Axa-Challenge

Course Project for Methods for Big Data Analytics @ Ecole polytechnique

Problem

Axa has 28 call centers, all having different behavior in time. Predicting in advance the number of incoming calls is of high importance for the labor management. When making predictions, from a business perspective, over-estimations are less damagable than under-estimations. Indeed, over-estimation means loss on the labor costs, but under-estimation means bad customer experience, and eventually client churn, which is far more costly than labor cost loss.

Hence, we use a dyssimetric loss measure:

$ LinEx(y, \hat {y}) = exp(\alpha (y - \hat {y}) - \alpha (y - \hat {y}) - 1$

with y the true value and $\hat{y}$ the predicted value. Specifically, an value of 0.1 for the alpha hyperparameter is used in the computation of the final loss value.

Method

Preprocessing

missing values group by for possible duplicates keep only relevant variables

Feature Engineering

See report

Feature selection

PCA, k-PCA, Lasso

Cross-Validation

Specific to Time Series (beware to temporal aspects, usual K-Fold CV does not apply)

Algorithms

ARMA (Time Series) Self-Training Lasso

Results

LinEx = 0.82