Measuring Airline Networks

Transcription

1 Measuring Airline Networks Chantal Roucolle (ENAC-DEVI) Joint work with Miguel Urdanoz (TBS) and Tatiana Seregina (ENAC-TBS) This research was possible thanks to the financial support of the Regional Council of Midi Pyrenees.

2 Airline networks Airline networks are complex and dynamic Number of Airports served Number of markets served Direct or connecting flights, frequencies, schedules 2

3 US airports 3

4 US domestic network in July

5 US domestic network in July

6 US domestic network in July

7 Airline networks Airline networks are complex and dynamic Number of Airports served Number of markets served Direct or connecting flights, frequencies, schedules Airlines make different choices 7

8 Frontier Airlines, July

9 Southwest Airlines, July

10 Airline networks Airline networks are complex and dynamic Number of Airports served Number of markets served Direct or connecting flights, frequencies, schedules Airlines make different choices and their decisions evolve over time 10

11 Evolution: Delta Airlines July

12 Evolution: Delta Airlines July

13 Evolution: Delta Airlines July

14 Airline network Airline networks are complex and dynamic Number of Airports served Number of markets served Frequencies, schedules Airlines make different choices and their decisions evolve over time Two questions to address: Network characterization Network evolution: what are the drivers of the choice? Usefulness of network analysis: Does the network structure affect costs, prices, profitability, delays? 14

15 Literature In most of the cases perfect hub-and-spoke networks (left) are compared with fully connected networks (right), For instance Brueckner (2004), Alderighi et al. (2005), Barla and Constantatos (2005), Flores Fillol (2009) or Silva et al. (2014). However reality is more complex Wojahn (2001) studies whether a mixed model can be preferred to minimize costs We want to get closer to this reality 15

16 Our objective Step 1: Network characterization Our approach: combine Graph theory and Principal Component Analysis (PCA) Graph theory: set of mathematical measures and tools to study networks Already used for airlines: Wandelt and Sun (2015), Dunn and Wilkinson (2016) or Du et al. (2016) study different network properties focusing on the country level. Burghouwt and Redondi (2013) present a compilation of connectivity indicators for airports built from graph theory. Lordan et al.(2016) study resilience with a sample of airlines from Europe, North America and China. PCA aims to explain most of the information of the dataset through a reduced number of new variables, called principal components, calculated as linear combinations of the original variables Main findings: Airline network could be characterized by three indicators: Hubness, Resilience, Size Traditional distinction between LCC and Legacies could be reconsidered 16

17 Our objective Step 2: Network evolution: what are the drivers of the choice? Our approach: explain the evolution of the three indicators over time Use of macroeconomic indicators, air market characteristics, airline type as explanatory variables Estimation of a system of equations on panel data Main findings: Network Hubness, Resilience and Size have distinct drivers Strategies in terms of network evolution depend on the type of airline 17

18 Data Official Airline Guide, OAG Worldwide data on frequencies, schedules and aircrafts for the last 10 years Monthly data for the third quarter Focus on the United States Domestic Market Advantage: most studied market with available data on fares Disadvantage: lack of international flights Data cleaning: Eliminating routes shorter than 200 miles or routes with less than 10 seats per flight Recoding of regional/feeder airlines Final data set: observations 28 operating carriers ranging from 19 to 25 per year City number from 413 to 517 and airports ranging from 435 to 537 per year. 18

19 Step 1: Network characterization Graph Theory measures: a lot of correlated indicators 19

20 Step 1: Network characterization Reduction of the information: use of PCA Three principal components explain 94.69% of the sample variability: Hubness, Resilience and Network Size Theoretical representation of network Hubness / Resilience map 20

21 Step 1: Network characterization PC2 : RESILIENCE point-to-point hub-and-spoke with several hubs hub-and-spoke with a unique hub PC1 : HUBNESS path or circle The distinction between LCCs and Legacies is nowadays unclear as highlighted in Jarach et al. (2009) or Bitzan and Peoples (2016). No distinction in term of Hubness (PC1) between Legacies and LCCs. Higher Resilience (PC2) values on average for LCCs 21

22 Step 1: Network characterization Network representation Size / Hubness and Size / Resilience maps WN WN When the network size increases, Hubness decreases to some level between 0 and -3, both for LCCs and Majors When the network size increases, Resilience seems to approach a level around -1, except for LCCs (Southwest case) 22

23 Step 2: Network evolution: what are the drivers of the choice? Simultaneous Equations Model G I ijt = α ij + β i1 t LCC + β i2 t L + β i3 y t 1 + β i4 f t 1 + β i5 d DLNW + β i6 d UACO + β i7 d WNFL + β i7 d AAUS + ε ijt where i {1,2,3} indexes one of three network indicators: Hubness, Resilience, Size j indexes the airlines t indexes the year Explanatory variables: Time trends: t LCC and t L depending on airline type Macroeconomic indicators: y G represents the output gap, and f the jet fuel prices, US domestic market characteristics: dummies to control for the 4 mergers occurred during the considered time frame 23

24 Hubness (i=1) Resilience (i=2) Size (i=3) β i1 t LCC (1.015) (1.613) (10.352)*** β i2 t L G β i3 y t 1 (2.399)** (2.895)*** (0.459) (4.689)*** (2.342)** (0.700) β i4 f t (2.686)*** (2.641)*** (0.521) β i5 d DLWN (3.283)*** (0.774) (4.101)*** β i6 d UACO (4.523)*** (0.602) (22.377)*** β i7 d WNFL (0.409) (0.985) (8.299)*** β i8 d AAUS (1.249) (3.828)*** (2.749)*** Constant WN Legacy AA (3.101)*** (9.959)*** (75.839)*** (4.752)*** (18.206)*** (2.089)** LCC B Legacy AS (24.247)*** (12.668)*** (48.688)*** (0.677) (15.322)*** (7.079)*** LCC F Legacy CO (44.737)*** (26.114)*** (75.420)*** (3.615)*** (22.771)*** (3.553)*** LCC FL Legacy HA (25.315)*** (33.692)*** (27.508)*** (12.375)*** (6.563)*** (28.786)*** LCC G Legacy DL (7.760)*** (22.403)*** (49.086)*** (2.281)** (21.224)*** (3.261)*** LCC NK Legacy NW (2.052)** (1.393) (68.473)*** (1.314) (24.469)*** (1.663)* LCC SY Legacy UA (6.298)*** (19.657)*** (37.913)*** (1.293) (21.894)*** (1.475) LCC VX Legacy US (11.992)*** (1.373) (54.925)*** (2.852)*** (19.464)*** (0.595) ρ Estimation results Observations * p<0.1; ** p<0.05; *** p<0.01

25 Hubness (i=1) Resilience (i=2) Size (i=3) β i1 t LCC (1.015) (1.613) (10.352)*** β i2 t L G β i3 y t 1 (2.399)** (2.895)*** (0.459) (4.689)*** (2.342)** (0.700) β i4 f t (2.686)*** (2.641)*** (0.521) β i5 d DLWN (3.283)*** (0.774) (4.101)*** β i6 d UACO (4.523)*** (0.602) (22.377)*** β i7 d WNFL (0.409) (0.985) (8.299)*** β i8 d AAUS (1.249) (3.828)*** (2.749)*** Constant WN Legacy AA (3.101)*** (9.959)*** (75.839)*** (4.752)*** (18.206)*** (2.089)** LCC B Legacy AS (24.247)*** (12.668)*** (48.688)*** (0.677) (15.322)*** (7.079)*** LCC F Legacy CO (44.737)*** (26.114)*** (75.420)*** (3.615)*** (22.771)*** (3.553)*** LCC FL Legacy HA (25.315)*** (33.692)*** (27.508)*** (12.375)*** (6.563)*** (28.786)*** LCC G Legacy DL (7.760)*** (22.403)*** (49.086)*** (2.281)** (21.224)*** (3.261)*** LCC NK Legacy NW (2.052)** (1.393) (68.473)*** (1.314) (24.469)*** (1.663)* LCC SY Legacy UA (6.298)*** (19.657)*** (37.913)*** (1.293) (21.894)*** (1.475) LCC VX Legacy US (11.992)*** (1.373) (54.925)*** (2.852)*** (19.464)*** (0.595) ρ Estimation results Observations * p<0.1; ** p<0.05; *** p<0.01 Main findings Hubness: initial gap between LCC and Legacies is vanishing over time

26 Hubness (i=1) Resilience (i=2) Size (i=3) β i1 t LCC (1.015) (1.613) (10.352)*** β i2 t L G β i3 y t 1 (2.399)** (2.895)*** (0.459) (4.689)*** (2.342)** (0.700) β i4 f t (2.686)*** (2.641)*** (0.521) β i5 d DLWN (3.283)*** (0.774) (4.101)*** β i6 d UACO (4.523)*** (0.602) (22.377)*** β i7 d WNFL (0.409) (0.985) (8.299)*** β i8 d AAUS (1.249) (3.828)*** (2.749)*** Constant WN Legacy AA (3.101)*** (9.959)*** (75.839)*** (4.752)*** (18.206)*** (2.089)** LCC B Legacy AS (24.247)*** (12.668)*** (48.688)*** (0.677) (15.322)*** (7.079)*** LCC F Legacy CO (44.737)*** (26.114)*** (75.420)*** (3.615)*** (22.771)*** (3.553)*** LCC FL Legacy HA (25.315)*** (33.692)*** (27.508)*** (12.375)*** (6.563)*** (28.786)*** LCC G Legacy DL (7.760)*** (22.403)*** (49.086)*** (2.281)** (21.224)*** (3.261)*** LCC NK Legacy NW (2.052)** (1.393) (68.473)*** (1.314) (24.469)*** (1.663)* LCC SY Legacy UA (6.298)*** (19.657)*** (37.913)*** (1.293) (21.894)*** (1.475) LCC VX Legacy US (11.992)*** (1.373) (54.925)*** (2.852)*** (19.464)*** (0.595) ρ Estimation results Observations * p<0.1; ** p<0.05; *** p<0.01 Main findings Hubness: initial gap between LCC and Legacies is vanishing over time Resilience: initial gap between LCC and Legacies remains over time

27 Hubness (i=1) Resilience (i=2) Size (i=3) β i1 t LCC (1.015) (1.613) (10.352)*** β i2 t L G β i3 y t 1 (2.399)** (2.895)*** (0.459) (4.689)*** (2.342)** (0.700) β i4 f t (2.686)*** (2.641)*** (0.521) β i5 d DLWN (3.283)*** (0.774) (4.101)*** β i6 d UACO (4.523)*** (0.602) (22.377)*** β i7 d WNFL (0.409) (0.985) (8.299)*** β i8 d AAUS (1.249) (3.828)*** (2.749)*** Constant WN Legacy AA (3.101)*** (9.959)*** (75.839)*** (4.752)*** (18.206)*** (2.089)** LCC B Legacy AS (24.247)*** (12.668)*** (48.688)*** (0.677) (15.322)*** (7.079)*** LCC F Legacy CO (44.737)*** (26.114)*** (75.420)*** (3.615)*** (22.771)*** (3.553)*** LCC FL Legacy HA (25.315)*** (33.692)*** (27.508)*** (12.375)*** (6.563)*** (28.786)*** LCC G Legacy DL (7.760)*** (22.403)*** (49.086)*** (2.281)** (21.224)*** (3.261)*** LCC NK Legacy NW (2.052)** (1.393) (68.473)*** (1.314) (24.469)*** (1.663)* LCC SY Legacy UA (6.298)*** (19.657)*** (37.913)*** (1.293) (21.894)*** (1.475) LCC VX Legacy US (11.992)*** (1.373) (54.925)*** (2.852)*** (19.464)*** (0.595) ρ Estimation results Observations * p<0.1; ** p<0.05; *** p<0.01 Main findings Hubness: initial gap between LCC and Legacies is vanishing over time Resilience: initial gap between LCC and Legacies remains over time Size: LCC increase their size over time; initial gap is vanishing

28 Hubness (i=1) Resilience (i=2) Size (i=3) β i1 t LCC (1.015) (1.613) (10.352)*** β i2 t L G β i3 y t 1 (2.399)** (2.895)*** (0.459) (4.689)*** (2.342)** (0.700) β i4 f t (2.686)*** (2.641)*** (0.521) β i5 d DLWN (3.283)*** (0.774) (4.101)*** β i6 d UACO (4.523)*** (0.602) (22.377)*** β i7 d WNFL (0.409) (0.985) (8.299)*** β i8 d AAUS (1.249) (3.828)*** (2.749)*** Constant WN Legacy AA (3.101)*** (9.959)*** (75.839)*** (4.752)*** (18.206)*** (2.089)** LCC B Legacy AS (24.247)*** (12.668)*** (48.688)*** (0.677) (15.322)*** (7.079)*** LCC F Legacy CO (44.737)*** (26.114)*** (75.420)*** (3.615)*** (22.771)*** (3.553)*** LCC FL Legacy HA (25.315)*** (33.692)*** (27.508)*** (12.375)*** (6.563)*** (28.786)*** LCC G Legacy DL (7.760)*** (22.403)*** (49.086)*** (2.281)** (21.224)*** (3.261)*** LCC NK Legacy NW (2.052)** (1.393) (68.473)*** (1.314) (24.469)*** (1.663)* LCC SY Legacy UA (6.298)*** (19.657)*** (37.913)*** (1.293) (21.894)*** (1.475) LCC VX Legacy US (11.992)*** (1.373) (54.925)*** (2.852)*** (19.464)*** (0.595) ρ Estimation results Observations * p<0.1; ** p<0.05; *** p<0.01 Main findings Hubness: initial gap between LCC and Legacies is vanishing over time Resilience: initial gap between LCC and Legacies remains over time Size: LCC increase their size over time; initial gap is vanishing Macroeconomic environment affect the network strategical decisions but not the size

29 Step 2: Network evolution: what are the drivers of the choice? Mergers mapping Following the merger: Increase in Size Decrease in Hubness the year of the merger Hubness recovers its initial level the years after 29

30 Conclusions and further research We propose a methodology to determine the drivers of network evolution Step 1: building three network indicators: Hubness, Resilience and Size Step 2: analysis of these indicators over time given macroeconomic and market conditions We apply the methodology to the US domestic networks Level and evolution of the networks depend on the type of airline LCCs and Legacies differ in terms of Resilience while there seems to be a convergence in terms of Hubness and Size. Next steps: study the impact of these indicators over the airline s cost structure Pels et al. (2000), the level of competition (Hendricks, Piccione, and Tan 1997), the prices (Tan and Samuel 2016), the level of congestion and delays (Mayer and Sinai 2003; Brueckner 2002; Fageda and Flores-Fillol (2016)) 30