Strategy Case Study: Fleet Customer Clone Modeling

The situation

A fast-growing fintech provider wanted to reach prospects in HVAC, construction, field services with 5 to 100 vehicles and offer a comprehensive program to manage fuel, fleet and employee expenses. These company vehicles are on the road often – at job sites, on emergency calls, or making service visits all day.

Their marketing agency proposed to reach the prospects through – gasp – direct mail! But if you think about it, that was brilliant. These prospects are not sitting in their office posting and networking on LinkedIn. They start the day early and are gone all day. The owners of these companies likely catch up on the weekends. It was best to get their attention when they had time to focus on the messages.

There were several hundred thousand prospects that fit their ICP. They applied several selects to narrow down the right prospects to 100K. For the mailing, as is typical practice, they chose multiple sources: three B2B company firmographic databases, an association list, another fleet data source, and ProsperFleet.

While the response was reasonably good in the initial mailing, they faced three difficulties to scaling:

  • No consistency to prioritize multiple list sources
  • Applying several selects caused unpredictable counts
  • Can’t execute quickly to meet print deadlines
Fleet Customer Clone Modeling Case Study

What we provided

The consensus among the customer, their direct mail agency, and ProsperFleet’s team led to proposing a customer clone model. The purpose of a model is to create a single, objective scoring using proven statistical methods. It can be applied to solve the three difficulties the customer was facing.

Our data science team extracted a representative sample from ProsperFleet of their total addressable market and created several derived variables and transformations customary in predictive modeling. We appended their customers that were in the “middle 80%” – we removed the bottom 10% and top 10% so as to not skew the results.

The model was built using a logistic regression. This approach was chosen due to the speed necessary, and the most reliable inference possible, considering the available data.

How our data science capabilities were used

The model found several statistically significant predictors from the ProsperFleet database:

  • Company revenues were sensitive in the lower end
  • Market penetration as defined by state and other geographies
  • Three strong SIC clusters within the service industries
  • Ratios of vehicle mix such light duty as a percent of fleet

As you can see, a model can create innumerable combinations from geography, firmographic, and fleet-specific data to create a much richer set of predictors.

The model found that the top 10% of the ProsperFleet database accounted for 43% of their ideal customers! That is what statisticians call “430% lift.” The model painted a clear picture of who is most like their customers. The team determined that the top 30% would produce at least 35% above average probability of cloning their best customers.

The model then scored the initial mailing and each source of the mailing individually. The test and validation split results showed no model bias.

Regarding results, the association file performed best – no surprise there, but it only made up 2% of the mailing because these types of lists usually have a smaller percent of a market. The three B2B company firmographic databases – let’s call them A, D & Z – performed marginally. ProsperFleet was the second best, but keep in mind this happened to be the largest segment with 29%. The other fleet data source was the worst performer.

The actions

With the right guidance confirmed, the entire ProsperFleet database relevant to the customer’s TAM was scored – about 550K company fleets. The acceptable performance threshold was determined to be decile 3, which is the top 30%. This exercise revealed that from the initial mailing, only 10% of the B2B database records were eligible for remaining, while about 45% of ProsperFleet records exceeded threshold.

For the planned next mailing of 100K, the customer proceeded to re-mail just 15% of prospects that were above this threshold. Another 85% were net new records that were selected from ProsperFleet’s top 30% that cleared the threshold.

The time to select a mailing universe was reduced significantly, from two weeks to just two hours. Both the customer and the direct mail agency were thrilled! Not only did they get a mail file ready on time – which itself is no insignificant feat, they were now able to score any list to those performance standards, and only select the prospects likely to perform the best – and the same standard was applied to ProsperFleet data as well.

There are many takeaways to digest here:

  • Typical list selection methods have limitations of speed, quality and subjectiveness
  • Use your ideal customer characteristics to find more like them with proven techniques
  • Analyze performance methodically, don’t rely only on brand name reputation of data providers
  • Don’t follow the herd with endless email campaigns; know what channel is best to reach your audience and don’t hesitate to execute at least a test

Do you need assistance with fleet market strategy? Contact us if you have a project to discuss.