esen s supplemen al
esul s. Sec ion SI.1 discusses da a sou ces and p o ides some summa y s a is ics.
Sec ion SI.2 gi es a me hodological in oduc ion o local linea causal ees. Sec ion
SI.3 con as s i wi h al e na i e app oaches. Sec ion SI.4 p o ides supplemen al
esul s on obus ness.
SI.1. Da a and desc ip i e s a is ics
SI.1.1. Gas consump ion da a
Ins an aneous gas consump ion me e ing o esiden ial and comme cial cus ome s
is s ill a e in Ge many, such ha accu a e day-by-day consump ion p o iles o
indi idual households o business uni s a e una ailable [1]. In he absence o di ec ly
me e ed da a, he Ge man Ne wo k Agency elies on esidual load da a published by
he Ge man gas exchange T ading Hub Eu ope (THE). The esidual load is de i ed
by aking he di e ence be ween gas in lows and gas ou lows om he ne wo k o
downs eam ne wo ks, s o ages, o he coun ies, o la ge-scale cus ome s [2]. These
da a a e by design o he whole Ge man ma ke a ea and hence ou analysis canno
ake in o accoun any spa ial di e en ia ion be ween consump ion pa e ns. Ve y
ecen da a a e subjec o e isions, and inal da a o a gi en da e a e only a ailable
a e ca. 1.5 mon hs.1The e o e, a he ime o w i ing, he las a ailable mon h
o inal da a is Decembe 2022. The publicly a ailable da ase includes he yea s
2018-2022.
1THE publishes inal da a acco ding o ‘M+2M-10WD’, which means ha inal da a o he
cu en mon h ‘M’ a e published wo mon hs la e (‘+2M’) minus 10 wo king days (‘-10WD’). This
in o ma ion is p o ided in a da a Excel ile a ailable a www. adinghub.eu/en-gb/Publica ions/
T anspa ency/Agg ega ed-consump ion-da a. Hence, o Decembe 2022, inal da a ha e been
a ailable since he 15 h o Feb ua y.
I
Supplemen al Tex and Figu es
SI.1.2. Wea he da a and o he con ols
Residen ial and comme cial gas demand is hea ily d i en by hea ing demand in-
ducing a high dependence on ou side ai empe a u es and o he wea he a iables.
Ge many’s Na ional Me eo ological Se ice (DWD) publishes dozens o wea he pa-
ame e s o hund eds o wea he s a ions daily. They a e a ailable h ough an ap-
plica ion p og amming in e ace (API) ha pe mi s downloading speci ic da a wi h
cus om p og amming sc ip s. We implemen ed ou download ou ine in Py hon (see
Sec ion SI.5 below). While he model desc ibed in he nex sec ion could po en ially
deal wi h a la ge numbe o co a ia es by means o egula isa ion me hods such as
a leas absolu e sh inkage and selec ion ope a o (LASSO), a egula ised eg ession
me hod ha cons ains he L1no m o he coe icien ec o helping o selec only
impo an eg esso s, we es ic ou sel es o a concise se ha accoun s o a e y
la ge sha e o he gas demand a ia ion in he con ol pe iod.
Fo each day and e e y wea he s a ion, we access he a e age empe a u e, as
well as he maximum and minimum empe a u es. The la e accoun o ex eme
empe a u e changes du ing a single day. To con ol o he mal ine ia, h ee lags
o a e age, minimum and maximum empe a u es a e added o he model2Sola
i adia ion migh be conduci e o hea ing demand educ ions no only h ough i s
e ec on ai empe a u es. We p oxy sola i adia ion by he sunshine du a ion pe
day in hou s. As discussed in he p e ious subsec ion, he gas demand da a is only
a ailable a a na ional le el. Hence, we need o agg ega e he co a ia es spa ially.
O he s udies use popula ion-weigh ing o a e age ac oss spa ially disagg ega ed e-
analysis da a, a blend o his o ical da a poin s and model ou pu s [3]. Fo simplici y,
we choose o ake he median ac oss wea he s a ions. We p e e he median o e
a simple a e age so as o no in oduce biases om ex eme obse a ions, such as
measu emen s om Ge many’s highes moun ain Zugspi ze. Las ly, we include ixed
e ec s o mon hs and weekends as well as na ional holidays.
2The Ge man building s ock equipped wi h gas- i ed hea ing has a ying deg ees o insula ion.
By allowing he model o choose he ela i e impo ance o empe a u es on p eceding days non-
pa ame ically, i accoun s lexibly o he a e age impac o insula ion on gas demand.
II
SI.1.3. Summa y s a is ics
In Table SI.I, we p esen a ew key s a is ics o ou da a se . We dis inguish be ween
2018-2021, he business-as-usual pe iod, and 2022, he yea subjec o beha iou al
sa ings. I is e iden om he op panel ha a e age gas consump ion is somewha
lowe in 2022 compa ed o p e ious yea s. Fo he wea he a iables in he ollowing
panels, he s a is ics a e qui e close o each o he , indica ing good o e lap, a key
equi emen o he alidi y o he me hod discussed in he nex sec ion [4]. Fo
exposi ion, we plo he mean empe a u e in he Sep embe o Decembe 2022 pe iod
agains he mean, minimum and maximum empe a u es o he same pe iod in 2018-
2021 in Figu e SI.I
Table SI.I: Selec ed summa y s a is ics
Va iable S a is ic 2018-2021 2022
a g 1088.44 966.66
Gas consump ion min 165.35 162.93
max 3273.74 2668.29
s d 764.83 726.95
a g 10.18 10.61
Mean empe a u e min -9.6 -6
max 27.2 26.2
s d 7.06 7.03
a g 5.63 5.83
Minimum empe a u e min -13.7 -10
max 18.2 16.95
s d 5.99 5.87
a g 14.8 15.43
Maximum empe a u e min -6 -3
max 35.5 35.6
s d 8.48 8.46
a g 4.96 5.54
Sunshine du a ion min 0 0
max 15.23 14.65
s d 4.33 4.38
III
Figu e SI.I: Mean empe a u es
0
10
20
Sep Oc No Dec Jan
2022
Mean daily empe a u e, °C
2022
A e age 2018−21
Range 2018−21
Figu e SI.II shows he ela ionship be ween gas consump ion and he daily mean
empe a u e di e en ia ed by calenda mon h (colou ) and pe iod (ma ke shape).
The igu e demons a es he non-linea ela ionship be ween mean empe a u e and
gas consump ion.
SI.2. Model desc ip ion
As e iden om Figu e SI.II, he ela ionship be ween wea he a iables and gas con-
sump ion is non-linea . T adi ional me hods, such as hea ing deg ee day co ec ions
o pa ame ic polynomial models, may in oduce biases, especially a he bounda y
o he suppo .
We deploy a ully da a-d i en, non-pa ame ic app oach ha can no only deal
wi h non-linea i ies in he ela ionship be ween co a ia es and gas consump ion bu
also wi h he e ogenei y in beha iou al sa ings condi ional on he co a ia es, such as
empe a u e o mon h. Non-pa ame ic models do no equi e he o mula ion o a
unc ional ela ionship be ween ele an ac o s, he co a ia es, and he a iable o
IV
1000
2000
3000
−10 0 10 20
Mean Tempe a u e, °C
Gas consump ion, GWh
Mon h
Jan
Feb
Ma
Ap
May
Jun
Jul
Aug
Sep
Oc
No
Dec
Pe iod
2018−2021
2022
Figu e SI.II: Rela ionship be ween mean empe a u e and gas consump ion by mon h and pe iod
in e es . Causal o es s pionee ed by [4], and e ined wi h doubly- obus echniques
in [5], ex end a classical machine lea ning me hod, andom o es s [6]. We p o ide
sho explana ions o hese e ms below.
SI.2.1. Random o es s
Random o es s p edic a a iable o in e es condi ional on a se o co a ia es by
a e aging o e he p edic ions o a po en ially la ge numbe o decision ees. A
decision ee spli s he da a se in o subse s, o neighbou hoods, in he co a ia e
domain. In ou case, a simple ee could i s di ide he da a se depending on
whe he a gi en obse a ion has a mean empe a u e abo e o below 10 deg ees
Celsius. In he below 10 deg ees subse , he nex spli could be based on whe he an
obse a ion is om a calenda mon h be o e Ma ch o no . A comple ely di e en
spli could di ide he abo e 10 deg ees subse . Fu he spli s may ollow. The
inal subse s, o neighbou hoods, a e called lea es. Fo each lea , he andom o es
algo i hm i s a local model. In he classic implemen a ion, his local model is a
simple a e age o he a iable o in e es o all obse a ions wi hin his lea . The
mo e e ined e sion used below i s a local linea model ins ead. The algo i hm
V
selec s spli ing ules in o de o minimise some p edic ion e o me ic, such as he
mean squa ed e o . Taken oge he , he collec ion o local models can ep esen
complica ed non-linea ela ionships wi hou ha ing o speci y a unc ional o m. As
shown [6] ha he a e age o a la ge numbe o decision ees es ima ed on boo s ap
samples imp o es he p edic i e powe , a o es usually consis s o a leas a ew
hund ed decision ees.
SI.2.2. Causal in e ence and causal o es s
Causal o es s use andom o es s o he p edic ion o ea men e ec s in a po en ial
ou comes amewo k (e.g. [7]). The undamen al p oblem o causal in e ence is ha
we canno obse e wha would ha e happened o a ea ed uni in he absence o he
ea men [8]. In ou case, he ea men co esponds o all ac o s discussed abo e
ega ding he looming supply c unch, assuming o s a a e 23/02/2022. In o de
o iden i y he ea men e ec , i.e. he beha iou al sa ings, modelle s ha e di e en
op ions anging om s uc u al models o andomised expe imen s. Obse a ional
s udies, like he one a hand, aim o emula e he andomisa ion o an expe imen , e.g.
by con olling o all ac o s ha a ec he p opensi y o being ea ed. P o ided we
can obse e all such ac o s, he ea men assignmen condi ional on he co a ia es
becomes as good as andom. Two me hods (o many) o con olling o he co a ia es
a ec ing ea men selec ion a e eg ession and in e se p opensi y sco e weigh ing
(IPW). Combining he wo leads o he class o doubly obus es ima o s ha ha e
he ad an age o eco e ing he ea men e ec e en i only one o he wo me hods
is co ec ly speci ied.
Much like a andom o es , a causal o es spli s he da a se based on ules
e e ing o co a ia e alues. Howe e , he objec i e sough o op imise by selec ing
he spli s is di e en . We do no ha e da a on he ue ea men e ec such ha
we canno op imise a p edic ion e o me ic. Ins ead, he causal o es algo i hm
aims o de e mine neighbou hoods in he co a ia e domain in such a way ha he
es ima ed ea men e ec s a e as simila as possible wi hin a neighbou hood and as
dissimila as possible be ween neighbou hoods [9]. The condi ional a e age ea men
VI
e ec s a e especially use ul in a con ex like ou s whe e he magni ude o beha iou al
sa ings is expec ed o a y signi ican ly by wea he condi ions.
SI.2.3. Model o mula ion and es ima ion
Le Y (1) be he gas consump ion in pe iod in he p esence o beha iou al sa ings
and Y (0) be he gas consump ion in he same pe iod in he absence o beha iou al
sa ings. Consequen ly, he obse ed consump ion can be exp essed by:
Y =W Y (1) + (1 −W )Y (0)
whe e W ∈ {0,1}indica es he p esence o beha iou al sa ings. In ou base case,
W = 1 o all ≥24/02/2022. We a e in e es ed in he e ec o beha iou al sa ings
condi ional on co a ia es X de ined by:
τ(x) = E[Y (1) −Y (0)|X =x] (A.1)
Ye , Y (1) and Y (0) a e no obse able a he same ime, such ha he unc ion
τ(x) is no di ec ly iden i iable.
We assume s ic exogenei y condi ional on he co a ia es X , i.e. he e a e no
unobse ed con ounde s o W and Y and a e con olling o he co a ia es he
ea men assignmen is as good as andom.
{Y (1), Y (0)} ⊥ W |X
We u he assume ha esiden ial and comme cial gas consump ion Y on a day
ollows he ollowing pa ially linea model:
Y =τ(X )W + (X ) + ε (A.2)
The e ec o beha iou al sa ings on consump ion is measu ed by a unc ion τ(·),
which may depend on he co a ia es X . (X ) is a po en ially complica ed unc-
ion o he co a ia es and ε is an independen ly dis ibu ed e o e m. Double
VII
da e by e- unning ou model wi h a mon hly sequence o s a da es beginning on
24 Sep embe 2021 and ending on 24 Augus 2022. Fo each i e a ion, we compu e
he o al cumula i e p edic ed sa ings in he pe iod om 1 Sep embe 2022 un il 31
Decembe 2022.
Figu e SI.VI: S a da e sensi i i y
−25
−20
−15
−10
−5
Oc 2021 Jan 2022 Ap 2022 Jul 2022
Assumed s a da e
Cum. sa ings Sep−Dec '22, TWh
Figu e SI.VI shows ha he esul s a e e y obus o a ia ions o he sa ings
s a da e in 2022. Fo s a da es in 2021, howe e , cumula i e es ima ed sa ings
om Sep embe 2022 o Decembe 2022 decline apidly, sugges ing ha households
and comme cial sec o s did no eac o he o eboding de elopmen s in wholesale
ma ke s a he ime.
SI.4.4. COVID-19
The p e-c isis pe iod o ou da a se includes he COVID-19 pandemic. As households
p ac ised social dis ancing, wo ked om home, and shops and o ices emained closed,
he hea ing beha iou o esiden ial and comme cial sec o s is likely o ha e changed
compa ed o he p e-COVID pe iod. Suppose ha ex ended pe iods o isola ion a
XIV
home ha e led o mo e gas consump ion, ce e is pa ibus, e en o se ing he educed
demand by comme cial buildings. I his we e ue, ou model may deli e biased
esul s as i exagge a es coun e ac ual gas demand compa ed o wha would ha e
been expec ed, as he wo ld has gone back o no mal in 2022, bu o he gas c isis
due o he Russian in asion o Uk aine.
The e o e, we conduc a sensi i i y es in o de o de e mine i ou esul s hinge
on a po en ial exagge a ion o sa ings due o lockdowns. In Ge many, he i s
lockdown s a ed on 22 Ma ch 2020 and ended on 4 May 2020. A second lockdown
began wi h ligh e es ic ions on 2 No embe 2020. By Janua y, igh e es ic ions
we e imposed and he lockdown was no li ed be o e 9 May 2021. Le L⊂Cbe
he se o lockdown days in ou p e-c isis da a se . In he i s s ep, we es ima e ou
model o e he se T ∪ C L. Fu he , we de ine a b oade se o pandemic days
ha comp ises all days be ween 1 Ma ch 2020 and 31 Decembe 2021. Le his se
be deno ed by P. In a second s ep, we compu e ou model es ima es o e he se
T ∪ C P.
Table SI.III: COVID-19 sensi i i y
Scena io Es . cum. beha iou al sa ings Change
Baseline 23.0 TWh
Excl. lockdown days (L) 22.7 TWh -1.58%
Excl. all pandemic days (P) 21.5 TWh -7.09%
As shown in Table SI.III, he e ec o excluding he lockdown pe iod is negligible.
The e ec o excluding he ull pandemic pe iod is a bi la ge a ca. 7%. Howe e ,
we do no hink i is easonable o exclude his pe iod en i ely. While i is e y likely
ha hea ing beha iou s ha e changed du ing he pandemic, i is also p obable ha
a leas a pa o hose changes con inues o ake e ec oday, e.g. due o lexible
wo king- om-home policies. We conclude ha ou model is no subs an ially biased
by he inclusion o he COVID-19 pe iod in he con ol se C.
XV
SI.5. Code
We w o e a Py hon sc ip o gas consump ion da a downloads and Deu sche We -
e diens API calls. All modelling s eps and cha ing we e conduc ed in R. We make
all code a ailable in his eposi o y: gi lab.com/diw-e u/p ojec s/gas-sa ings.
XVI
Re e ences
[1] ACER, Annual Repo on he Resul s o Moni o ing he In e nal Elec ici y
and Na u al Gas Ma ke s in 2021 (Oc . 2022).
URL www.ace .eu opa.eu/si es/de aul / iles/documen s/
Publica ions/MMR_2021_Ene gy_Re ail_Consume _P o ec ion_Volume.
pd
[2] BDEW, BDEW/VKU/GEODE Bes P ac ice Guidelines ma ke p ocesses o
he managemen o gas balancing g oups - pa 1 (Ma . 2021).
URL www. adinghub.eu/Po als/0/Downloadcen e %20-%
20Koope a ions e einba ung%20und%20Lei %C3%A4den/KoV%20englisch/
KoV_XII_LF_BKM_Gas_Teil1_EN.pd ? e =iImXwzb JXgHmEwqSMHgQ%3D%3D
[3] O. Ruhnau, C. S iewe, J. Muessel, L. Hi h, Gas demand in imes o c isis:
ene gy sa ings by consume g oup in Ge many, ZBW - Leibniz In o ma ion
Cen e o Economics P ep in (261082) (2022).
URL h ps://ideas. epec.o g/p/zbw/esp ep/261082.h ml
[4] S. Wage , S. A hey, Es ima ion and In e ence o He e ogeneous T ea men E -
ec s using Random Fo es s, Jou nal o he Ame ican S a is ical Associa ion
113 (523) (2018) 1228–1242. doi:10.1080/01621459.2017.1319839.
[5] S. A hey, J. Tibshi ani, S. Wage , Gene alized andom o es s, The Annals o
S a is ics 47 (2) (Ap . 2019). doi:10.1214/18-AOS1709.
[6] L. B eiman, Random Fo es s, Machine Lea ning 45 (1) (2001) 5–32. doi:10.
1023/A:1010933404324.
[7] D. B. Rubin, Causal In e ence Using Po en ial Ou comes, Jou nal o he
Ame ican S a is ical Associa ion 100 (469) (2005) 322–331. doi:10.1198/
016214504000001880.
[8] P. W. Holland, S a is ics and Causal In e ence, Jou nal o he Ame ican S a is-
ical Associa ion 81 (396) (1986) 945–960. doi:10.2307/2289064.
XVII
[9] J. Tibshi ani, S. A hey, E. S e d up, S. Wage , A g guided ou .
URL www.g -labs.gi hub.io/g /a icles/g _guide.h ml
[10] R. F iedbe g, J. Tibshi ani, S. A hey, S. Wage , Local Linea Fo es s, Jou nal o
Compu a ional and G aphical S a is ics 30 (2) (2021) 503–517. doi:10.1080/
10618600.2020.1831930.
[11] S. A hey, G. W. Imbens, The S a e o Applied Econome ics: Causali y and
Policy E alua ion, Jou nal o Economic Pe spec i es 31 (2) (2017) 3–32. doi:
10.1257/jep.31.2.3.
XVIII
