Fast-tracking Model Selection for New Sales Channel

At an energy company moving SME sales to an online channel, pricing models to be deployed were compared using Multi-Armed Bandit. The best model helped to generate a 6% better conversation rate within only three months.

Challenge

Testing multiple models for a new online sales channel.

For a new sales channel, created in the transformation by going from telesales to online, new price prediction models were needed. Since it was a completely new sales channel, no historical data was available at the time of the model development. Right after the go-live, the best performing model needed to be selected to meet the business goals on conversion and margin.

Approach

Using Multi-Armed Bandit (MAB) models to eliminate the less performing models.

In the development phase, we created variations of the models that were used on the telesales channel. In the sales CPQ, a Multi-Armed Bandit (MAB) was integrated with a direct feedback loop to the conversion. The MAB made us capable of selecting the best performing models and abandon the less performing ones.

After three weeks, only the best performing model was left. Now we used the same MAB method the finetune the hyperparameters of the best performing model to even further improve the accuracy of the model.

Methodologies 

  • Multi-Armed Bandits
  • Thompson sampling

Results

  • A good performing model in a new market within one month.
  • A better 6% better conversion rate in the online channel compared to existing telesales channels within three months.
  • After one year, 70% of all the sales were generated via the newly created online channel.
To know more
Contact Us