News

I set out to build a realistic simulation of an airline, something that could mimic the operational, logistics and business aspects as well. That meant not only creating some random routes and inventory. I wanted to simulate airline strategy: which cities should be connected? should an airline select? What would the operating costs be—from airport handling to fuel consumption? Where should catering stations be located to guarantee availability and service quality—especially for meals, whose quality degrades with time since preparation? And quickly, I ran into a basic, blocking question: I had no idea how many people would want to fly from Madrid to Hamburg in December.

That simple question opened a rabbit hole.

What drives passenger demand between two cities? Is it population size? Income? Distance? Seasonality? I started digging through datasets, reading old aviation studies, and piecing together what I could from open sources. Slowly, it turned into something bigger than a simulation detail—it became a standalone model. A way to estimate how many people are likely to fly between any two airports, in any given month, based only on fundamental variables like demographics and geography.

This is the story of how that model came to life.

At its core, the goal was simple: estimate how many passengers travel between two airports in a given month. But to do that, I had to decide which variables actually drive that behavior.

I started with a few intuitive ones:

• Month of the year — to capture seasonality. Travel demand isn’t constant across the year; it fluctuates with holidays, tourism cycles, and school calendars. This was encoded using monthly dummy variables.
• Population at the origin and destination cities — assuming larger population centers naturally generate and attract more travel.
• GDP per capita (PPA-adjusted) — as a proxy for economic capacity. The idea was that regions with higher income levels are more likely to support discretionary travel.
• Distance between airports — under the hypothesis that there’s an ideal range where air travel is most attractive: too short, and people drive or take trains; too long, and demand might taper off due to cost or complexity.

These variables fed into a Ridge regression model. I used standardization for numeric inputs and one-hot encoding for categorical ones. Ridge regularization helped avoid overfitting and encouraged the model to stay generalizable rather than memorize noise in the data.

Gathering the data

Before anything could be modeled, I needed to define what each variable actually meant in practical terms—and find real-world data to represent it. Population, for instance, sounds straightforward, but airports rarely serve a single city. A major hub like Frankfurt draws passengers from hundreds of kilometers away, while smaller regional airports might only cover a fraction of their immediate surroundings. So I had to translate “population” into something measurable and comparable across the globe.

Population

For that, I turned to GeoNames, a freely available geographic database that lists millions of populated places worldwide along with their coordinates and population estimates. The idea was simple: for each airport, estimate the total population living within a certain radius—essentially, the “catchment area.” People within that area would represent the potential users of that airport.

Defining the radius wasn’t trivial. A flat 50 km boundary would treat all regions equally, but in reality, access depends on infrastructure and geography. In densely connected regions like Western Europe, passengers might have several airports within a short distance; in sparsely populated areas, one airport might serve an entire province. To keep things manageable, I started with a fixed distance threshold and left adjustments for future refinement.

Using GeoNames data, I aggregated populations of all cities and towns located within the chosen distance from each airport’s coordinates. The resulting figure approximated the “population served” by each airport—both for the origin and the destination in every city pair. This variable, while imperfect, became a crucial foundation: it translated abstract geography into a concrete measure of potential demand.

Airports

The next challenge was compiling a reliable and consistent database of airports. At first glance, this might seem simple—after all, airports are well-documented infrastructure. But in practice, it turned out to be one of the most tedious parts of the process. The reason is that there’s no single, universal standard for identifying airports. The most commonly used systems—ICAO and IATA codes—both fall short of true uniqueness. Airports can change their codes over time, and in some cases, entirely different facilities may share similar identifiers.

To assemble a usable dataset, I had to combine multiple public sources:

• OSM World Airports (https://www.openstreetmap.org)
• OpenFlights (https://openflights.org/data.html)
• ICAO public data (https://www.icao.int)
• PyPI airport-related packages (https://pypi.org)

My approach was incremental. I began with OpenStreetMap (OSM) as the base dataset, since it provides the most complete geographic coverage, including smaller regional airports often missing from official records. Then, I progressively enriched and corrected this dataset using the others as complementary sources.

Identifying duplicates was a particular headache. I used a combination of the IATA code, ICAO code, and geographic distance to detect likely matches. For the distance component, I calculated the great circle distance between airport coordinates—the shortest path along the surface of the Earth—using the standard haversine-based formula. This metric worked well to detect airports that were clearly the same facility despite slightly differing coordinate references.

However, the method wasn’t foolproof. Large airports can span several kilometers, and not all datasets reference the same point within an airport’s perimeter. At the same time, entirely separate airports can exist just a few kilometers apart. This made automatic matching unreliable in edge cases. Whenever ambiguity arose, I resorted to manual verification using Wikipedia, cross-checking names, locations, and served cities. In those cases, I manually updated or corrected the records to ensure each airport in the final database truly represented a distinct facility. The end result was a consolidated, cleaned dataset that became the geographic backbone of the model.

Economic Capacity

The next variable I needed was a measure of economic capacity—something that could explain differences in travel demand beyond simple population size. For that, I used GDP per capita adjusted by purchasing power parity (PPA), which accounts for differences in local price levels and better reflects real income and consumption potential. The data came from the OECD database (https://data.oecd.org/gdp/gross-domestic-product-gdp.htm), which offers consistent and comparable figures across many countries.

The dataset covered 292 national entities, including both full countries and territories. However, not all of them had complete or up-to-date data. Some entries stopped reporting altogether—often a telltale sign of governments discontinuing the publication of official economic statistics when performance deteriorates. As a result, a few regions had data gaps or outdated values.

To keep the model consistent, I took a pragmatic approach. For each country, I used the latest available data point, regardless of year. Since the model’s purpose was to capture structural economic differences rather than short-term fluctuations, I assumed that temporal variation in GDP per capita would not exceed ±5% per year. This simplification allowed me to maintain coverage without introducing artificial noise or unrealistic extrapolations.

Each airport was then associated with the GDP per capita value of its corresponding country. This step introduced a coarse-grained but valuable proxy for the propensity to travel, especially for discretionary trips such as leisure or business travel beyond necessity. Together with population and distance, it added an economic dimension to what would otherwise be a purely geographic model.

Number of passengers carried

To train and validate the model, I needed a reliable measure of how many passengers actually travel between airports. For that, I relied on data from Eurocontrol, specifically the avia_par dataset, which provides detailed aviation statistics for European countries (https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Air_passenger_transport_statistics). The data is organized as one file per country, each containing a set of related series such as total flights, number of passengers, and cargo volumes. In total, I downloaded and processed 39 national files.

Each file included several variations of passenger statistics—namely “on board” and “carried” passengers. The distinction between them is important:

• On board represents the total number of passengers physically on the aircraft, including those in transit who remain onboard during stopovers.
• Carried refers only to passengers who either boarded or disembarked at the airport in question, which better reflects real origin–destination activity.

For the purpose of this model, I selected “carried” passengers. This choice eliminated the double counting that occurs with transit passengers and aligned the data with the concept of city-pair demand—that is, how many people actually begin or end their journey at each airport.

Storing and Processing the Data

I developed a C# program to automate the ingestion and parsing of the 39 country files. The script read each dataset, normalized field names, and stored the relevant data in PostgreSQL tables. The most relevant dimensions were origin airport, destination airport, month, and number of passengers carried. In many cases, datasets included separate figures for departing and arriving passengers, which allowed me to reconstruct both directions of travel between airport pairs. When data for one side of a route was missing—for example, when no U.S. dataset existed—I could still estimate transatlantic flows by using the arrival data from European airports. In other words, if 50,000 passengers arrived in Frankfurt from “non-European origins,” I treated that as an approximation of the outbound traffic from those same origins to Frankfurt.

Intra-European routes added another layer of complexity. Some countries reported passenger flows both from the origin and from the destination, sometimes with discrepancies between the two. When such inconsistencies appeared, I adopted a consistent rule: the origin data prevailed, as it more directly reflects the actual number of passengers boarded on flights departing from that airport.

The end result was a harmonized dataset of passenger flows that covered most of Europe and its major international connections. It provided a realistic picture of observed demand between airports—one that could be compared against the model’s predictions later on. Despite gaps outside the Eurocontrol network, this dataset was sufficient to establish a strong empirical basis for training the model and validating its assumptions about what drives passenger demand between any two cities.

To store and manipulate the data, I selected PostgreSQL because it offers a powerful relational database with geographic capabilities through the PostGIS extension (https://postgis.net). All datasets—airports, cities, populations, and GDP per capita—were imported into separate tables. PostGIS allowed me to store coordinates as spatial data and perform geographic queries directly, such as calculating great circle distances between airports or aggregating populations within a defined radius. This setup provided a consistent and efficient environment to prepare the data needed for modeling.

All variables were consolidated into a single database view containing, for each airport pair, the origin and destination airports, the population within each airport’s catchment area, the GDP per capita of their respective countries, and the distance between them. This view provided a unified, queryable structure that combined demographic, economic, and geographic factors into a single dataset ready for modeling.

The model

I started with a pretty simple model: it used a logarithmic Ridge regression, applying the natural logarithm to all continuous variables—population, GDP per capita, and distance—to linearize multiplicative relationships and reduce skewness. This approach better captured how passenger volumes scale with underlying factors: doubling a city’s population doesn’t double its air traffic, but it increases it proportionally on a log scale. By regressing the log of passengers carried against the log-transformed predictors, the model estimated elasticities that describe how sensitive demand is to each variable while maintaining interpretability and numerical stability:

$${\small \begin{aligned}\log(\hat{Y}_{ij})&=\beta_{0}+\beta_1\log(Pop_i)+\beta_2\log(Pop_j) \\&+\beta_3\log(PPP_i)+\beta_4\log(PPP_j) \\&+\beta_5\log(Dist_{ij})+\beta_6Year \\&+\sum_{m=2}^{12}\gamma_mMonth_m+\epsilon_{ij}\end{aligned}}$$

where:

\({\small \hat{Y}_{ij}}\) is the estimated number of passengers between airports i and j.

\({\small Pop_i,Pop_j}\) are the populations around origin and destination airports.

\({\small PPP_i,PPP_j}\) are GDP per capita adjusted by power parity at origin and destination.

\({\small Dist_{ij}}\) is the distance between the two airports.

\({\small Year}\) starts in 1 with the first datapoint (1993).

\({\small M_m}\) is a monthly dummy variables capturing seasonality.

\({\small Inter_{ij}}\) is a dummy variable capturing the interaction between origin and destination.

\({\small \beta_k,\gamma_m}\) are the model coefficients.

\({\small \epsilon_{ij}}\) is the model residual.

The model above produced quite disappointing results, with an R² of around 0.13. While it captured general trends driven by population, income, and distance, it failed to explain much of the real variability in passenger flows. The reason quickly became clear: air travel demand isn’t purely a function of geography and economics. What truly shapes the intensity of connections between cities are the interactions between them—the invisible network of social, cultural, and economic ties that drive people to move.

These interactions include a wide range of factors. Cultural and linguistic affinity often makes travel more frequent between certain countries, as do historical and colonial ties, migration patterns, and economic relationships such as trade or tourism partnerships. Even shared business ecosystems—like finance between London and Frankfurt, or technology between Helsinki and Tallinn—create travel patterns that simple demographic variables can’t predict. In short, the model treated cities as independent points on a map, but in reality, they’re deeply connected through human and economic networks.

To address this, I introduced interaction dummies that represent these latent links between airport pairs. They act as fixed effects for pairs that have historically shown above- or below-average demand relative to what the baseline model predicts. While this doesn’t explicitly model each underlying cause, it allows the regression to partially learn and account for those asymmetric relationships, greatly improving the realism of the simulated demand patterns. The complete model can be described with this equation:

$${\small \begin{aligned}\log(\hat{Y}_{ij})&=\beta_{0}+\beta_1\log(Pop_i)+\beta_2\log(Pop_j) \\&+\beta_3\log(PPP_i)+\beta_4\log(PPP_j) \\&+\beta_5\log(Dist_{ij})+\beta_6Year \\&+\sum_{m=2}^{12}\gamma_mMonth_m \\&+\sum_{ij}^{i\not=j}\delta_{ij}Inter_{ij}+\epsilon_{ij}\end{aligned}}$$

where:

\({\small Inter_{ij}}\) is a dummy variable capturing the interaction between origin and destination.

\({\small \delta_{ij}}\) is the coefficient of the Interaction variable.

With the interaction variables included, the model’s performance improved dramatically. The final version achieved an R² of about 0.65, compared to just 0.13 before, showing that it now captures most of the explainable variation in passenger flows. To interpret these results properly, it’s useful to understand what each metric reflects:

• RMSE (Root Mean Squared Error) measures the typical difference between predicted and actual values—in this case, in the logarithmic scale of passenger counts. Lower RMSE means predictions are closer to reality. The model’s RMSE_test_log = 0.68 indicates relatively tight clustering around the observed values after taking logs, a strong sign of improved accuracy.
• R² (Coefficient of Determination) indicates how much of the variance in the data the model explains. An R²_test_log of 0.65 means the model accounts for roughly 65% of the observed variability in passenger demand between airport pairs—an excellent fit for this type of behavioral, aggregate data.
• MAPE (Mean Absolute Percentage Error) expresses the average prediction error as a percentage of the actual value. However, because air traffic is heavily skewed—with many small routes and a few extremely busy ones—MAPE tends to overemphasize errors on low-volume routes. The model’s MAPE_test ≈ 307% seems high, but that reflects the distribution’s asymmetry rather than poor performance overall.
• sMAPE (Symmetric Mean Absolute Percentage Error) offers a more balanced measure by normalizing errors with respect to both predicted and actual values. Here, the sMAPE_test ≈ 37% gives a more realistic sense of scale: on average, the model’s predictions differ from reality by about one-third, which is quite reasonable given the inherent volatility and unobserved behavioral factors in air travel data.

Together, these metrics show that the enhanced model not only fits the data far better but also generalizes well, capturing the structural drivers of passenger demand while tolerating the unavoidable randomness in real-world mobility patterns.

To understand how each variable contributed to the model’s explanatory power, I ran a sensitivity analysis by training alternate models, each time removing one variable and measuring the change in test R². The results were:

Variable Removed	R2	ΔR2
none	0.639867	0.000000
distance	0.639831	0.000037
ppp	0.639866	0.000001
pob	0.638180	0.001687
od	0.129997	0.509870
month	0.598472	0.041395
year	0.633293	0.006574
ppp,distance	0.639827	0.000040

The sensitivity analysis confirmed what intuition already suggested: the interaction variables are by far the most influential. Removing them caused the R² to collapse from 0.64 to just 0.13—a loss of nearly 0.5 points—showing that socio-economic and cultural links between cities dominate travel behavior.

Other variables had much smaller effects. Population showed a modest impact, confirming its role as a broad demand driver, while GDP per capita and distance contributed little once the interactions were accounted for. Month dummies were still important, explaining around 4% of the variance through seasonality patterns, while the year variable added a smaller but consistent improvement.

Overall, this experiment validated that structural connectivity between cities—more than pure size or wealth—defines modern air travel demand.

Limitations

Although there are more than 92 million possible airport pairs worldwide, my model only covers around 24,150 routes. Not only because air travel is inherently sparse, but because free, reliable, consolidated data simply doesn’t exist for most of the world. I based the model on Eurostat’s dataset, which is one of the very few publicly available and structured sources that report passenger flows between airports at a monthly level. Extending this to global coverage would mean sourcing, cleaning, and validating passenger data from over 300 national and territorial aviation authorities—a task that is not only monumental but practically beyond human capacity without a dedicated international effort. Just identifying which datasets exist, what formats they use, and how trustworthy they are is a research project in itself.

It’s important to note that the model’s performance drops significantly when predicting routes not seen during training. For these out-of-sample airport pairs, the R² falls to around 0.13, meaning the model captures very little of the variance in demand. This suggests that while the model generalizes well to known routes, it struggles to extrapolate demand for entirely new origin–destination combinations. This isn’t surprising—route-specific dynamics are captured through dummy variables, and without historical representation, the model has no “memory” of those pairs. It’s a clear reminder of the importance of route-level information in demand modeling.

Comparison with Related Work

A relevant reference point for this project is the 2020 paper “Demand for Air Travel in the United States: Bottom-Up Estimates and Implications for Forecasting” by Serowsky, Leung, and Mahmassani (Northwestern University). While focused on domestic air travel within the United States, their approach shares important similarities with my work—particularly in the goal of estimating passenger demand between airport pairs using exogenous variables. However, the differences in scope, modeling choices, and data availability reflect important trade-offs and challenges in demand modeling.

Both studies share the intent of building predictive models from the ground up (“bottom-up”): instead of projecting from aggregate totals, we independently estimate demand at the airport-pair level. This is critical for capturing the heterogeneity in origin-destination dynamics and understanding the role of structural variables such as population or economic activity. In both cases, airport-pair-level demand is not treated as noise around a global trend but as something systematically explainable based on attributes of the route and context.

However, there are key differences in methodology and ambition. The Northwestern study had access to comprehensive, high-quality data from the U.S. Bureau of Transportation Statistics (BTS), including monthly enplanement data for over 18,000 domestic pairs. This is far richer than the dataset I worked with, which includes just 24,150 airport pairs extracted primarily from the Eurostat avia_par database. Importantly, my dataset is inherently biased toward Europe and lacks coverage for many parts of the world—simply because no open-access, harmonized global dataset exists. The challenge of identifying consistent passenger statistics across more than 300 national and territorial aviation systems proved to be insurmountable for a single researcher. Thus, while their dataset covers nearly all active domestic routes in the U.S., mine captures only a tiny fraction of global passenger flows.

Methodologically, the Northwestern study used a generalized linear model (GLM) framework, testing several specifications including log-linear models with fixed effects and interaction terms. In contrast, I adopted a Ridge regression approach with extensive use of one-hot encoding for categorical variables like route and month. The rationale was practical: Ridge regression handles multicollinearity gracefully and produces interpretable coefficients even with high-cardinality categorical inputs like 10,000+ airport pairs. Furthermore, I intentionally avoided temporal fixed effects (like year dummies) in the final version of my model to ensure generalization to future years.

In terms of explanatory variables, both models incorporate population and GDP per capita, though the U.S. model benefits from fine-grained county-level data while mine relies on public international sources like World Bank and OECD. One of the most striking differences is in the treatment of distance: the Northwestern study finds a strong, typically negative, log-linear effect of distance on demand. In my model, I experimented with several distance transformations, including logarithmic and asymmetric bell-shaped curves, but ultimately found that distance added very little explanatory power once route-level fixed effects were included.

Predictive performance also differs. My final model, applied to held-out data, achieves an R² of 0.64 and a sMAPE of ~40%. This is strong considering the heterogeneity and sparsity of the international data. However, this level of accuracy was only achieved for routes already present in the training dataset. For routes unseen during training, the predictive performance drops sharply, with R² estimates around 0.12—something that was less of a concern in the U.S. study given their more complete route coverage.

In conclusion, while the work by Serowsky et al. benefits from a more complete dataset and more detailed regional economic information, my model offers a flexible, extensible structure suitable for global analysis under severe data limitations. Both works illustrate the value of bottom-up modeling, and together they suggest that scalable, route-level demand estimation remains a promising avenue for both academic research and aviation applications.

Compared to the study by Mao et al. (2015), which investigates air passenger demand in China using economic indicators and advanced spatiotemporal models, my work takes a different approach both in scope and methodology. Mao’s study is focused primarily on domestic air traffic within China, incorporating data from national statistics to explore how demand responds to urban hierarchy, income distribution, and network centrality. Their methodology leverages gravity models enhanced by spatial autocorrelation and policy factors like high-speed rail competition, producing insights that are tightly bound to China’s unique infrastructure and policy environment.

In contrast, my model is based on a much broader geographic scope, encompassing international airport pairs primarily derived from Eurostat data, and including any non-European airport as long as it is connected to a European one. Instead of focusing on domestic policy or internal connectivity, my objective was to build a generalizable, data-driven estimator of monthly passenger volumes based on fundamental variables like city population, air route identifiers, seasonal patterns, and limited economic indicators.

Crucially, while GDP per capita in purchasing power parity (PPP) was included in the initial design of my model—as inspired by studies like Mao et al.—sensitivity analysis revealed that it contributes virtually no predictive power. Its removal caused no significant loss in the model’s performance. Distance, another common component in spatial demand models, also had negligible effect, with near-zero change in R² upon exclusion. The dominant factors in my model turned out to be the identifiers of the origin-destination airport pairs (as dummy variables), along with year and month, which together accounted for most of the model’s explanatory power.

While Mao’s study benefits from high-resolution economic data within a single national framework, mine operates under the constraint of fragmented international statistics and yet is capable of explaining over 63% of the variance in log-transformed passenger volumes (R²_test ≈ 0.64). This is remarkable given that the model handles over 24,000 routes and over a million observations, with only basic geographic and demographic inputs. Whereas Mao’s work provides rich policy insight for a national context, my model aims to be a practical forecasting engine usable in simulations, operational planning, or exploratory aviation analytics beyond any single country.

Mao et al. (2015) and the present study share a common objective—estimating passenger flows between airports—but diverge sharply in scope, methodology, and purpose. Mao’s work, developed at the University of Florida, aimed to model monthly global air passenger flows by refining an existing gravity framework into a temporally resolved model. It used Poisson regression to estimate flows between directly connected airports in 2010, integrating a wide range of explanatory variables: demographic, economic, climatic, and network-based (e.g., population, purchasing power, airport capacity, betweenness centrality, temperature, humidity, and distance). The result was a global, high-level view of the world airline network’s monthly evolution, optimized for completeness and public accessibility rather than predictive precision.

In contrast, my model pursued a narrower but more data-grounded goal: to estimate actual passenger volumes between observed airport pairs using open European data, focusing on predictive performance rather than synthetic completeness. Where Mao relied on modeled passenger flows derived from partial U.S. data and OAG capacity figures, my model trained directly on Eurocontrol’s avia_par dataset—real observed passenger movements. The dataset, while geographically limited to Europe and connected routes, provided verifiable ground truth and monthly granularity that enabled statistical validation.

Methodologically, Mao’s approach followed the classical gravity paradigm—multiplicative relationships between population, economic potential, and distance, transformed through Poisson or log-normal regression. My model began similarly, with a logarithmic Ridge regression on population, GDP per capita, and distance, but departed by introducing high-dimensional interaction dummies for airport pairs. These variables captured latent socio-economic, cultural, and migratory ties—elements Mao’s model treated only indirectly through network metrics. The effect was substantial: R² rose from 0.13 to 0.65, showing that the structure of intercity relations, not just mass or proximity, drives air travel behavior.

Mao’s model included meteorological and network measures like temperature and centrality, achieving strong qualitative insights but limited transferability; its validation relied primarily on correlation with annual airport statistics (r > 0.95). My Ridge model, instead, prioritized generalization and interpretability: it explicitly estimated elasticities for each explanatory variable and used cross-validation over thousands of real airport pairs, achieving RMSE_test_log ≈ 0.68 and sMAPE ≈ 37%. The sensitivity analysis confirmed that population exerted moderate influence, while GDP per capita and distance had negligible impact once interaction effects were introduced.

Conceptually, Mao’s study focused on mapping dynamics—how the world airline network evolves monthly—while mine focused on understanding determinants—why passenger volumes between specific cities take the values they do. Mao’s framework produced synthetic data useful for global modeling in fields like epidemiology, but not necessarily for forecasting real traffic on unseen routes. My model, in contrast, produces empirically grounded predictions suitable for operational and strategic airline simulations, though limited to the subset of routes represented in training data.

In essence, Mao’s work advanced the spatio-temporal modeling of global air travel, while my study advanced its predictive econometrics. His model emphasized completeness and accessibility; mine emphasized accuracy and interpretability under real-world data constraints. Together, they illustrate complementary facets of air travel modeling—one describing the shape of the global network, the other its behavioral underpinnings.

Conclusion

What began as a missing parameter in a simulation—how many people would fly from one city to another—ended up becoming a research project of its own. The final model, while built from open data and simple principles, achieved a level of realism that makes it genuinely useful for simulation and forecasting. It showed that air travel demand cannot be explained by geography or income alone: it is fundamentally social, shaped by cultural proximity, economic ties, and shared history between regions.

There are still clear limitations. Coverage remains partial, restricted mainly to Europe and connected airports, and the model performs poorly for completely unseen routes. But the framework is extensible. With better access to passenger statistics and richer representations of intercity relationships—trade flows, migration, tourism, and digital connectivity—it could evolve into a global estimator of air travel demand, open and transparent.

In the end, this was never just about predicting numbers. It was about understanding what connects people across distances—and turning that understanding into something measurable, reproducible, and, hopefully, useful for how we design the next generation of airline systems.