Case Study: partnership between DataAi and Digital Catapult UK
Scoring of clients for subprime credit card
Problem: determine the probability that an applicant will default in the next 6 months to 1 year
Challenges: very high default rates, hundreds of variables,data inconsistency, poor statistics
Solution: We boost the accuracy by more than 30% and reduce losses by 18%
Technology: substantial pre-processing and data cleaning and classification with Random Forest
Churn prediction
Problem: based on CDR files (contact detailed records), to predict the probability of a customer to churn
Solution: we matched the accuracy of industry competitors (Kxen) with a less computational intensive approach
Technology: we used complex network analysis with weighted graphs to develop a diffusion model.
Customer segmentation
Problem: based on transactional data from credit cards, predict customer behaviours and categorisation.
Solution: we were able to predict when a customer will become inactive or travel abroad over the next 30 days.
Technology: Self-Organized Maps, Neural Networks and Random Forests algorithms.
Hotel demand prediction
Problem: Predict the demand of hotel bookings, occupancy rate and optimize the RPR.
Solution: we were able to predict the sales up to one month ahead with 89% accuracy.
Technology: Deep learning neural networks for time-series prediction corrected for seasonality effects based on historical data, competition and Google trends.
Image recognition and multimodal learning
Problem: based on facial images, determine emotional states and personality traits.
Solution: we developed an algorithm that achieves human level accuracy in a large database of 30 000 images.
Technology: Convolutional Neural Networks and Stacked Denoising Auto-Encoders