How AI Could Improve Airline Network Forecasting

By Jeff Popple

Artificial Intelligence is rapidly changing how businesses make decisions, and airline network planning is no exception. But will AI replace traditional passenger forecasting models, or will its greatest value come from making those models better?

I’m certainly not a data scientist, but having spent many years working in airline network forecasting, I’ve been thinking about how AI might improve the way airlines evaluate network changes.

The Limitations of Traditional Forecasting

Most airline network forecasting today relies on passenger choice models using QSI or Logit methodologies. These models evaluate itineraries based on characteristics such as elapsed time, departure time, aircraft type, airline presence, and connection quality.

One of their greatest strengths is explainability. When a forecast predicts a change in passenger share, an analyst can generally identify the factors driving that change.

However, these models have limitations.

Calibration can be time-consuming, and models risk becoming overfitted to historical behavior that may not reflect future economic or competitive conditions.

More importantly, traditional passenger choice models generally distribute a predetermined number of passengers across available itineraries. They don’t inherently account for changes in total market demand or fares resulting from new service, additional capacity, or increased competition.

Market stimulation, particularly when introducing the first nonstop service between two cities, often requires separate assumptions and considerable analyst judgment.

Where AI Could Make a Difference

AI and machine learning offer opportunities to address these limitations.

Techniques such as gradient-boosted trees and hierarchical time-series models could improve O&D demand forecasting by incorporating economic conditions, fares, capacity, competitive dynamics, and other market characteristics.

AI could also help predict market stimulation by identifying patterns from previous network changes. Rather than relying primarily on broad assumptions, models could evaluate how factors such as carrier type, frequency, and travel time improvements influence demand.

Machine learning may also help identify groups of markets with similar passenger behavior, improving how traditional QSI or Logit parameters are applied.

Of course, AI introduces its own challenges. Complex models can be difficult to explain, and a forecast that cannot be understood or trusted has limited value when making multimillion-dollar network decisions.

Techniques that improve explainability and distinguish causation from correlation will be critical.

Moving Beyond Point Forecasts: Understanding Risk and Uncertainty

Another potentially significant advancement is the use of probabilistic forecasting.

Traditional network forecasts typically produce a single estimate of expected passengers, revenue, or profitability. But anyone who has worked in network planning knows that the actual outcome rarely matches the forecast exactly.

Probabilistic approaches, including Bayesian models, can provide a distribution of possible outcomes rather than just a single number.

Instead of forecasting that a new route will generate $20 million in annual revenue, a probabilistic model might estimate a range of potential outcomes, along with the likelihood of achieving each.

This distinction is particularly important when making investment decisions.

Two proposed routes might have the same expected profitability but very different risk profiles. One might offer relatively predictable returns, while the other has greater upside potential but also a much higher probability of losing money.

Article content
Two routes. The same expected profitability. Very different risk profiles.

Understanding those differences allows planners to evaluate not just expected returns, but the risks associated with committing scarce aircraft and capital.

The objective should not simply be to predict the most likely outcome, but to understand the range of possible outcomes and make better decisions under uncertainty.

The Future Is Probably a Hybrid Approach

I don’t believe traditional passenger choice models are going away.

They remain valuable for understanding how passengers select among competing itineraries, particularly for near-term schedule changes.

Instead, I envision a hybrid approach where AI improves forecasts of total market demand, fares, and stimulation, while passenger choice models determine how that demand distributes across available itineraries.

Probabilistic methods could add another dimension by quantifying uncertainty around those forecasts and helping planners evaluate the risk associated with different network alternatives.

The appropriate methodology will depend on the planning horizon and the type of network change being evaluated.

The Bigger Opportunity: Network Optimization

While AI could significantly improve forecasting accuracy, I believe its greatest potential lies in something even more valuable: the ability to evaluate hundreds of network alternatives quickly.

One of the biggest limitations in traditional network planning is the time required to construct and forecast individual scenarios. As a result, airlines often evaluate only a fraction of the possible network alternatives.

Imagine combining AI-enhanced demand forecasting, passenger choice models, probabilistic forecasting, and optimization into an engine capable of rapidly evaluating different routes, frequencies, departure times, and aircraft assignments.

Instead of simply forecasting the impact of a proposed schedule change, airlines could systematically search for the network configuration that produces the best overall result, while accounting for the uncertainty and risk associated with each alternative.

This could fundamentally change the decision-making process. Rather than selecting the network with the highest forecast profitability under a single set of assumptions, planners could identify networks that deliver attractive returns across a range of potential market conditions.

The best network isn’t necessarily the one with the highest point forecast. It may be the one that offers the best balance of expected profitability, risk, and resilience.

Final Thoughts

I don’t see the future of airline network forecasting as AI versus traditional models.

I see it as combining the strengths of multiple approaches.

Traditional passenger choice models provide structure and explainability. AI offers opportunities to better predict changing demand, market stimulation, and complex relationships. Probabilistic models help quantify uncertainty and evaluate risk.

But the real breakthrough may come when these capabilities are integrated with optimization, allowing airlines to move beyond simply forecasting network changes to identifying better network decisions.

The most exciting application of AI in network planning may not be producing a more accurate point forecast. It may be helping airlines make better investment decisions and ultimately build better networks.

I’d be interested in hearing how others (especially those who know more about data science than I do) in airline network planning see AI changing the forecasting and decision-making process.


 

About the Author:

Jeff Popple is Director of Customer Success at Zulu Airline Systems.