Flight delays cost airlines
In the United States, more than 8 million commercial flights take to the air each year. 25% of those flights experience delays, an ongoing problem that costs airlines more than $20 billion yearly. Although airlines, airports, and researchers often attempt to accurately predict and mitigate delays, they have been unable to customize and retrain their predictions as available data changes and grows.
Although predicting arrival delays of flights is a widely researched topic, there exist challenges in application of models in real world scenarios due to following shortcomings:
- Ability to predict most likely cause of delays which can be used effectively to take mitigating actions
- Ability to predict flight delays anywhere from 2 weeks to 6 months ahead to be used by Airlines or Airports for better planning and scheduling of flights
- Individual flight delays predicted across a wide range of Airports
- Ability to customize and retrain the models in quick time through integration of internal and external data
One airline seeks a new solution
Hoping to gain accurate data to make real-time flight decisions and reduce delays, one airline decided to turn to AI. They wanted to predict individual flight delays across a broad range of airports and predict the most likely cause of potential delays.
Although the airline has global operations, the costs due to delays are the most pronounced due to US domestic flights and especially so with short-haul flights and so that ended up being the focus of the initial predictive models.
Our approach
CrowdANALYTIX first leveraged its community of 25,000+ data scientists to identify external data sources that would be relevant and can be combined with the client’s internal flight operations data.
Once the data sources were identified, ETL algorithms were developed by CrowdANALYTIX to structure and normalize the data from the wide-variety of data sources. These ETL algorithms were deployed for continuous processing of raw data into a format that can be both used to train the predictive models and provide predictions in near real-time once the flight delay prediction application is deployed.
The predictive algorithm that was eventually deployed was picked after evaluating hundreds of approaches and ended up being an ensamble of the top 3 solutions submitted by our community of data scientists.
Impact
The model was a vast improvement on Client’s current methods of predicting flight delays. Not just that, the model was able to provide the top 4-5 factors that were the most indicative of delays at different airports and for different flight durations, allowing the airlines to have customized operations plans for each flight based on the recommendations of the tool.
In early field tests, the airline was clearly able to achieve operational cost reductions of more than 4%. Even if the airline can achieve 70-80% of that reduction when the solution is expanded to other airports, this could result in millions of dollars of bottom line gains.
