1. Biography for William Swan Currently the “Cheap” Economist for Boeing Commercial Aircraft. Previous to Boeing, worked at American Airlines in Operations Research and Strategic Planning and United Airlines in Research and Development. Areas of work included Yield Management, Fleet Planning, Aircraft Routing, and Crew Scheduling. Also worked for Hull Trading, a major market maker in stock index options, and on the staff at MIT’s Flight Transportation Lab. Apparently has a hard time holding a steady job. Education: Master’s, Engineer’s Degree, and Ph. D. at MIT. Bachelor of Science in Aeronautical Engineering at Princeton. Likes dogs and dark beer. © Scott Adams

2. What Airplanes Should Cost William M. Swan Chief Economist Boeing Commercial Airplanes Marketing August 2003

3. Two Things that “bug” Me Thing One 1.Network Design would be very easy if: Cost per flight were linear with size Also handy if it were linear with distance

4. Two Things that “bug” Me Thing 2 2.Policy tests would be very easy if: Cost per ASK were linear with distance Also handy if it were linear with airplane size

5. Why are Networks “Easy”? Linear cost with size implies –A fixed cost per link or per frequency (in our example equal to about 100 seats) –Constant marginal cost for capacity –Capacity and Link Existence separable Linear cost with distance implies –Mileage and Departure costs are separable Separable and linear costs allow Linear Programming formulations of Networks

6. Why are Policy Tests “Easy” Known cost vs. distance allows adjustments –Compare efficiency at different stage lengths $.083/seat-km @500km same as $.050 @1550km –Linear means average does not depend on mix Known cost vs. size also allows adjustments $.050/seat-km @150 seats same as $.045 @200 –Although sizes are less significantly different Real data points too few to test policy and calibrate cost function simultaneously

7. Thing 1 “bugs” me Because I “know” that the distance relationship is:

8. Thing 2 “bugs” me Because I “know” that the size relationship is:

9. Arrogance Invented: Why do I “know” these things? 1.At M.I.T. we found costs linear with with size and distance for parametric designs of airplanes with identical missions and levels of the three basic technologies of propulsion, structure, and drag

10. Arrogance Revisited: Why do I “know” these things? 2.At United Airlines we found costs linear with distance for existing airplanes used in assignments over the network for route planning and with size using the same program with fleet planning costs.

11. Arrogance Revisited Again: Why do I “know” these things? 3.At Boeing the relationships are linear with distance and size based on data generated by a standard cost program used both for sales and for product planning. This program has accumulated over 30 years of calibration and use.

12. The Challenge: Share the Results without sharing proprietary data We can use cost “index” values –Applications require relative costs, not absolute We can fit trends to data points –No need to publish airplane-model level values We can discuss motivation for shapes –Engineering cost functions: each formula has a reason for its algebraic structure

13. First: Imagine the data Cost Per Airplane Trip Trip Distance (KM) SEATS5001000150020002500300040005000 1002.53.54.55.56.57.59.511.5 1503.85.36.88.39.811.314.317.3 2005.07.09.011.013.015.019.023.0 2506.38.811.313.816.318.823.828.8 3007.510.513.516.519.522.528.534.5 40010.014.018.022.026.030.038.046.0

14. About the Data Each cell is generated by one “run” of the Cost Model Able to get data covering entire operating range an airplane is capable of Able to get data for airplanes of many different sizes Able to generate as many points as we want

15. About the Airplanes Boeing Airplanes only –Covers all sizes from smallest to largest Competitive product in all categories –No biases due to manufacturer preference Credibility not subject to whinging suspicions –Conclusions checked offline using Airbus too Airbus data less certain, increased scatter Did not change conclusions

16. Comparable Seat Count Rules Desired “fair” measure of airplane capacity Seat counts at comparable “use”: 1.Same pitch and width by class 2.Same mix of classes onboard 3.Same ratios of lavatories, galleys, closets Counts were “theoretical” adjusted closest configuration to targets above results were fair relative sizes Results were near practical counts

17. Public Sources Failed Seating density changes with usage –The longer the haul, the lower the density –More low-cost usage, the higher the density Tried and failed –Most popular/Median scheduled seat count –Median seats on ownership records (Airclaims) –Either manufacturers’ claims

18. “Fair” Seat Counts are Hard to Get airplaneShort-HaulLong-Haul 737-70012288 737-800153110 737-900165119 757 200187135 767 200204147 757 300226163 767 300250180 767 400286206 777 200385277 777 300481346 747-400536386

19. “Fair” Seat Counts are Hard to Get airplaneShort-HaulLong-Haul A31810374 A31912187 A320140101 A321177127 A310-200/300218157 A300263189 A330 200291210 A340 200304219 A330/340 300335241 A340-500367264 A340-600418301

20. Comparable Design Missions Within a type (“737”) and model (“-700”) –Different engines, Different weight options Select for Comparable missions –As close to 5000km max range for short-haul –As close to 10000km range for long-haul 757s used although they fell short on range 767s at highest weights 777s and 747 at lowest weights

21. Major Cost Components: Crew Flight Crew cost per block hour –Hourly costs increase with airplane size –Pilots are about 12% of airplane costs Cabin Crew cost per block hour –Hourly costs independent of airplane –Crew in proportion to seat count (about 1:40) –Cabin crew are about 10% of airplane costs

22. Major Cost Component: Fuel Fuel cost nearly linear with distance –Cost per departure/landing cycle –Cost per cruise kilometer –Very small increase in cost for very long haul Due to cost of carrying weight of fuel used at end Fuel cost about 12% of total costs Actual use widely and accurately available

23. Major Cost Component: Maintenance Maintenance on Engines and Airframe –Per departure-landing cycle and per hour –Includes labor, parts, and facilities Periodic Maintenance accrued –Major checks every 3-4 years Steady-state values used –5-year Maintenance holiday from new is over Maintenance averages 13% of airplane costs

24. Major Costs: Ownership Ownership at 32% for new designs –Estimate based on lease rates –0.8%-0.9% of market price, per month Lease rates include reported and unreported costs: –Return to capital –Tax benefits –Depreciation

25. Allocating Ownership Lease rates allocated based on trips per year Formula is: –Trips = 4560/(1.5+800/distance(km)) Matches hours and cycles by tail number No allocation of ownership to peak season or peak flight of the day Works out to ownership per km and per cycle

26. Cost Categories: Fees Landing fees based on weight Air Traffic Control based on weight & hrs. Run 8%-14% of airplane costs –Outside of US (US is lower) Security and other passenger charges not included in airplane costs –These costs are growing

27. Cost Categories: Overheads Overheads not covered in this analysis Airplane costs above are 60% of total costs Additional 20-30% is General and Administrative –Thought to be proportional to airplane costs Additional 20% is passenger costs –Mostly proportional to revenues, or costs –Some on a per-passenger-trip basis

28. Rule Sets for Costing Methodology Cost formulas have various rule sets –US domestic, European Short-haul,....... European Short-Haul used under 5000km –Representative for all non-US regional flying International Long-Haul used over 5000km Choice of rule sets secondary for relative costs of different airplanes or ranges.

29. Results Short-Haul Cost per Airplane trip: $ = k * (Dist + 722) * (Seats + 97) Dist is trip distance in km Seats are airplane size in short-haul seat count k is about $0.02 Long-Haul Cost per Airplane trip: $ = K * (Dist + 530) * (Seats + 205) Dist is trip distance in km Seats are airplane size in long-haul seat count K is about $0.03 Long-haul seats = 72% of short-haul seats

30. Calibration of Results Linear Regression of the form: –$ = a + b * dist + c * seats + d * (seats*dist) Two-stage calibration used to get planar form: $ = k * (dist + c) * (seats + b) $ = k* {c*b + b*dist + c*seats + dist*seats} Stage 1: for each airplane size $ = z * (dist + c) Stage 2: pick z = k * (seats + b) least squares unbiased estimate of k and b

31. Comments on Calibration Values almost unchanged doing size first, then range rather than range, then size Values almost unchanged using least squared percentage error, or summed absolute error Matrix of input data points had correct averages for seats and range Matrix of input data points covered wide spectrum of ranges and seats

32. Linear Regression Comparison Run against same data points Format: $ = a + b * dist + c * seats + d * (seats*dist) Formulated to test whether “d” is different from planar Results: short $ = long $ =

33. Cobb-Douglas Comparison Run against same data points Gives Elasticities from “certain” data –Very broad spectrum of ranges –Very broad spectrum of airplane sizes Results: –Short $ = x * dist  *  seats  –Long $ = x * dist  *  seats 

34. Cobb-Douglas is no Less Accurate Plot C-D vs Planar --just less linear and therefore less convenient

35. C-D elasticities compare to other work Note other work results. Why different? –Few points’ –Little change in range or size –Step changes in labor costs, year to year Based on contract renegotiations

36. Why I like the Planar results Points are “perfect” – no complicating factors Points cover entire range of size and distances No false correlation between size and range No false correlation between range and seating density No airport-specific effects

37. What is NOT in the soup Factor input prices the same for all points –Includes labor, fuel, capital, airports Per passenger (therefore load factor) out –No passenger-related costs, just seats Supplemental crews at long distances

38. What it is Good For Great for adjusting different airlines’ or years’ costs to same stage length and seating capacity—for comparing firm- specific costs Great for designing networks based on typical cost functions

39. Adjusting for Southwest Ownership (trips per year) Seating count Factor inputs: same

40. Adjusting for Pacific Seat count Crew costs?

41. Conclusion Linear works –Close enough for network design –Makes network problem much simpler Relative costs work –Good for normalizing firm-specific –Helps overcome small sample problems (eliminates size, range variables)

42. William Swan: Data Troll Story Teller Economist