(with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads ›...
Transcript of (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads ›...
![Page 1: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/1.jpg)
Total
A
AA AB AC
B
BA BB BC
C
CA CB CC
1
George Athanasopoulos(with Rob J. Hyndman, Nikos Kourentzes andFotis Petropoulos)
Forecasting hierarchical (andgrouped) time series
![Page 2: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/2.jpg)
Outline
1 Hierarchical and grouped time series
2 Optimal forecasts
3 Approximately optimal forecasts
4 Temporal hierarchies
5 References
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 2
![Page 3: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/3.jpg)
Summary
3 key developments:
1 We generalise the forecasting process in"properly" accounting for grouped data.(Empirical Application 1).
2 We advance the “optimal combination”approach by proposing two new estimatorsbased on WLS.Both now implemented in the hts package.
3 We introduce temporal hierarchies.(Empirical Application 2).
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 3
![Page 4: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/4.jpg)
Summary
3 key developments:
1 We generalise the forecasting process in"properly" accounting for grouped data.(Empirical Application 1).
2 We advance the “optimal combination”approach by proposing two new estimatorsbased on WLS.Both now implemented in the hts package.
3 We introduce temporal hierarchies.(Empirical Application 2).
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 3
![Page 5: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/5.jpg)
Summary
3 key developments:
1 We generalise the forecasting process in"properly" accounting for grouped data.(Empirical Application 1).
2 We advance the “optimal combination”approach by proposing two new estimatorsbased on WLS.Both now implemented in the hts package.
3 We introduce temporal hierarchies.(Empirical Application 2).
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 3
![Page 6: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/6.jpg)
Summary
3 key developments:
1 We generalise the forecasting process in"properly" accounting for grouped data.(Empirical Application 1).
2 We advance the “optimal combination”approach by proposing two new estimatorsbased on WLS.Both now implemented in the hts package.
3 We introduce temporal hierarchies.(Empirical Application 2).
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 3
![Page 7: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/7.jpg)
Hierarchical versus grouped
Table: Geographical Hierarchy
Level Total series per levelAustralia 1States and Territories 7
VIC, NSW, QLD, SA, WA, NT, TAS;
Zones 27VIC (5): Metro, West Coast, East Coast, Nth East, Nth West;
NSW (6): Metro, Nth Coast, Sth Coast, Sth, Nth, ACT;
QLD (4): Metro, Central Coast, Nth Coast, Inland;
Regions 76Metro VIC: Melbourne, Peninsula, Geelong;
Metro NSW: Sydney, Illawarra, Central Coast;
Metro QLD: Brisbane, Gold Coast, Sunshine Coast;
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 4
![Page 8: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/8.jpg)
Australian domestic tourism
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 5
Hierarchical:Australia (1)
States (7)
Zones (27)
Regions (76)
Total: 111 series
![Page 9: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/9.jpg)
Australian domestic tourism
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 5
Purpose of travel (PoT):Holiday
Visiting Friends and Relatives
Business
Other
![Page 10: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/10.jpg)
Australian domestic tourism
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 5
Hierarchical:Australia (1)
States (7)
Zones (27)
Regions (76)
Total: 111 series
PoT (×4):AustraliaPoT (4)
StatesPoT (28)
ZonesPoT (108)
RegionsPoT (304)
Total: 444 series
![Page 11: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/11.jpg)
Australian domestic tourism
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 5
Hierarchical:Australia (1)
States (7)
Zones (27)
Regions (76)
Total: 111 series
PoT (×4):AustraliaPoT (4)
StatesPoT (28)
ZonesPoT (108)
RegionsPoT (304)
Total: 444 series
GroupedGrand total: 555 series
![Page 12: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/12.jpg)
Forecasting such structures
Existing methods:
ã Bottom-up
ã Top-down
ã Middle-out
Key idea: forecast reconciliationå Ignore structural constraints and forecast
every series of interest independently.
å Adjust forecasts to impose constraints.
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 6
![Page 13: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/13.jpg)
Forecasting such structures
Existing methods:
ã Bottom-up
ã Top-down
ã Middle-out
Key idea: forecast reconciliationå Ignore structural constraints and forecast
every series of interest independently.
å Adjust forecasts to impose constraints.
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 6
![Page 14: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/14.jpg)
Hierarchical data
Total
A B C
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 7
Yt : observed aggregate of allseries at time t.
YX,t : observation on series X attime t.
Bt : vector of all series atbottom level in time t.
![Page 15: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/15.jpg)
Hierarchical data
Total
A B C
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 7
Yt : observed aggregate of allseries at time t.
YX,t : observation on series X attime t.
Bt : vector of all series atbottom level in time t.
![Page 16: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/16.jpg)
Hierarchical data
Total
A B C
Yt =
YtYA,tYB,tYC,t
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 7
Yt : observed aggregate of allseries at time t.
YX,t : observation on series X attime t.
Bt : vector of all series atbottom level in time t.
![Page 17: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/17.jpg)
Hierarchical data
Total
A B C
Yt =
YtYA,tYB,tYC,t
=
1 1 11 0 00 1 00 0 1
︸ ︷︷ ︸
S
YA,tYB,tYC,t
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 7
Yt : observed aggregate of allseries at time t.
YX,t : observation on series X attime t.
Bt : vector of all series atbottom level in time t.
![Page 18: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/18.jpg)
Hierarchical data
Total
A B C
Yt =
YtYA,tYB,tYC,t
=
1 1 11 0 00 1 00 0 1
︸ ︷︷ ︸
S
YA,tYB,tYC,t
︸ ︷︷ ︸
Bt
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 7
Yt : observed aggregate of allseries at time t.
YX,t : observation on series X attime t.
Bt : vector of all series atbottom level in time t.
![Page 19: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/19.jpg)
Hierarchical data
Total
A B C
Yt =
YtYA,tYB,tYC,t
=
1 1 11 0 00 1 00 0 1
︸ ︷︷ ︸
S
YA,tYB,tYC,t
︸ ︷︷ ︸
BtYt = SBt
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 7
Yt : observed aggregate of allseries at time t.
YX,t : observation on series X attime t.
Bt : vector of all series atbottom level in time t.
![Page 20: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/20.jpg)
Hierarchical dataTotal
A
AX AY AZ
B
BX BY
C
CX CY
Yt =
YtYA,tYB,tYC,tYAX,tYAY,tYAZ,tYBX,tYBY,tYCX,tYCY,t
=
1 1 1 1 1 1 11 1 1 0 0 0 00 0 0 1 1 0 00 0 0 0 0 1 11 0 0 0 0 0 00 1 0 0 0 0 00 0 1 0 0 0 00 0 0 1 0 0 00 0 0 0 1 0 00 0 0 0 0 1 00 0 0 0 0 0 1
︸ ︷︷ ︸
S
YAX,tYAY,tYAZ,tYBX,tYBY,tYCX,tYCY,t
︸ ︷︷ ︸
Bt
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 8
![Page 21: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/21.jpg)
Hierarchical dataTotal
A
AX AY AZ
B
BX BY
C
CX CY
Yt =
YtYA,tYB,tYC,tYAX,tYAY,tYAZ,tYBX,tYBY,tYCX,tYCY,t
=
1 1 1 1 1 1 11 1 1 0 0 0 00 0 0 1 1 0 00 0 0 0 0 1 11 0 0 0 0 0 00 1 0 0 0 0 00 0 1 0 0 0 00 0 0 1 0 0 00 0 0 0 1 0 00 0 0 0 0 1 00 0 0 0 0 0 1
︸ ︷︷ ︸
S
YAX,tYAY,tYAZ,tYBX,tYBY,tYCX,tYCY,t
︸ ︷︷ ︸
Bt
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 8
![Page 22: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/22.jpg)
Hierarchical dataTotal
A
AX AY AZ
B
BX BY
C
CX CY
Yt =
YtYA,tYB,tYC,tYAX,tYAY,tYAZ,tYBX,tYBY,tYCX,tYCY,t
=
1 1 1 1 1 1 11 1 1 0 0 0 00 0 0 1 1 0 00 0 0 0 0 1 11 0 0 0 0 0 00 1 0 0 0 0 00 0 1 0 0 0 00 0 0 1 0 0 00 0 0 0 1 0 00 0 0 0 0 1 00 0 0 0 0 0 1
︸ ︷︷ ︸
S
YAX,tYAY,tYAZ,tYBX,tYBY,tYCX,tYCY,t
︸ ︷︷ ︸
Bt
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 8
Yt = SBt
![Page 23: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/23.jpg)
Grouped dataTotal
A
AX AY
B
BX BY
Total
X
AX BX
Y
AY BY
Yt =
YtYA,tYB,tYX,tYY,tYAX,tYAY,tYBX,tYBY,t
=
1 1 1 11 1 0 00 0 1 11 0 1 00 1 0 11 0 0 00 1 0 00 0 1 00 0 0 1
︸ ︷︷ ︸
S
YAX,tYAY,tYBX,tYBY,t
︸ ︷︷ ︸
Bt
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 9
![Page 24: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/24.jpg)
Grouped dataTotal
A
AX AY
B
BX BY
Total
X
AX BX
Y
AY BY
Yt =
YtYA,tYB,tYX,tYY,tYAX,tYAY,tYBX,tYBY,t
=
1 1 1 11 1 0 00 0 1 11 0 1 00 1 0 11 0 0 00 1 0 00 0 1 00 0 0 1
︸ ︷︷ ︸
S
YAX,tYAY,tYBX,tYBY,t
︸ ︷︷ ︸
Bt
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 9
![Page 25: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/25.jpg)
Grouped dataTotal
A
AX AY
B
BX BY
Total
X
AX BX
Y
AY BY
Yt =
YtYA,tYB,tYX,tYY,tYAX,tYAY,tYBX,tYBY,t
=
1 1 1 11 1 0 00 0 1 11 0 1 00 1 0 11 0 0 00 1 0 00 0 1 00 0 0 1
︸ ︷︷ ︸
S
YAX,tYAY,tYBX,tYBY,t
︸ ︷︷ ︸
Bt
Forecasting hierarchical (and grouped) time series Hierarchical and grouped time series 9
Yt = SBt
![Page 26: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/26.jpg)
Outline
1 Hierarchical and grouped time series
2 Optimal forecasts
3 Approximately optimal forecasts
4 Temporal hierarchies
5 References
Forecasting hierarchical (and grouped) time series Optimal forecasts 10
![Page 27: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/27.jpg)
Optimal forecasts
Key idea: forecast reconciliationå Ignore structural constraints and forecast
every series of interest independently.
å Adjust forecasts to impose constraints.
Let Yn(h) be vector of initial h-step forecasts,made at time n, stacked in same order as Yt.
Yt = SBt . So Yn(h) = Sβn(h) + εh .
βn(h) = E[Bn+h | Y1, . . . ,Yn].εh has zero mean and covariance Σh.Estimate βn(h) using GLS?
Forecasting hierarchical (and grouped) time series Optimal forecasts 11
![Page 28: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/28.jpg)
Optimal forecasts
Key idea: forecast reconciliationå Ignore structural constraints and forecast
every series of interest independently.
å Adjust forecasts to impose constraints.
Let Yn(h) be vector of initial h-step forecasts,made at time n, stacked in same order as Yt.
Yt = SBt . So Yn(h) = Sβn(h) + εh .
βn(h) = E[Bn+h | Y1, . . . ,Yn].εh has zero mean and covariance Σh.Estimate βn(h) using GLS?
Forecasting hierarchical (and grouped) time series Optimal forecasts 11
![Page 29: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/29.jpg)
Optimal forecasts
Key idea: forecast reconciliationå Ignore structural constraints and forecast
every series of interest independently.
å Adjust forecasts to impose constraints.
Let Yn(h) be vector of initial h-step forecasts,made at time n, stacked in same order as Yt.
Yt = SBt . So Yn(h) = Sβn(h) + εh .
βn(h) = E[Bn+h | Y1, . . . ,Yn].εh has zero mean and covariance Σh.Estimate βn(h) using GLS?
Forecasting hierarchical (and grouped) time series Optimal forecasts 11
![Page 30: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/30.jpg)
Optimal forecasts
Key idea: forecast reconciliationå Ignore structural constraints and forecast
every series of interest independently.
å Adjust forecasts to impose constraints.
Let Yn(h) be vector of initial h-step forecasts,made at time n, stacked in same order as Yt.
Yt = SBt . So Yn(h) = Sβn(h) + εh .
βn(h) = E[Bn+h | Y1, . . . ,Yn].εh has zero mean and covariance Σh.Estimate βn(h) using GLS?
Forecasting hierarchical (and grouped) time series Optimal forecasts 11
![Page 31: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/31.jpg)
Optimal forecasts
Key idea: forecast reconciliationå Ignore structural constraints and forecast
every series of interest independently.
å Adjust forecasts to impose constraints.
Let Yn(h) be vector of initial h-step forecasts,made at time n, stacked in same order as Yt.
Yt = SBt . So Yn(h) = Sβn(h) + εh .
βn(h) = E[Bn+h | Y1, . . . ,Yn].εh has zero mean and covariance Σh.Estimate βn(h) using GLS?
Forecasting hierarchical (and grouped) time series Optimal forecasts 11
![Page 32: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/32.jpg)
Optimal forecasts
Key idea: forecast reconciliationå Ignore structural constraints and forecast
every series of interest independently.
å Adjust forecasts to impose constraints.
Let Yn(h) be vector of initial h-step forecasts,made at time n, stacked in same order as Yt.
Yt = SBt . So Yn(h) = Sβn(h) + εh .
βn(h) = E[Bn+h | Y1, . . . ,Yn].εh has zero mean and covariance Σh.Estimate βn(h) using GLS?
Forecasting hierarchical (and grouped) time series Optimal forecasts 11
![Page 33: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/33.jpg)
Optimal combination forecasts
Yn(h) = Sβn(h) = S(S′Σ†hS)−1S′Σ†hYn(h)
Optimal P = (S′Σ†hS)−1S′Σ†h.
Σ†h is generalized inverse of Σh.
Revised forecasts unbiased: SPS = S.
Revised forecasts minimum variance:
Var[Yn(h)|Y1, . . . ,Yn] = S(S′Σ†hS)−1S′.
Problem: Σh hard to estimate.Forecasting hierarchical (and grouped) time series Optimal forecasts 12
![Page 34: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/34.jpg)
Optimal combination forecasts
Yn(h) = Sβn(h) = S(S′Σ†hS)−1S′Σ†hYn(h)
Initial forecasts
Optimal P = (S′Σ†hS)−1S′Σ†h.
Σ†h is generalized inverse of Σh.
Revised forecasts unbiased: SPS = S.
Revised forecasts minimum variance:
Var[Yn(h)|Y1, . . . ,Yn] = S(S′Σ†hS)−1S′.
Problem: Σh hard to estimate.Forecasting hierarchical (and grouped) time series Optimal forecasts 12
![Page 35: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/35.jpg)
Optimal combination forecasts
Yn(h) = Sβn(h) = S(S′Σ†hS)−1S′Σ†hYn(h)
Initial forecasts
Optimal P = (S′Σ†hS)−1S′Σ†h.
Σ†h is generalized inverse of Σh.
Revised forecasts unbiased: SPS = S.
Revised forecasts minimum variance:
Var[Yn(h)|Y1, . . . ,Yn] = S(S′Σ†hS)−1S′.
Problem: Σh hard to estimate.Forecasting hierarchical (and grouped) time series Optimal forecasts 12
![Page 36: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/36.jpg)
Optimal combination forecasts
Yn(h) = Sβn(h) = S(S′Σ†hS)−1S′Σ†hYn(h)
Revised forecasts Initial forecasts
Optimal P = (S′Σ†hS)−1S′Σ†h.
Σ†h is generalized inverse of Σh.
Revised forecasts unbiased: SPS = S.
Revised forecasts minimum variance:
Var[Yn(h)|Y1, . . . ,Yn] = S(S′Σ†hS)−1S′.
Problem: Σh hard to estimate.Forecasting hierarchical (and grouped) time series Optimal forecasts 12
![Page 37: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/37.jpg)
Optimal combination forecasts
Yn(h) = Sβn(h) = S(S′Σ†hS)−1S′Σ†hYn(h)
Revised forecasts Initial forecasts
Optimal P = (S′Σ†hS)−1S′Σ†h.
Σ†h is generalized inverse of Σh.
Revised forecasts unbiased: SPS = S.
Revised forecasts minimum variance:
Var[Yn(h)|Y1, . . . ,Yn] = S(S′Σ†hS)−1S′.
Problem: Σh hard to estimate.Forecasting hierarchical (and grouped) time series Optimal forecasts 12
![Page 38: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/38.jpg)
Optimal combination forecasts
Yn(h) = Sβn(h) = S(S′Σ†hS)−1S′Σ†hYn(h)
Revised forecasts Initial forecasts
Optimal P = (S′Σ†hS)−1S′Σ†h.
Σ†h is generalized inverse of Σh.
Revised forecasts unbiased: SPS = S.
Revised forecasts minimum variance:
Var[Yn(h)|Y1, . . . ,Yn] = S(S′Σ†hS)−1S′.
Problem: Σh hard to estimate.Forecasting hierarchical (and grouped) time series Optimal forecasts 12
![Page 39: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/39.jpg)
Optimal combination forecasts
Yn(h) = Sβn(h) = S(S′Σ†hS)−1S′Σ†hYn(h)
Revised forecasts Initial forecasts
Optimal P = (S′Σ†hS)−1S′Σ†h.
Σ†h is generalized inverse of Σh.
Revised forecasts unbiased: SPS = S.
Revised forecasts minimum variance:
Var[Yn(h)|Y1, . . . ,Yn] = S(S′Σ†hS)−1S′.
Problem: Σh hard to estimate.Forecasting hierarchical (and grouped) time series Optimal forecasts 12
![Page 40: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/40.jpg)
Optimal combination forecasts
Yn(h) = Sβn(h) = S(S′Σ†hS)−1S′Σ†hYn(h)
Revised forecasts Initial forecasts
Optimal P = (S′Σ†hS)−1S′Σ†h.
Σ†h is generalized inverse of Σh.
Revised forecasts unbiased: SPS = S.
Revised forecasts minimum variance:
Var[Yn(h)|Y1, . . . ,Yn] = S(S′Σ†hS)−1S′.
Problem: Σh hard to estimate.Forecasting hierarchical (and grouped) time series Optimal forecasts 12
![Page 41: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/41.jpg)
Outline
1 Hierarchical and grouped time series
2 Optimal forecasts
3 Approximately optimal forecasts
4 Temporal hierarchies
5 References
Forecasting hierarchical (and grouped) time series Approximately optimal forecasts 13
![Page 42: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/42.jpg)
Approx. optimal forecasts
Forecasting hierarchical (and grouped) time series Approximately optimal forecasts 14
Yn(h) = S(S′Σ†hS)−1S′Σ†hYn(h)
Solution 1: OLSAssume εh ≈ SεB,h where εB,h is the forecasterror at bottom level.
If Moore-Penrose generalized inverse used,then (S′Σ†S)−1S′Σ† = (S′S)−1S′.
Yn(h) = S(S′S)−1S′Yn(h)
![Page 43: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/43.jpg)
Approx. optimal forecasts
Forecasting hierarchical (and grouped) time series Approximately optimal forecasts 15
Yn(h) = S(S′Σ†hS)−1S′Σ†hYn(h)
Solution 2: RescalingSuppose we approximate Σh by its diagonal.
Let Λ =[diagonal
(Σ1
)]−1contain inverse
one-step ahead in-sample forecast errorvariances.
Yn(h) = S(S′ΛS)−1S′ΛYn(h)
![Page 44: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/44.jpg)
Approx. optimal forecasts
Forecasting hierarchical (and grouped) time series Approximately optimal forecasts 16
Yn(h) = S(S′ΛS)−1S′ΛYn(h)
Solution 3: AveragingIf the bottom level error series areapproximately uncorrelated and have similarvariances, then Λ is inversely proportional tothe number of series contributing to eachnode.
So set Λ to be the inverse row sums of S:
Λ = diag(S× 1)−1
where 1 = (1,1, . . . ,1)′.
![Page 45: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/45.jpg)
Outline
1 Hierarchical and grouped time series
2 Optimal forecasts
3 Approximately optimal forecasts
4 Temporal hierarchies
5 References
Forecasting hierarchical (and grouped) time series Temporal hierarchies 17
![Page 46: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/46.jpg)
Temporal hierarchies: quarterly
Annual
Semi-Anual1
Q1 Q2
Semi-Anual2
Q3 Q4
Basic idea:å Forecast series at each available
frequency.
å Optimally combine forecasts within thesame year.
Forecasting hierarchical (and grouped) time series Temporal hierarchies 18
![Page 47: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/47.jpg)
Temporal hierarchies: quarterly
Annual
Semi-Anual1
Q1 Q2
Semi-Anual2
Q3 Q4
Basic idea:å Forecast series at each available
frequency.
å Optimally combine forecasts within thesame year.
Forecasting hierarchical (and grouped) time series Temporal hierarchies 18
![Page 48: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/48.jpg)
Temporal hierarchies: monthly
Annual
Semi-Anual1
Q1
M1 M2 M3
Q2
M4 M5 M6
Semi-Anual2
Q3
M7 M8 M9
Q4
M10 M11 M12
Aggregate: 3, 6, 12Alternatively: 2, 4, 12.How about: 2, 3, 4, 6, 12?
Forecasting hierarchical (and grouped) time series Temporal hierarchies 19
![Page 49: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/49.jpg)
Temporal hierarchies: monthly
Annual
FourM1
BiM1
M1 M2
BiM2
M3 M4
FourM2
BiM3
M5 M6
BiM4
M7 M8
FourM3
BiM5
M9 M10
BiM6
M11 M12
Aggregate: 3, 6, 12Alternatively: 2, 4, 12.How about: 2, 3, 4, 6, 12?
Forecasting hierarchical (and grouped) time series Temporal hierarchies 19
![Page 50: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/50.jpg)
Temporal hierarchies: monthly
Annual
FourM1
BiM1
M1 M2
BiM2
M3 M4
FourM2
BiM3
M5 M6
BiM4
M7 M8
FourM3
BiM5
M9 M10
BiM6
M11 M12
Aggregate: 3, 6, 12Alternatively: 2, 4, 12.How about: 2, 3, 4, 6, 12?
Forecasting hierarchical (and grouped) time series Temporal hierarchies 19
![Page 51: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/51.jpg)
Monthly data
ASemiA1
SemiA2
FourM1
FourM2
FourM3
Q1
...Q4
BiM1
...BiM6
M1
...M12
︸ ︷︷ ︸
(28×1)
=
1 1 1 1 1 1 1 1 1 1 1 11 1 1 1 1 1 0 0 0 0 0 00 0 0 0 0 0 1 1 1 1 1 11 1 1 1 0 0 0 0 0 0 0 00 0 0 0 1 1 1 1 0 0 0 00 0 0 0 0 0 0 0 1 1 1 11 1 1 0 0 0 0 0 0 0 0 0
...0 0 0 0 0 0 0 0 0 1 1 11 1 0 0 0 0 0 0 0 0 0 0
...0 0 0 0 0 0 0 0 0 0 1 1
I12
︸ ︷︷ ︸
S
M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
M11
M12
︸ ︷︷ ︸
Bt
Forecasting hierarchical (and grouped) time series Temporal hierarchies 20
![Page 52: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/52.jpg)
Experimental setup:
M3 forecasting competition (Makridakisand Hibon, 2000, IJF). In total 3003 series.
1,428 monthly series with a test sample of12 observations each.
756 quarterly series with a test sample of8 observations each.
Forecast each series with ETS (ARIMA)models. Methods performed well in theactual competition.
Forecasting hierarchical (and grouped) time series Temporal hierarchies 21
![Page 53: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/53.jpg)
Experimental setup:
M3 forecasting competition (Makridakisand Hibon, 2000, IJF). In total 3003 series.
1,428 monthly series with a test sample of12 observations each.
756 quarterly series with a test sample of8 observations each.
Forecast each series with ETS (ARIMA)models. Methods performed well in theactual competition.
Forecasting hierarchical (and grouped) time series Temporal hierarchies 21
![Page 54: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/54.jpg)
Experimental setup:
M3 forecasting competition (Makridakisand Hibon, 2000, IJF). In total 3003 series.
1,428 monthly series with a test sample of12 observations each.
756 quarterly series with a test sample of8 observations each.
Forecast each series with ETS (ARIMA)models. Methods performed well in theactual competition.
Forecasting hierarchical (and grouped) time series Temporal hierarchies 21
![Page 55: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/55.jpg)
Experimental setup:
M3 forecasting competition (Makridakisand Hibon, 2000, IJF). In total 3003 series.
1,428 monthly series with a test sample of12 observations each.
756 quarterly series with a test sample of8 observations each.
Forecast each series with ETS (ARIMA)models. Methods performed well in theactual competition.
Forecasting hierarchical (and grouped) time series Temporal hierarchies 21
![Page 56: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/56.jpg)
Results: Monthly
Forecast Horizon (h)sMAPE Annual SemiA FourM Q BiM M Average(obs) (1) (2) (3) (4) (6) (12)
ETS
Initial 9.66 9.18 9.76 10.14 10.82 12.85 10.40
Bottom-up 8.38 9.14 9.78 10.06 11.04 12.85 10.21
OLS 7.80 8.64 9.39 9.72 10.68 12.68 9.82Scaling 7.64 8.44 9.15 9.49 10.45 12.40 9.60Averaging 7.51 8.31 9.05 9.38 10.34 12.30 9.48
Forecasting hierarchical (and grouped) time series Temporal hierarchies 22
![Page 57: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/57.jpg)
Results: Quarterly
Forecast Horizon (h)sMAPE Annual Semi-Ann Quart Average(obs) (2) (4) (8)
ETS
Initial 10.50 9.97 9.84 10.10
Bottom-up 8.87 9.35 9.84 9.35
OLS 9.31 9.78 10.28 9.79Scaling 8.75 9.19 9.70 9.21Averaging 8.81 9.26 9.78 9.28
Forecasting hierarchical (and grouped) time series Temporal hierarchies 23
![Page 58: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/58.jpg)
Outline
1 Hierarchical and grouped time series
2 Optimal forecasts
3 Approximately optimal forecasts
4 Temporal hierarchies
5 References
Forecasting hierarchical (and grouped) time series References 24
![Page 59: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/59.jpg)
More information
Forecasting hierarchical (and grouped) time series References 25
Vignette on CRAN
![Page 60: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/60.jpg)
References
RJ Hyndman, RA Ahmed, G Athanasopoulos, and HL
Shang (2011). “Optimal combination forecasts for
hierarchical time series”. Computational Statistics and
Data Analysis 55(9), 2579-2589.
G Athanasopoulos, RA Ahmed, RJ Hyndman,(2009).
“Hierarchical forecasts for Australian domestic tourism”.
International Journal of Forecasting 25, 146-166.
RJ Hyndman, et al., (2014). hts: Hierarchical time series.
cran.r-project.org/package=hts.
RJ Hyndman and G Athanasopoulos (2014). Forecasting:
principles and practice. OTexts. www.otexts.org/fpp/.
Forecasting hierarchical (and grouped) time series References 26
![Page 61: (with Rob J. Hyndman, Nikos Kourentzes and Fotis Petropoulos ... › wp-content › uploads › gravity_forms › 7-2a51… · Fotis Petropoulos) Forecasting hierarchical (and grouped)](https://reader033.fdocuments.net/reader033/viewer/2022060407/5f0fb0937e708231d4456712/html5/thumbnails/61.jpg)
References
RJ Hyndman, RA Ahmed, G Athanasopoulos, and HL
Shang (2011). “Optimal combination forecasts for
hierarchical time series”. Computational Statistics and
Data Analysis 55(9), 2579-2589.
G Athanasopoulos, RA Ahmed, RJ Hyndman,(2009).
“Hierarchical forecasts for Australian domestic tourism”.
International Journal of Forecasting 25, 146-166.
RJ Hyndman, et al., (2014). hts: Hierarchical time series.
cran.r-project.org/package=hts.
RJ Hyndman and G Athanasopoulos (2014). Forecasting:
principles and practice. OTexts. www.otexts.org/fpp/.
Forecasting hierarchical (and grouped) time series References 26
å Email:
Acknowledgments: Rob J Hyndman, Roman
Ahmed, Han Shang, Shanika Wickramasuriya,
Nikolaos Kourentzes, Fotios Petropoulos.
Acknowledgments: Excellent research
assistance by Earo Wang.