How walking through the “Random Forest” helps you select the right business opportunities, optimize your costs and reach higher performance.

At Karetis, we are “entropy reducers”. We take messy, complicated situations and work to simplify as well as clarify them – to help our clients make decisions. To achieve that, analytics, business acumen and experience will often do the trick. But some situations are too complex for human intelligence alone.

We very recently helped a large pharmaceutical company in China. They had a simple question: how to identify opportunities to increase sales force productivity? The management’s base hypothesis was that too many reps were assigned to big, well-established accounts while newer and growth-driving accounts were not given enough attention. But our client needed to validate this intuition and quantify its magnitude. From an Operations perspective, they also needed to identify which accounts required more or less effort to boost sales growth and increase reps productivity.



Figure 1: Jean-François, Analyst at Karetis Paris

The problem is by nature quite complex – considering China’s scale, it is exceedingly complex. Several thousand sales reps are promoting a dozen of major brands to tens of thousands of accounts across hundreds of cities.

For starters, we know that sales productivity is influenced by many factors such as the listing and reimbursement situation in each city or province, the type of account, the type of geography, the profile of sales reps doing the promotion, the amount and quantity of staff meetings, congress invitations in the account or in the city, etc. But how do we move forward? How do we develop analyses to convince sales management to act on our recommendations?

In this context, Karetis brought in the power of machines. Sifting through gigabytes of data to find patterns is a hopeless task for a human brain. Artificial intelligence powered by a computer brain however will make excellent work of these huge amounts of data: the more information at hand, the smarter it becomes.



Figure 2: Lee, Analyst at Karetis Singapore

We had assembled our small army of robots: if you think we could just let them loose and sit back until they solved the problem, think again.

Before you target the issue at hand with robots, it is crucial that you equip them with the right type of Artificial Intelligence (AI). The brain of an AI runs on a machine learning algorithm, which many models can be found in free libraries on the Internet.

Like a Mad Max car, these models can then be tweaked and tuned to achieve any outcome. A linear regression model is like an old Japanese katana: traditional, simple, ruthlessly efficient but not very subtle. A deep neural network is like thermonuclear warfare: high-tech, hard to develop and handle, deadly over a large scale.

Our choice was a Random Forest model. Random Forest is an ensemble of hundreds of decision trees, each slicing and dicing the data in a different way. Individual trees are crude devices, flailing their leaves at the problem with poor precision. Assembled as a forest however, the decision trees will reach a consensus as wise and nuanced as the magical forest of Broceliande.

Growing an effective random forest takes far more work than watering cactuses. We fertilized the model with carefully selected amounts of rich data. We trimmed our decision trees with precision and periodically removed weeds. Once the sylvan artificial intelligence was fully grown however, its insights proved immensely rewarding.

The AI could rank all the factors impacting field force productivity by order of importance. It assessed the most appropriate balance between face to face calls and staff meetings. More impressive, it determined for every single account in China whether the level of promotion was appropriate, too high, too low, and how to optimize it.

Incredulous, we pitted the machine against an experienced sales manager on his own turf. For every account we tested, the human and the machine reached the same conclusion – only the AI was faster and better quantified the results.



Figure 3: Ranking of Variables’ Importance

Equipped with these results, we knew the brands and geographies on which we were over-investing, and where it would be most effective to redeploy resources. We proposed and evaluated two scenarios:

1. Keep overall resources constant but reallocate investment between accounts

      • Estimated 5-10% increase in sales and productivity

2. Reduce overall level of investment and reallocate resources more efficiently to maintain the sales trend

      • Estimated 15-20% savings in costs

Backed by the power of machines, we could confidently start the process of finalizing the scenarios and implementing the changes with local sales teams. Back to human intelligence…



Figure 4: Increase of market share as a function of total promotional effort and % of spending on staff meetings

Follow us on LinkedIn and Twitter to be sure not to miss our latest articles.

To know more about Karetis’ machine learning and promotion response expertise, feel free to contact us and ask for the Karetis Promotion Response White Paper. If you’re ready to start improving your sales & marketing productivity, we will be happy to meet you and discuss your specific situation and needs.