Transcription
Organizational Structure and Pricing:Evidence from a Large U.S. AirlineAli Hortaçsu, University of Chicago and NBEROlivia R. Natan , University of California, BerkeleyHayden Parsley, University of Texas, AustinTimothy Schwieg, University of Chicago, BoothKevin R. Williams, Yale School of Management and NBER*November 2021AbstractWe study how organizational boundaries affect pricing decisions using comprehensive data provided by a large U.S. airline. We show that contrary to prevailingtheories of the firm, advanced pricing algorithms have multiple biases. To quantifythe impacts of these biases, we estimate a structural demand model using sales andsearch data. We recover the demand curves the firm believes it faces using forecastingdata. In counterfactuals, we show that correcting biases introduced by organizationalteams individually have little impact on market outcomes, but addressing all biasessimultaneously leads to higher prices and increased dead-weight loss in the marketsstudied.JEL Classification: C11, C53, D22, D42, L10, L93Keywords: Pricing Frictions, Organizational Inertia, Dynamic Pricing, Revenue Management, Behavioral IO* The views expressed herein are those of the authors and do not necessarily reflect the views of theNational Bureau of Economic Research. We thank the anonymous airline for giving us access to the dataused in this study. Under the agreement with the authors, the airline had "the right to delete any trade secret, proprietary, or Confidential Information" supplied by the airline. We agreed to take comments in goodfaith regarding statements that would lead a reader to identify the airline and damage the airline’s reputation. All authors have no material financial relationships with entities related to this research. We alsothank the seminar participants at Yale University, University of Virginia, Federal Reserve Bank of Minneapolis, University of Chicago, University of California-Berkeley, the NBER 2021 Summer Institute, and theVirtual Quantitative Marketing Seminar for comments. We gratefully acknowledge financial support fromthe NET Institute, www.NETinst.org. Emails: [email protected], [email protected], timothys[email protected], [email protected], [email protected]
1IntroductionDramatic decreases in the cost of computation and data storage, along with algorithmicinnovations, have increasingly allowed firms to develop data-driven decision optimizationsystems. Data and algorithms now play a key role in driving firm decisions across industries. This is especially relevant in the airline context where firms must match fixed flightcapacity with dynamically evolving demand. To solve this difficult allocation problem,airlines have developed sophisticated pricing systems over the last several decades. Thesesystems depend on inputs from multiple organizational teams. How airlines allocate decision rights across teams within the firm is not unique to the industry. Hotels, cruises, carrentals, entertainment venues, and retailers have all adopted features of the airline pricingmodel. Given the investments firms have made into these decision machines and their wideuse across industries, we may expect that prices are close to optimal.In this paper, we study how organizational boundaries affect pricing decisions by leveraging a data partnership with a large international air carrier based in the United States.1The granularity of the data allow us to understand the firm’s incentives to adjust priceswithout needing to assume prices are optimally set. We show that contrary to prevalenttheories of the firm, the pricing at a sophisticated firm—one that employs advanced optimization techniques and has a heavy reliance on automation—does not appear to react tosome important market fundamentals. This includes not internalizing consumer substitution to other products, using persistently biased forecasts, and not responding to changes inopportunity costs driven by scarcity. We show that these biases are introduced by separateteams within the firm. What happens to prices and allocative efficiency if organizationalteams address the pricing frictions that we document? Using a new technique to estimatedemand and detailed forecasting data to infer the firm’s beliefs about the demand it faces,we find that correcting pricing frictions introduced by teams individually does little to affect market outcomes. However, we also show that addressing biased introduced acrossorganizational teams simultaneously can result in increased price targeting and higher rev1 Theairline has elected to remain anonymous.1
enues, but also higher dead-weight loss for the routes studied. Our results highlight hownon-coordinating teams with complementary functions can have significant consequenceson firm performance.We begin by providing an empirical glimpse under the hood of dynamic pricing solutions used by airlines. In addition to observing prices and quantities, we also observegranular demand forecasts, output of the pricing and allocation algorithms, the optimization code itself, and clickstream data that detail all consumer interactions on the airline’swebsite. The core data cover hundreds of thousands of flights spanning hundreds of domestic origin-destination pairs. We document the main organizational details of how pricingdecisions are made within the firm and provide insights on the incentives to adjust pricesover time in Section 2 and Section 3, respectively. We show that all major airlines havesimilar organizational structures. Therefore, we believe our discussion and subsequent empirical findings likely hold for the airline industry broadly and perhaps for other industriesthat have adopted similar pricing technologies and organizational structures.In Section 4, we discuss data patterns that suggest the airline could be doing more withits data. We show that prices do not necessarily adjust when the value of remaining capacitychanges. This is caused by coarse pricing, or the use of “fare buckets." However, the factopportunity costs may adjust by hundreds of dollars without triggering a price adjustmentsuggests a mismatch between the fares chosen by one organizational team and demandfundamentals. We establish that the forecasts maintained by a separate team respond todemand “surprises” too little and too late. These demand forecasts are biased upward intwo years of data. We establish that the pricing algorithm itself is biased by showing thatcross-price elasticities are not considered when setting prices. Finally, we show that thefirm chooses prices on the inelastic side of the demand curve—according to their ownestimates of demand—in more than half of the data sample.In the second stage of our analysis, we quantify the impacts of these observed pricingfrictions on welfare. To do this, we estimate a structural model of consumer demand using arecently proposed demand methodology (Hortaçsu, Natan, Parsley, Schwieg, and Williams,2021). In Section 5, we consider a model in which “leisure” and “business” travelers2
arrive according to independent and time-varying Poisson distributions in discrete time.Consumers know their preferences and solve discrete choice maximization problems. Eachconsumer chooses among the available flight options or an outside option. We provideevidence to motivate some of our demand assumptions, including that consumers do notappear to be betting on price and consumer arrivals are not endogenous to price.We estimate the model using consumer search and bookings data. Aggregate searchcounts calculated from the clickstream data inform the overall arrival process, and weidentify the price coefficient using instrumental variables (see Section 6). The estimatespresented in Section 7 reveal meaningful variation in demand, with a general increase insearch for travel as the departure date approaches and substantial changes in the overallprice sensitivity of consumers over time. We discuss similarities and differences in modelestimates across routes.Given the demand estimates, we then ask: what does the firm believe its demand curveslook like? We call these “firm beliefs” and we recover them in Section 8 using detailedforecasting data and output from an algorithm that classifies search and bookings as comingfrom a “leisure” or “business” traveler. Relative to our baseline demand estimates, we findthat the firm’s beliefs about their demand has more compressed demand elasticities bothwithin and across routes, more elastic demand near departure, and consumer types thatare “closer together" in terms of preferences. Using these recovered demand curves, weconfirm our descriptive finding that prices are often too low: nearly 30% of observed pricesare below the optimal price if capacity costs were zero.In Section 9, we perform counterfactual exercises using a pricing model that closely follows the heuristic the firm uses. First, we isolate pricing frictions individually—removingforecast bias and mismatched fare choices to the forecast, separately. We show that outcomes are largely unchanged because the pricing heuristic commonly defaults to the lowestfiled fare because it expects that future demand can be satisfied with remaining capacity.However, we also show internalizing complementarities across teams—correcting the forecast and inputting fares into the algorithm tailored to this forecast—yields very differentoutcomes. Coordination guarantees that fares never drop below the optimal price if capac3
ity costs were zero. This raises the distribution of fares offered and allows the firm to targetbusiness travelers with higher fares. Revenues increase substantially—upward of 17% forsome markets. Dead-weight loss also increases by over 10%. The fact that that the firmcould extract additional surplus but has chosen not to do so is puzzling. We argue that this ispossible due to under-experimentation across organizational teams that we quantify in thedata. We also hypothesize the firm may consider the implicit cost of regulatory oversightor long-term competitive responses as alternative explanations.1.1 Related LiteratureThis paper contributes to a growing literature in behavioral industrial organization that examines pricing frictions, including DellaVigna and Gentzkow (2019) and Hitsch, Hortaçsu,and Lin (2021) in retailing, Huang (2021) in peer-to-peer markets, Ellison, Snyder, andZhang (2018) in online retailing, and Cho and Rust (2010) in rental cars. We confirm pricing frictions also impact firms that have pioneered advanced pricing algorithms. Our workalso contributes to research on miscalibrated firm expectations, as forecasts are persistentlybiased (Massey and Thaler, 2013; Akepanidtaworn, Di Mascio, Imas, and Schmidt, 2019;Ma, Ropele, Sraer, and Thesmar, 2020). Dubé and Misra (2021) provide an example wherea firm has not priced optimally and pricing to the correct demand curve greatly increasesrevenue. In our setting, pricing to the correct demand curves is insufficient to greatly impactrevenues if other pricing frictions are not addressed.Our work also contributes to the literature in organizational economics. The adoptionof information technology (IT) can increase productivity when complementary organizational and management practices are implemented alongside these investments (Bresnahan,Brynjolfsson, and Hitt, 2002; Bloom, Sadun, and Van Reenen, 2012).2 Although Atheyand Stern (2002) do not find strong complementarities in one IT setting, we find substantialcomplementarities in the context of pricing algorithms. Our work complements empiricalresearch on the impacts of decentralized-decision making on firm performance (Aguirregabiria and Guiton, 2020; Filippas, Jagabathula, and Sundararajan, 2021) and theoretical2 Brynjolfssonand Milgrom (2012) provide an overview of this and related work.4
work on the inputs to organizational teams, including Che and Yoo (2001), Siemsen, Balasubramanian, and Roth (2007), Alonso, Dessein, and Matouschek (2008, 2015).Finally, this paper quantifies the effectiveness of pricing heuristics proposed in operations research using airline data (Littlewood, 1972; Belobaba, 1987, 1989; Brumelle,McGill, Oum, Sawaki, and Tretheway, 1990; Belobaba, 1992; Wollmer, 1992). Additionalwork on the welfare consequences of dynamic pricing include Sweeting (2012); Cho, Lee,Rust, and Yu (2018); Hendel and Nevo (2013); D’Haultfœuille, Février, Wang, and Wilner(2020); Castillo (2019); Aryal, Murry, and Williams (2021); and Williams (forthcoming).2Organizational Structure and Division ResponsibilitiesWe study the US airline industry, an industry that directly supports over two million jobsand contributes over 700 billion to the US economy.3 In 2019 alone, 811 million passengers flew within the United States.4 . In addition to being an important industry in its ownright, airlines have influenced the development of pricing technologies that are now used inother sectors—for example, in hospitality, retailing, and entertainment and sports events.Although the sophistication of these technologies has improved, many of the original yieldmanagement ideas described in McGill and Van Ryzin (1999) and Talluri and Van Ryzin(2004) remain in place today.Fares at our air carrier depend on the actions of managers in several distinct departments, and these departments do not explicitly coordinate their actions. Generally, decisions become increasingly granular, taking all previous departments’ decisions as given.First, network planning decides the network, flight frequencies, and capacity choices. Second, the pricing department sets a menu of fares and fare restrictions for all possibleitineraries. Finally, the revenue management (RM) department decides the number of seatsto sell for every fare-itinerary combination.The RM group maintains the demand forecasting model and pricing algorithm, but does3 See viation/. July 1, 2021.4 See full-year. July 1, 2021.5
not have control over the fares inputted into these algorithms. The pricing department setsfares, but does not use the forecasting information.5 The forecasting model incorporateshistorical information and current bookings information to predict flight-level sales. Thepricing algorithm allocates remaining inventory given the fare menu. A commonly usedpricing heuristic in the industry is Expected Marginal Seat Revenue (EMSR-b), whichclosely approximates the algorithm used by the firm. We provide additional details ofthe EMSR-b in Online Appendix A and outline the algorithm here. EMSR-b belongs to aclass of static optimization solutions. Dynamics are removed because it assumes all futuredemand will arrive tomorrow. The key trade-off, therefore, is to offer seats today versusreserve them for tomorrow. Given all pricing inputs, it calculates the opportunity cost ofa seat and then assigns the number of seats it is willing to sell at all price levels. Lowestpriced units are assumed to sell first. If expected future demand is high (low), it will restrict(not restrict) inventory at lower prices today.Fare BucketFigure 1: Fare Bucket Availability and Lowest Available ucket6Bucket5Bucket4Bucket3Bucket2Bucket1120Lowest Available Price100806040Days Before Departure200Note: Image plot of fare availability over time as well as the active lowest available fare. Bucket1 is the least expensive; Bucket12 isthe most expensive fare class. The color depicts the magnitude of prices—blue are lower fares, red are more expensive. White spacedenotes no fare availability. The white line depicts the lowest available fare.We demonstrate how inputs impact prices for an example flight at our airline in Figure 1. On the vertical axis, we show the anonymized fare buckets decided by the pricing5 Eachfiled fare contains an origin, destination, filing date, class of service, routing requirements, andother ticket restrictions. A common fare restriction decided by the pricing department is an advance purchasediscount, which specifies an expiration date for a discounted fare to be purchased by. These discounts arecommonly observed seven, 14, and 21 days before departure.6
department, with bucket one being the least expensive and bucket twelve being the mostexpensive. The bottom right of the graph shows that the pricing department restricts theavailability of the lowest fares close to the departure date. There is relatively little variationin prices for a given bucket over time. Given fares and the forecast (not shown), the whiteline marks to lowest available price (LAP) offered to consumers. This is an output of thealgorithm maintained by the RM group. All flights, regardless of market structure, flightfrequencies, etc., are priced using the same algorithm.2.1 Potential Pricing Bias with Uncoordinated InputsPricing heuristics can be sensitive to algorithm inputs. To see this, consider a firm selling15 units over two sequential markets. Demand in the first period is equal to Q1 (p1 ) 10 10p1 , and demand in the second period is Q2 (p2 ) 10 p2 . If the firm maximizestotal revenues subject to the capacity constraint, the capacity constraint will not bind, andoptimal prices are equal to (p1 , p2 ) (0.5, 5). This outcome can be also obtained using thepricing and revenue management roles and the pricing algorithm EMSR-b described aboveif the pricing department assigns prices to be {0.5, 5} and {5}, and the RM group “forecasts”demand to be the functions above.EMSR-b decides the number of seats that can be sold at each input price to ensurethat future demand can be satisfied. Lower prices are restricted only in situations wherefuture demand cannot be satisfied. In this case, five seats are needed for period two, andthe algorithm will appropriately allocate all seats to 0.5 in the first period. Suppose insteadthat the pricing department did not coordinate with the RM group and set prices equal to{.2, .5, 5} and {5}. Note that all first-period prices leave sufficient capacity available forthe second period, which means EMSR-b will allocate all seats at the lowest price, 0.2.Consequently, the heuristic will choose a suboptimal price even though the optimal price,0.5, is included in the choice set.7
2.2 All US Airlines have the same Organizational StructureOur description of airline pricing is not unique to our airline—all airlines have the sameorganizational structure and use similar pricing techniques. We show this by collecting jobpostings information for all the major carriers in the U.S.6 We confirm that Alaska, American, Delta, JetBlue, Southwest, and United have a network planning, pricing, and revenuemanagement department. As an example, JetBlue Airlines job postings show that the firmhas three teams related to pricing: Future Schedules, Revenue Management-Pricing, andRevenue Management-Inventory. Job details delineate team responsibilities. The Revenue Management department at JetBlue has two separate teams, Pricing and Inventory.The Pricing team has ownership over fares by “monitoring industry pricing changes filedthrough a clearinghouse throughout the day, and determining and executing JetBlues response.”7 The Inventory team uses “inventory controls to determine the optimal fare to sellat any given moment in time to maximize each flights revenue.”8 American Airlines managers describe how inventory controls are implemented in Smith, Leimkuhler, and Darrow(1992)—they outline EMSR-b. Because all carriers have the same organizational structureand use similar algorithms, we believe our analysis characterizes the entire industry, ratherthan the perspective from a single firm.3Data and Summary AnalysisWe use data provided by a large international air carrier based in the United States. Tomaintain anonymity, we exclude some data details. In Online Appendix B, we describe ourroute selection criteria.6 Screenshotsof the job postings are available on request.7 See nalyst-Revenue-Management-NY-11101/737962800/.8 See y 1, 2021.July 1, 2021.
3.1 Data TablesWe combine several data sources, which we commonly refer to as: (1) bookings, (2) inventory, (3) search, (4) fares, and (5) forecasting data.(1) Bookings data: The bookings data contain details for each purchased ticket, regardless of booking channel, e.g., the airline’s website, travel agency, etc. Key variablesincluded in these data are the fare paid, the number of passengers involved, the particularflights included in the itinerary, the booking channel, and the purchase date.9 Our analysisconcentrates on nonstop bookings and economy class tickets.(2) Inventory data: The inventory data contain the decisions made by the RM group.Inventory allocation is conducted daily. The data include the number of seats the airline iswilling to sell for each fare class in economy and aircraft capacity. We also observe outputfrom the pricing algorithm, including the opportunity cost of a seat.(3) Search data: We observe all consumer interactions on the airline’s website for twoyears. The clickstream data include search actions, bookings, and referrals from otherwebsites. Tracking occurs regardless as to whether an individual has a consumer loyaltyaccount or is logged in.(4) Filed fares data: The filed fares data contain the decisions made by the airline’spricing department. A filed fare contains the price, fare class, and all ticket restrictions,including any advance purchase discount requirements.(5) Forecasting data: The air carrier forecasts future demand at granular levels. Weobserve these predictions down to the flight-passenger type-price level. In addition to thebaseline forecast, we also observe all managerial adjustments to the forecasts.3.2Data SummaryTable 1 provides a basic summary of the nearly 300,000 flights in our cleaned sample. Wefocus on the last 120 days before departure due to the overwhelming sparsity of search andsales observations earlier in the booking horizon.9 Wedocument facts using nonstop bookings, however, our measure of remaining capacity adjusts for alltickets sold, e.g., connections, reward tickets, and consumers altering tickets, etc.9
Table 1: Summary StatisticsData SeriesVariableMeanStd. Dev. Median5th pctile95th pctileOne-Way Fare ( )Num. Fare ChangesFare Changes Inc.Fare Changes 09.0FaresBookingsBooking Rate-ODBooking Rate-AllLoad Factor (%)SearchesSearch RateSummary statistics for the data sample. Fares are for nonstop flights only. The initial load factor is the percentage of the number of seatsoccupied 120 days before departure. The booking rates are for non-award, direct travel on nonstop flights and for all traffic on nonstopflights, respectively . The number of passengers denotes the number of passengers per booking. The ending load factor includes allbookings, including award and connecting itineraries. The search rate is for origin-destination queries at the daily level. The number ofpassengers is the number of passengers per request.Average flight fares in our sample are 201, with large dispersion across routes andover time. Typically, prices for a particular flight adjust nine times and double in 120 days.Many adjustments occur at specified times, such as after expiration of advance purchasediscount opportunities. However, over 60% of price adjustments occur before the first APfares expires. This is because inventory (and therefore, prices) is re-optimized daily.In our sample, the average load factor is 82.2%.Although overselling is possible, weabstract from this possibility because we do not observe denied boarding/no show information. Our notion of capacity will be actual plane capacity plus the number of seats theairline is willing to sell over capacity (if any)—the observed “authorized” capacity.3.3Empirical Facts that Motivate Demand AssumptionsWe summarize search and purchase patterns to motivate some of our demand assumptions.The bookings data suggest that unit demand is a reasonable assumption. The averagepassengers per booking is 1.3. In addition, the bookings data confirm that overwhelmingly,consumers purchase the lowest available fare even though several fares may be offered atany point in time. We find that 91% of consumers purchase the lowest available fare. Using10
a separate data base that contains an indicator for corporate bookings under special fares,we find that corporate discounts are not a concern for the routes studied.Bookings and searches are sparse, which motivates using a model model that accountsfor low daily demand. 60-80% of observations involve zero observed searches. The fractionof zero sales is even higher (80% zeros). Zeros are not just present because we focus onnonstop demand. The fraction of zero sales for any itinerary involving a particular flightranges between 40-80%.We adopt a two-type consumer model, corresponding to “leisure” and “business” travelers, because that is how the firm considers demand. The airline maintains separate forecastsfor these consumer types, and an algorithm classifies every search and booking into thesetwo categories.10 We explore the predictions of this algorithm in Section 8.Figure 2: Search and Booking Facts to Motivate Demand r Re-Searching Same Itnerary0.0(c) Channel Booking DistributionsPercentage of Bookings(b) CDF of Similar Itin. SearchesCDFCDF(a) CDF of Same Itin. Searches0246810Number DDs Searched for a Given OD4.03.53.02.52.01.51.00.50.0DirectOTAsAgency120 10080 60 40 20Days From Departure(a) Empirical CDF of the number of days from departure searchers appear for a given itinerary. (b) Empirical CDF of the number ofdeparture dates a given searcher looks for. (c) Percentage of Bookings, across days from departure, for each Channel. Direct refersto bookings that occur on the air carrier’s website, OTAs is purchases made on online travel agencies, and Agency are bookings madethrough travel agencies.Figure 2-(a) and (b) motivate our assumption that consumers solve static optimizationproblems. We investigate the tendency for consumers to return to search for tickets for consumers who were not referred to the airline from other websites. Panel (a) shows the CDFof number of times that consumers search for the same itinerary across days. 90% of consumers search for an itinerary (OD-DD pair) once. Panel (b) shows the CDF for the numberof different departure dates (for the same OD) that consumers search for within a cookie.10 Theairline does consider additional types of passengers, but these categorizations are very small relativeto the two we consider. If we observe any searches or sales from other categories, we reassign them to beleisure travelers.110
82% of customers search a single departure date. The average time lag between searchesfor different departure dates is 45 days, which may suggest entirely different purchasingopportunities (different trips).Figure 2-(c) motivates adjusting our model for non-observed searches differently overtime. The figure shows the distributions of bookings across booking channels over time.OTAs, or online travel agencies, closely follows the distribution of bookings via the direct channel. However, the agency curve—which includes corporate travel bookings—ismore concentrated closer to departure. We discuss this adjustment in Section 6.1. Notethat Figure 2-(c) shows some bunching in bookings immediately before advance purchaseopportunities expire. Although this may suggest consumers strategically time their purchasing decisions—they are forward looking—we find evidence that supports certain daysbefore departure simply have higher demands. Using the search data, we split the sampleinto two groups, one that includes routes that never have 7-day AP requirements, and onethat includes these requirements. We find that that search activity (and purchases) bunch atthe 7-day AP requirement, regardless of their existence. Because arrivals increase regardless of price changes, we maintain the commonly used assumption that the market size isnot endogenous to price. Instead, we flexibly estimate arrivals as a function of time and thedeparture date that allows for this bunching.4Pricing Frictions Across Organizational TeamsIn this section, we document several pricing frictions that suggests that the firm does notprice optimally.4.1 Coarse Pricing and Not Responding to All Available InformationFigure 3 demonstrates that the firm has access to, and indeed generates, payoff relevantinformation that its pricing algorithms do not respond to. In panel (a), we plot the fraction of flights that experience changes in price or marginal costs (the shadow value on thecapacity constraint as reported by the pricing algorithm) over time. The figure shows that12
costs change at a much higher frequency than do prices. This occurs because of the industry practice of using a discrete set of fares (fare buckets). That is, it is possible thatmarginal costs change by 1 but the next fare is 20 more expensive. Our analysis suggeststhis friction is significantly more important. In panel (b), we run a flexible regression ofthe change in costs on an indicator function of a price adjustment occurring. As the figureshows, changes in marginal costs exceeding 150 only lead to price adjustments with 50%probability. This may suggest alternative fares could lead to higher revenues.Figure 3: Fare Adjustments in Response to Opportunity Cost Changes(a) Fare vs. Shadow Price Changes801.0Changes in FaresChanges in Shadow Values0.8ProbabilityPercentage of Flights100(b) Probability of Fare Change6040200Pr(Fare Change)0.60.40.2120
Although the sophistication of these technologies has improved, many of the original yield management ideas described in McGill and Van Ryzin (1999) and Talluri and Van Ryzin (2004)
