Work / Datalex

Datalex: turning airline demand signals into smarter pricing decisions

Navi Insights came in to further develop the capability of Pricing AI, Datalex's dynamic pricing engine for airlines. That meant setting up a machine-learning workflow that makes predictions and then checks them against realised outcomes.

Client: Datalex
Sector: Airline retail and e-commerce
Status: In production

The opportunity

Airline pricing decisions have to respond to continually changing demand, booking behaviour, capacity and the time left before departure. The engine was developed to strengthen this decision-making with machine learning, using production airline data to predict key demand signals.

It is meant to augment revenue-management expertise, not replace it. The goal is a scalable capability that can evaluate large numbers of flights and routes consistently, and measure performance leg by leg, not only as a network average.

Navi Insights came in to further develop the engine's capability.

The scope

What the work covered

01Machine-learning predictions of demand signals
02Validation of those predictions at network and individual-leg level
03Automated, repeatable testing

How it was solved

Further developing the capability, in three parts

Part 1 of 3 · Predicting demand signals

Navi Insights established an end-to-end production machine-learning and validation workflow. It draws on historical and production airline data for its predictions, which then feed into pricing, gold-standard validation and performance monitoring.

Part 2 of 3 · Validating predictions at network and individual-leg level

The gold-standard validation compares predictions with realised outcomes at network and individual-leg level.

Performance is measured with WAPE (weighted absolute percentage error), MAPE (mean absolute percentage error), SMAPE (symmetric mean absolute percentage error), R² and bias. R² is the share of the variation in realised outcomes that the predictions explain, where 1.00 means all of it. Bias shows whether the predictions run above or below realised outcomes on average. The strongest production results are set against those in more challenging conditions, such as harder or lower-volume legs. In the strongest results, predicted and actual averages and standard deviations were almost identical across many routes. The broader production evaluation also exposed where calibration is valuable: some configurations systematically over-forecast and were weaker on particular routes.

Performance at a glance, from the production evaluation

Part 3 of 3 · Automated testing and commercial evaluation

Navi Insights also automated the gold-standard testing, so that new production runs can be evaluated consistently. This creates a feedback loop between model output and observed performance, making it possible to identify route-specific weaknesses, changes in accuracy and systematic bias.

The production workflow, from production data to performance monitoring

The programme evaluated commercial outcomes through controlled testing. The analysis looked at revenue uplift alongside conversion-related measures, using frequentist significance testing, multiple-testing correction, statistical power analysis and Bayesian analysis. The commercial question is whether additional revenue is achieved without an unacceptable impact on conversion.

The result

Accuracy, checked leg by leg

Around 1 to 2%WAPE on most evaluated legs in the strongest production results, with R² around 0.99 to 1.00

Typically around 6 to 10%WAPE in more challenging conditions, where harder legs can be higher, with R² commonly around 0.88 to 0.94 and positive bias observed on a number of legs

Automated validation
Network and individual-leg level
In production
AWS

The strongest evaluated models achieved approximately 1 to 2 per cent WAPE, with R² close to 1.00, on many legs. The monitoring also identifies the routes and operating conditions where performance is weaker and further calibration is required. This transparency is a feature of the validation approach, not simply a headline accuracy exercise.

Commercially, better prediction of how demand is developing gives airlines a stronger basis for deciding when and where pricing should change. They can respond more quickly to changing booking and demand patterns and create a measurable foundation for improving pricing and revenue outcomes.

WAPE: Strongest production resultsaround 1 to 2% on most evaluated legs; some around 4% More challenging conditionstypically around 6 to 10%; harder legs can be higher
R²: Strongest production resultsaround 0.99 to 1.00 More challenging conditionscommonly around 0.88 to 0.94

Stack

AWS
S3
Athena
Lambda
KMS
Automated gold-standard testing

Datalex's engine operates within an AWS-based production environment. Components used in the workflow include Amazon S3 for data and model-output storage, Amazon Athena for querying production datasets, AWS Lambda for automated processing and scheduled execution, and AWS KMS controls around secured data access.

Next case