Sales-target-oriented rule-based dynamic pricing system design for airlines

Abstract

In today’s airline industry, the shift toward dynamic pricing—where prices evolve continuously rather than being fixed in fare classes—has become more than a trend; it represents a pivotal transformation in revenue management (RM). Yet, many existing systems still rely on legacy multi-fare-class reservation frameworks and conventional business logic, leaving them ill-equipped to handle modern demands. These systems often struggle with integrating competitive pricing data, capturing nuanced passenger behavior across web and mobile platforms, and reacting swiftly to real-time market shifts. More critically, they lack the capacity for processing the vast, granular datasets essential for artificial intelligence (AI)-powered revenue strategies. To bridge this gap, we developed a rule-based dynamic pricing algorithm that not only aligns with classic bid-price principles but also leverages both real-time data and the practical expertise of revenue managers. This algorithm is sales-target-driven and designed for responsiveness in live environments. Building on this foundation, we introduced a multi-layered software architecture tailored for airlines. Additionally, we offer practical recommendations for database structuring—both logical and physical—and propose streamlined business processes to enhance responsiveness under high computational loads and complex system integrations. Notably, a major Chinese airline has implemented our system prototype, marking a significant advancement in its RM capabilities.

Introduction

Over the past few decades, revenue management (RM) has become a cornerstone of airline operations, delivering tangible financial benefits and gaining widespread recognition across the global aviation industry. In practice, traditional leg-based seat inventory control—rooted in multi-fare-class management—has gradually been replaced by more refined origin-and-destination (O&D) pricing strategies that leverage the bid-price approach. More recently, with the rapid global rollout of the New Distribution Capability (NDC) initiated by the International Air Transport Association (IATA), a shift toward dynamic pricing models that operate independently of fare classes has captured the attention of airlines worldwide. This approach is widely regarded as the future direction of airline RM.

Despite significant theoretical advances, bringing dynamic pricing into real-world airline operations has proven difficult. The barriers stem from outdated system infrastructures and long-standing practices that are deeply embedded in current RM frameworks. Key challenges include:

(i) A heavy reliance on airlines’ internal historical data, limiting their ability to respond quickly to competitive changes in the market;

(ii) The growing influence of sales channels like online travel agency (OTA), direct websites, and mobile apps—alongside evolving passenger behaviors—makes real-time adaptation essential, yet difficult to achieve;

(iii) As flight frequency and route density continue to rise, systems must now meet stricter demands for real-time responsiveness;

(iv) The growing role of artificial intelligence (AI) in pricing makes the need for large-scale, fine-grained data collection unavoidable, posing significant technical and organizational hurdles.

Today, only a few global carriers—including Lufthansa, Finnair, American Airlines, Air France, and Koninklijke Luchtvaart Maatschappij (KLM)—have successfully embedded dynamic pricing into their core RM systems, thanks to early and proactive adoption of NDC standards. ¹ Most other airlines remain constrained by legacy computer reservation systems (CRSs) and instead rely on add-on modules to simulate dynamic pricing. In China, the three major airlines—Air China, China Southern, and China Eastern—each introduced O&D versions of the revenue management system (RMS) developed by U.S.-based Profit & Revenue Optimization Software (PROS) Corporation over the past two decades. Subsequently, more than a dozen other Chinese airlines followed suit, though many continue to depend on seat inventory–focused models. Yet, due to significant differences in China’s regulatory policies, competitive dynamics, and market maturity, implementing PROS’s RMS has proven difficult. Furthermore, critical technical components of the system—namely leg-level forecasting, bid-price calculations at the leg level, and O&D-based dynamic pricing—are proprietary and lack transparency, which limits their adaptability and precision in practical applications. Consequently, despite the significant initial investment in procuring the RMS, airlines often find themselves forced to allocate additional resources to tailor and extend the system in order to meet their specific operational needs.

Given this landscape, there is a growing consensus that Chinese carriers must develop homegrown dynamic pricing systems (DPSs) that are lightweight, practical, and well-suited to local operational practices. These systems should seamlessly integrate with existing CRSs, offer high computational efficiency, and be adaptable to the workflow habits of revenue managers. In response to this need, this paper proposes a sales-target-oriented, rule-based dynamic pricing algorithm tailored to the specific challenges faced by Chinese airlines. It also outlines a corresponding software architecture, database design, and computational optimization strategies aimed at delivering a high-performance, locally adaptable RM solution.

Related work

RM traces its roots back to the deregulation of the U.S. airline industry in 1978, which opened the door to more flexible and data-driven pricing strategies. ¹ Technically, RM evolved along two major lines: inventory control and price control. Early successes were largely driven by inventory control models such as EMSRb, which gained traction in practical airline applications.^2,3 However, as global air travel expanded and RM theory matured, the industry began shifting toward price control strategies—especially those based on bid-price logic. ⁴ With the rise of IATA’s NDC, the focus has further moved toward dynamic pricing models that allow prices to change continuously over time, decoupled from traditional fare classes. This shift reflects a broader transformation in how airlines aim to manage revenue, moving from static models to real-time, responsive systems.

According to Wittman and Belobaba, ⁵ dynamic pricing can be categorized into three main mechanisms: assortment optimization, dynamic price adjustment, and continuous pricing. Most existing implementations fall into the first two categories—well-aligned with fare-class-based CRSs—and have been supported by previous research.^6–10 In contrast, continuous pricing, which allows price setting to evolve moment-by-moment, offers greater flexibility and responsiveness but remains under active exploration. ¹¹ Despite a wealth of academic research on dynamic pricing,^4,12–14 real-world adoption still faces several hurdles. Most CRSs remain deeply rooted in fare-class structures and are not yet designed for large-scale dynamic pricing integration. As a result, airlines often rely on external software tools to bridge this gap. On the algorithmic side, traditional approaches use Markov decision processes (MDPs) to simulate passenger arrivals, incorporating different customer choice models and variations of the Bellman equation. Each arriving passenger is then matched with a suitable price point from a discrete set. While theoretically sound, these methods are computationally intensive and struggle to keep up with the real-time demands of today’s high-frequency flight environments, especially under volatile market conditions.

A typical RMS includes two key modules: demand forecasting and pricing strategy development. ¹⁵ The PROS system exemplifies this structure, applying bid-price logic at the leg level and enabling dynamic pricing at the O&D level via its real-time dynamic pricing (RTDP) module. However, the proprietary nature of these algorithms limits transparency and customization, preventing airlines from tailoring the system to their own operational realities. Meanwhile, the modern airline environment has grown more complex. Beyond flight schedules and historical demand trends, systems must now process real-time booking data, competitor activity, and user behavior across multiple digital sales channels—all under high-concurrency conditions.^16–19 Leading carriers like Lufthansa, British Airways, and American Airlines have already adopted dynamic pricing on select routes. For instance, Lufthansa uses NDC APIs for real-time pricing in intra-European markets and is working to expand this to other platforms. ²⁰ Low-cost carriers (LCCs) such as Ryanair, easyJet, and Southwest have also built dynamic pricing engines based on rules and historical data, gradually incorporating customer segmentation, time sensitivity, and market dynamics. ²¹ In contrast, most Chinese airlines still depend on O&D versions of the PROS system and have only limited dynamic pricing capabilities, particularly on domestic routes. The imported systems, while robust in international settings, often underperform in China due to differences in regulation, market behavior, and consumer expectations. Implementation challenges, closed-source algorithms, and the high cost of localization have discouraged full utilization. A few carriers, such as China Eastern Airlines, have attempted to build their own RMS platforms to better align with local needs. In general, Chinese airlines continue to lag behind their leading international counterparts in several key areas, including the implementation of dynamic pricing strategies, real-time system responsiveness, and seamless integration with existing CRSs.

Against this backdrop, the development of a localized DPS with full intellectual property rights has become both necessary and inevitable for Chinese airlines. These systems must handle diverse data sources, support rapid responses, reduce reliance on manual intervention, and incorporate expert knowledge from RM teams. In this paper, we introduce a sales-target-oriented, rule-based DPS designed to meet these needs—capable of real-time processing, multi-operator support, integration of historical and live data, and compatibility with both domestic and international market conditions.

Dynamic Pricing Algorithm

Algorithm design concept

In airline revenue analysis, two core metrics—revenue per available seat kilometer (RASK) and load factor (LF)—are commonly used to evaluate the performance of a given route. Building on firsthand insights from revenue managers at a major carrier, we developed a rule-based dynamic pricing algorithm specifically designed to align with sales targets. This algorithm emphasizes computational efficiency, rapid responsiveness, and real-time adaptability to both internal booking dynamics and external market competition. An overview of its conceptual framework is shown in Figure 1.

Figure 1.

Illustration of the sales-target-oriented rule-based dynamic pricing algorithm.

The process begins with defining sales targets, which include three main components: RASK, LF, and a minimum discount rate (MDR). Of these, RASK typically takes precedence as the primary control objective. On monopoly or strongly held routes, RASK alone may suffice as the key target. However, for competitive routes, a combined focus on RASK and LF is often necessary to strike a balance between revenue and market share. On underperforming or ”disadvantaged” routes—especially those experiencing unusually low sales—an MDR threshold is introduced to prevent the algorithm from generating overly aggressive discounts that could trigger a downward price spiral. Due to constraints in current CRSs, large-scale historical data cannot always be processed in real time. Instead, data is aggregated and analyzed at specific intervals known as data collection points (DCPs), which serve as control checkpoints for forecasting and adjusting future demand. ²² Sales targets are recalibrated at each DCP, and vary by route type: monopoly routes often carry high RASK expectations, competitive routes require adaptive targets that reflect market changes, and disadvantaged routes typically prioritize LF. External factors like holidays, schedule variations, and special events are also incorporated into the target-setting logic.

In the second stage, data pipelines—such as booking records, demand forecasts, and inventory updates—are handled differently depending on the airline’s existing RMS. Each DCP triggers a recalculation of bid-prices for flight legs, which remain in effect until the next DCP update.

Finally, when a customer initiates a booking request, the system fetches the current booking status of the flight and available inventory data from competitor flights. These inputs, together with the DCP-defined sales targets and precomputed bid-prices, are fed into a rule engine. This engine dynamically generates real-time pricing decisions, allowing the airline to respond instantly to market signals while remaining anchored to strategic revenue goals.

Sales target setting

In Figure 1, sales target setting forms the backbone of the proposed dynamic pricing algorithm. These sales targets are not static; they are adjusted across different DCPs based on several contextual factors—such as route type, the level of market competition, and seasonal or holiday-related demand variations tied to the flight’s departure date. Within each DCP cycle, our algorithm first estimates regression model parameters for the primary metric, RASK, using historical sales data drawn from flights with similar characteristics. To complement this, the LF target is set at the 70th percentile of the LF distribution derived from the same dataset, while the MDR threshold is defined at the 30th percentile. These initial values provide a data-driven foundation, but revenue managers are encouraged to apply their operational judgment and make targeted adjustments before finalizing the targets.

The selection of historical flight data for training the model adheres to strict criteria: the flight must match in day of week (DOW), seasonal pattern (i.e., peak, shoulder, or off season), and route classification (monopoly/dominant, competitive, or disadvantaged). For special cases such as national holidays or large-scale events, the dataset is categorized separately to capture atypical demand patterns. Each flight record in the dataset includes a range of descriptive variables—such as airline brand influence, flight schedule characteristics, share of route capacity, seat supply, number of seats sold, average LF, average discount rate, the lowest available fare offered by competitors, and the number of seats offered at that fare. These features are expressed as a tuple: DataSet(x₁, x₂, …, x₉), Using this structured dataset, a regression model is constructed to estimate the expected RASK value, as shown in equation (1):

$RASK={\beta }_0+\sum _{i=1}^9{\beta }_i\cdot x_i+ϵ$

(1)

where $({\widehat{\beta }}_1,{\widehat{\beta }}_2,\dots ,{\widehat{\beta }}_9)$ can be estimated using the least squares method on the DataSet and ϵ is the random errorterm in equation (1). The initial value of the RASK sales target at the current DCP can thereby be obtained.

Sales-target-oriented rule-based dynamic pricing algorithm

In Figure 1, the ”target-driven rule engine” forms the operational core of our proposed dynamic pricing algorithm. At each DCP, the engine makes real-time pricing decisions by aligning pre-defined sales targets with the latest sales activity. It leverages rule sets established by revenue managers and continuously incorporates dynamic inputs—both internal and external—to adjust prices in a responsive, data-informed manner. The overall process flow is illustrated in Figure 2.

Figure 2.

Flowchart of the sales-target-oriented rule-based dynamic pricing algorithm.

Algorithm inputs:

• Sales target, including object RASK (ObjectRASK), object LF (ObjectLF), and object MDR (ObjectMDR);

• Current booking volume of the flight, average discount rate of seats already sold, historical sales performance for the same period, and demand forecasting situation (the seat demand forecast for the period from the current DCP to the flight’s departure);

• Current and historical sales status of competing flights.

Algorithm outputs:

• Recommended current MDR (LD);

• Recommended sales volume constraint at the MDR (LAU).

In Figure 2, the algorithm follows a three-stage process, each designed to iteratively refine pricing decisions based on sales targets, market data, and real-time dynamics.

Stage 1: The algorithm begins by revisiting the initially defined sales targets. These targets—centered around RASK, LF, and MDR—are dynamically adjusted using the real-time sales data. Revenue managers retain the ability to make manual corrections based on their judgment and ongoing observations, ensuring that strategic intent is preserved even under rapidly changing conditions.

Stage 2: With the updated ObjectRASK and ObjectLF values in place, the system calculates the required sales volume and estimates a feasible minimum price. This step leverages the EMSRc model, ²³ which is applied to the remaining unsold seats. The model aggregates expected revenues in descending order, progressively including lower-fare seats until the sales targets are met.

Stage 3: Based on a comparative analysis of historical booking trends for the same period, current real-time sales performance, along with the competitors’ real-time lowest fares and their overall sales volume, the algorithm dynamically determines the MDR by incorporating the results from Stage 2 and the predefined MDR constraints. The EMSR model is then reapplied to determine the maximum number of seats that can be sold under the updated MDR threshold.

While the algorithm is primarily applied at the leg level, it can also support O&D pricing scenarios. In such cases, leg-level pricing outputs are aggregated using leg occupancy patterns to produce bid prices for complete O&D itineraries. One of the key strengths of this approach lies in its balance between revenue maximization and market responsiveness. By combining rule-based logic with real-time data inputs and human oversight, the algorithm avoids the computational overhead of traditional operations research methods, while still maintaining strategic flexibility and execution speed. This makes it both practica and adaptable for real-world airline RM environments.

Notably, a similar multi-stage, data-driven decision-making approachframeworks has have been applied in various fields to addressmanage complex uncertainties and optimizeenhance decision-making quality. ²⁴

System implementation and performance evaluation

Software system architecture design

Building on the dynamic pricing algorithm introduced earlier, we developed a layered software architecture (Figure 3), consisting of four functional tiers: the data processing layer, the predictive analytics layer, the decision analytics layer, and the results layer. This design accommodates a wide range of airline information technology (IT) environments, including those with an existing RMS and those operating without such infrastructure. For airlines equipped with an RMS, the architecture allows direct integration by importing forecasting results and bid-price outputs from the existing system. For those without an RMS, it provides a full-stack framework covering the entire workflow—from data ingestion and demand forecasting to bid-price calculation and pricing decision-making. Once the system has access to forecasting data and bid prices, either from internal modules or external systems, it incorporates real-time sales dynamics and passenger behavioral insights. These inputs are processed using the sales-target-oriented rule-based algorithm described earlier, enabling the system to generate dynamic pricing decisions in real time that align closely with operational goals and strategic revenue targets.

Figure 3.

Software system architecture design diagram.

In Figure 3, the system adopts a four-layer architecture, each tier playing a distinct role in the overall dynamic pricing workflow.

• Data Processing Layer: This foundational layer collects flight sales data from a variety of sources and performs data cleaning, integration, and alignment to transform raw or unstructured inputs into structured formats suitable for algorithmic modeling. Data alignment occurs at each DCP, capturing key metrics such as booking volume by fare class (including competitor data), cancellations during the DCP window, and changes in fare class availability. Special attention is given to DCP = −1, which refers to data collected 1 day post-departure, including final passenger counts, no-shows, go-shows, and denied boardings (DB). Additional data—such as full flight schedules, passenger name records (PNRs), and frequent flyer information—is used for passenger segmentation and estimating passenger choice (PSC) model parameters. Real-time inventory (INV) data provides the current number of bookings, cancellations, and fare class status, while contextual data like holidays and events are also included to support forecasting and pricing. The processed output is organized by flight, segment, and fare class, and stored in a structured database.

• Predictive Analytics Layer: Building on the cleaned and structured data from the previous layer, this module estimates passenger arrival parameters for each fare class using a range of models, including linear regression, additive models, log-based regressions, and exponential smoothing. These are optionally supplemented with modern AI methods for enhanced accuracy. The layer produces forecast parameters across three dimensions: segment, flight, and fare class. If the airline already operates PROS O&D systems, the forecast data can be imported directly from the existing platform.

• Decision Analytics Layer: This is the system’s algorithmic core. It takes inputs from the Data Processing and Predictive Analytics layers to compute bid prices, expected seat values, and protection levels for virtual fare classes. If PROS O&D is in place, bid-price results can be imported directly. These inputs, combined with real-time INV data from the CRS, are processed through the sales-target-oriented, rule-based algorithm to generate dynamic pricing decisions for each flight.

• Results Layer: This layer delivers the final output of the system—the recommended minimum discount—based on the calculations from the Decision Analytics Layer. The result is transmitted to the airline’s direct sales channels, such as its official website or app, enabling real-time application of dynamic pricing decisions in customer-facing environments.

Challenges in system design and proposed solutions

As the core of an airline’s marketing software suite, the DPS must interface with a variety of platforms—including the CRS, global distribution system (GDS), OTA portals, and multiple ticketing applications—while supporting a large user base and handling high-frequency interactions. At each DCP, the system is required to retrieve booking data from the CRS. As the number of flights increases, the volume of data grows rapidly, placing significant demands on both software performance and hardware capacity. In real-world operations, the DPS must respond to thousands of concurrent booking queries or purchase requests within seconds, which imposes strict requirements for system timeliness, responsiveness, and concurrent processing capability in a highly interconnected environment. To address these challenges, we implemented three key design strategies, as detailed below.

Adopting a Layered Architecture to Manage Complex System Interactions

The DPS is directly linked to the CRS, airline-owned sales channels, and OTA platforms via both web and mobile applications. To manage these complex integrations, the system adopts a layered software architecture (Figure 3). This design approach follows the principle of separation of concerns, using modular components and encapsulated interfaces to isolate different functional layers. As a result, the system achieves high cohesion and low coupling, which enhances maintainability and adaptability to evolving business requirements. The modular structure also improves code reusability and lowers both development and long-term maintenance costs.

Combining Logical and Physical Database Structures to Handle High-Density, High-Concurrency Operations

The database structure is carefully aligned with the layered architecture and is composed of four logical layers: foundation, organization, prediction, and decision. Data tables within and across these layers are linked via referential integrity rules, with higher-layer tables typically dependent on lower-layer data. Given the multi-dimensional nature of flight data—typically defined by combinations such as ”flight number, departure and arrival airports, DOW, and seasonal characteristics”—we designed multi-level composite indexes at the foundation layer. These indexes cover route-level, flight-level, flight-with-date, and flight-with-date-and-characteristics combinations, and are converted into unique long-integer IDs, which significantly enhance the speed of multi-table joins and high-frequency queries.

Operationally, airlines often segment their route networks into regions, with different RM teams handling distinct route groups. Forecasting and pricing tasks are typically executed at the flight level. Reflecting this business logic, the physical database allocates separate tablespaces to each route region. Each tablespace contains flight booking data, route-level booking summaries, forecasting parameters, forecast results, bid-price tables, and dynamic pricing outputs. To optimize database performance under high load, we implemented a partitioned storage strategy based on leg (O&D pair) and flight date. This approach minimizes full-table scans, reduces disk input/output (I/O) pressure, and dramatically improves both query and update efficiency. Moreover, it enables the system to store detailed passenger behavior data from both web and app sources, forming a data infrastructure ready for future AI-powered pricing applications.

Optimizing Computation Time to Meet Operational Deadlines

Consider a mid-sized Chinese airline operating flights between 60 cities and covering approximately 120 round-trip legs. Each day, the system must process data across 24 DCPs, covering 185 days of flight schedules. For each flight, this includes four virtual fare classes, competitor fare classes, leg-level demand forecasts, and remaining seat availability—resulting in extensive calculations. Dynamic pricing involves EMSRb and bid-price computations for each fare class, assuming an average of 100 unsold seats per flight. Before optimization, the total computational time for daily processing reached nearly 5 h. For instance, if data collection begins at 5:30 a.m., the system could not complete all calculations before 10:30 a.m.—failing to meet the required 8:00 a.m. readiness deadline.

Upon analyzing the algorithm workflow, we identified inefficiencies. Forecasting parameter recalculations—based on DOW and seasonal patterns—produced only four new data points per month. Thus, these calculations did not need to be performed daily. Additionally, competitor data was primarily used for manual post-analysis, allowing it to be calculated after the airline’s own forecasts. To improve efficiency, we revised the computational sequence as follows:

(i) Each night, the system sequentially performs the following tasks: data preprocessing, demand forecasting, and dynamic pricing computation for the airline itselft; followed by data cleaning and preparation for competing airlines, and their corresponding demand forecasting.

(ii) The recalculation of forecasting parameters is reduced to once per week, with the airline’s parameters updated on Saturdays and competitors’ on Sundays.

Through this set of design enhancements—including a layered system architecture, detailed logical and physical database structure design, and optimization of computational workflows—the DPS achieved substantial gains in system responsiveness, concurrent data handling, and task scheduling efficiency, making it both scalable and operationally reliable.

System performance evaluation

Following the careful design of both the system architecture and underlying database, as well as the implementation of the sales-target-oriented, rule-based dynamic pricing algorithm, the developed system has successfully met practical operational requirements—particularly in terms of processing speed and revenue performance. Compared with the original EMSR-based system that relied heavily on manual input and static logic, the new system has demonstrated measurable improvements, detailed as follows:

Time Performance Improvement

During the system preparation phase, the time required for completing dynamic pricing calculations per flight was reduced from 130 min to under 100 min. As a result, pricing tasks for the airline’s own flights could be completed before 7:00 a.m. In parallel, the statistical analysis and forecasting for competing airlines required approximately 45 min. Altogether, the system is now capable of achieving full operational readiness by 8:00 a.m. on a daily basis. In live operation scenarios—including high-concurrency periods—the system continues to exhibit strong real-time responsiveness. Pricing decisions for individual flights can be computed in under 3 s, ensuring seamless user interaction and rapid adaptation to booking dynamics.

Revenue Improvement

Revenue performance was validated using both empirical comparisons and controlled simulations. From an empirical perspective, flight revenue generated using the new dynamic pricing method was compared with that from the same period in the previous year, when the system had not yet been deployed. The analysis showed an average increase of 0.53% in total revenue per flight. In simulation analysis, a controlled environment was built using real operational data from the airline over the past two years. A total of 500 simulation runs were conducted across three typical routes: business, tourism, and standard. Compared to the EMSRb method (i.e., commonly used by airlines RM), the proposed dynamic pricing method resulted in average revenue increases of 3.23% on business routes, 8.06% on tourism routes, and 4.78% on standard routes, respectively.

Conclusion

This paper responds to the pressing demand among Chinese airlines for a practical and adaptable DPS. To address this, we proposed a sales-target-oriented, rule-based dynamic pricing algorithm that draws on both established theoretical foundations and the operational expertise of RM professionals. The algorithm integrates real-time sales dynamics, competitive market conditions, historical demand patterns, and data from multiple sources and distribution channels to deliver pricing decisions that are both timely and operationally effective. To support the implementation of this algorithm in a real-world airline environment, we further introduced multi-layered software architecture designed to handle high system complexity, large-scale data storage, and strict real-time performance requirements. Key technical components include the layered system architecture design, logical and physical database structure design, as well as an optimized computational workflow tailored for high-frequency pricing tasks. The prototype system developed under this framework has already been successfully adopted by a major Chinese airline as part of its RMS upgrade, demonstrating both the practical viability and commercial value of the proposed approach.