Conclusion
Non-Technical Summary
The analysis of U.S. flight delays shows that delays became noticeably worse after the COVID-19 pandemic but have slowly improved in recent years. The aviation system is stabilizing overall, but delays are still common at certain airports and certain airlines. Busy travel periods such as summer and major holidays also continue to contribute to delays. To better understand and predict delays, several machine learning approaches were tested and provided different insights. The trend model confirmed that delays increased sharply right after COVID and have gradually decreased over time. The decision tree model showed that using only basic pre-departure information, such as airline, airport, and time of day, makes it possible to identify flights that will depart on time in most cases. However it still remains very difficult to accurately identify the flights that will end up delayed. The random forest model performed much better when predicting late-aircraft delays, particularly when each departing flight was linked to the arrival of the same aircraft earlier in the day. This reflects the real-world pattern that a late arrival often causes the next flight to depart late as well. Clustering and association rule mining added further context by identifying groups of airlines and airports with similar delay behaviors and revealing that delays are more common during certain times of day, such as evening hours.
Key Insights and Discoveries
Across all models, flight delays are partially predictable and highly context dependent. At the system level, linear regression showed a clear post-COVID spike in average monthly delays, followed by a gradual downward trend, suggesting the aviation system is still recovering but slowly stabilizing. At the individual flight level, simple pre-departure information like carrier, airport, and scheduled time, were not enough to reliably predict which flights would be delayed. The Decision Tree model mostly learned to identify on-time flights and missed many true delays. However, when the problem was narrowed to late aircraft delays and incorporated paired arrival–departure data and engineered features like scheduled turnaround time, Random Forest models performed substantially better. Unsupervised methods added another layer of insight. K-Means clustering revealed distinct groups of airlines and airports, separating relatively on-time operators from those with systematically higher delay rates, while Apriori rules confirmed that evening departures and certain carrier–airport combinations are more delay prone than the baseline.
Explaination of Real-world Impact of Results
These findings can inform how airlines, airports, and even travelers think about delay risk. The post-COVID trend analysis highlights that delays are not just random but follow structural patterns over time. This can support long-term planning around staffing, crew availability, and infrastructure. The late-aircraft models show that propagation delays or in other words when one late flight causes another, are more predictable than generic delays. This suggests that airlines could use similar models to prioritize gate turns, add buffer time for at-risk rotations, or proactively rebook passengers when an inbound flight is severely late. Clustering and frequent-pattern mining provide practical guidance on where and when to expect trouble. The delay-prone airports, carriers, and evening time windows could be targeted for extra staffing or improved ground operations. Even if perfect prediction is not always viable every delayed flight, these models highlight high-risk contexts where interventions and communication are likely to mitigate delays.
Discussion of Limitations, Improvements, and Future Work
The analysis is constrained by several important limitations. This analysis worked with a restricted sample and did not have real-time operational data such as detailed weather conditions, crew scheduling, or air-traffic control constraints. Additionally, the data is imbalanced with more on-time flights than delayed ones, hurting classification performance and simple models overpredicting the majority class. Finally, interpretable baseline models like Decision Trees, Random Forests, K-Means, Apriori were mainly used and did not explore more complex architectures models. Future work could incorporate external weather APIs and operational data as well as implement advanced models. Future work could also design user-facing visualizations or decision tools for airport personnel and travelers.