Maximizing Lead Conversion and Marketing Effectiveness through Advanced Analytics — An Automobile Industry Context

A Major challenge for organizations today is reacting and reaching out to right customers at right time using various touch points to improve customer experience and build customer loyalty. Ability to identify right target segment and right solution offerings for promotion campaign is a common business objective in every industry, like Manufacturing, Telecom, BFSI, Retail etc.

Vehicle manufacturers and service providers are struggling to identify the right mix of customers for each vehicle family/type of vehicle(SUV, Sedan,MUV etc.) type and what influences customer to buy vehicle(Driver of Sales) the best, to optimize product/service mix and campaign design. Advanced analytics on the helm of Big Data and cutting edge machine learning techniques enables organizations to understand the customer than ever before to find answers for following business problems:

1) Who are the target customer with high propensity to buy for better conversion to make a promotion/campaign effective

2) What are drivers of effective and impactful campaign?

3) What is the correct product mix a particular customer segment?

With the evolution of new ‘Big Data’ era the organizations are flooded with useful data, in the form of in the form of structured, unstructured and semi structured, collected in various touch points to have 360 degree view of customer. Using this customer touch point data customized advanced analytics solution can be built to bring value to entire ecosystem of marketing and CRM function.

Fig 1: Breadth of Analytics Solutions to Marketing and CRM value chain

As the maturity and complexity of the analytics has increased the organizations are capable of devising right customer acquisition, retention and relationship strategy by precisely identifying right target segment for promotion campaign, right channel mix to reach right product/service at the right time with right price points. The increased efficiency in marketing function backed by data driven insights and predictive models has helped organization to do more- increased revenue($)/sales(volume) with less resource- reduced cost, manpower etc.

Case Study Brief

Client: A major automobile giant having predominant presence in North America and Europe.

Business Need: Based on following available data elements, the marketing and campaign relationship team like to identify and target right customer(most likely)with right product offerings to boost sales and improve campaign effectiveness(reduce average customer acquisition cost).

1) Old lead data: Consists of Customer details, Demographic and Personal Interests detail of the customers who enquired about the vehicles.

2) Old Sales Data :- Consists of Customers who purchased a vehicle((Leads Converted from Old Leads Data),Purchased Vehicle Details and list of features

3) New Leads Data: Consists of new customers (Fresh leads who yet to make a decision) who enquired about the vehicle and their demographic and personal interests.

Derived Analytics Problem and Solution Framework:

1) Propensity to buy model: By calculating Purchase Probability score for new leads to optimize lead conversion by devising customized campaign strategy for different score bucket. Typically binary classification technique(logistic regression, SVM, Decision Tree etc.)is widely accepted method here.

Model Equation:

Y=(Intercept) + (* IV1) + ( * IV2) + ( * IV3) +……. + ( * IVn)

Purchase Probability = exp(Y) / (1 + exp (Y))

2) Product Recommendation System: It’s a classic multinomial classification problem,building a multi class model(Multinomial Logistic Regression, CHAID, Random Forrest etc.)to build probability score for each class and selecting the class having maximum probability value.Also other recommendation methods(Collaborative Filtering, Content-based filtering etc) can be used.

Fig 2: Multinomial Classification Model and Product Recommendation System

Solution Approach:

Propensity to Buy Predictive Model:-. It’s a classic binary classification model using old lead data features and mapping buy flag(if there is a buy then 1 or 0) against old customer from old sales data.We had total buy for 4300 customer out of total 20000 leads(Event rate-21.5%). Following classification techniques are used and the final model is an average ensemble model of all models to increase accuracy and reduce variance(overfit).Then the model applied to new lead data to calculate likelihood to buy score for each new customer and class prediction(Buy or No Buy).

1) Logistic Regression

2) Tree based Models

3) SVM

4) Random Forrest

5) Adaptive Boosting

(I will discuss individual model building techniques in a separate post.)

Fig 3: Propensity to Buy Model Framework

Model Performance: We have four performance criteria of model goodness of fit, A) AUC ROC (Area under receiver operating curve), B) Cumulative Lift/Gain(Cumulative accuracy profile), C) Overall Accuracy(Total Correct Classification/Total No of cases),D) Sensitivity(True Positive/Actual Positive). Although not exhaustive following are some model performance result on out-of-sample validation data (5000 customers).

Fig 4:Model Performance Comparison — Metrics,Cumulative Lift Chart and AUC-ROC

From the lift and ROC it’s clearly visible almost all the models have done their job pretty well in differentiating (rank ordering)likely buyer from non-buyer. Almost all the models have cumulative lift score of more than 2.5 in top decile(10% of total population), which means given scoring model is 2.5 times better than the random model at this score range.

Product Recommendation System:From old sales data sample we found total 430 vehicle buy for following four classes. With this historical data using other customer demographic and interest features we have built multinomial classification models to predict probable class for new leads

1) Sports

2) Next Gen

3) 200S

4) Sedan

Following four types of models have been built and the final probability score is average ensemble of these to increase accuracy in prediction and reduce variance.

1) Multinomial Logistic Regression

2) CHAID

3) Random Forest

4) Gradient Boosting(GBM)

Fig 5: Multinomial classification Model Building and Performance Metrics

Actionable Insights and Business Value Creation:From propensity model output our client gets important some of actionable insights to take corrective action fine tune their campaign strategy and focus more towards the customers, having maximum likelihood to buy vehicle. Also identifying the leads having less propensity to buy in advance there is an opportunity to launch customized promotion(reward points, price markdown etc.) for this segment to maximize the conversion rate.

Fig 6: Propensity Model Insight

From above ranking plot business gets a clear insight about buying propensity for each of its customer segment. From cumulative gain chart marketing team can form strategy regarding whom to target for campaign to maximize campaign ROI. From the cumulative gain chart we can see by targeting only top 20% of total leads sales team can achieve more than 80% of total sales volume to achieve a pretty high ROI of 400%. If would like to acquire 95% of total buyers, it has to focus on top 40% of leads, based on propensity score, to achieve a 240% ROI.

Now, it is natural to ask what is the financial impact of all these models and insights to be able to decide what the worthiness this solution is. Following is an illustrative calculation of financial gain using some simplifying assumptions., although in real business the approach may be slightly more complex.

If our partner has total 50000 leads(enquiry)/month and average cost of communication(phone, mail, direct visit etc.) is $2/, then without having propensity model the customer relationship team have to target all leads with having probability of conversion 0.22. So here the cost of marketing is 100000$ with approx. 11000(0.22 × 50000) sales. Now with this propensity model with targeting only top 20%(10000),top 40%(20000) and 45%(22500)prospective customers leads to 8800(80% of true buy),10450(95% of true buy) and 10945(99.5% of true buy) vehicle sales. While the last two numbers are very close to the total true sales figure in first case there is a substantial saving of USD 60000/month(30000 × 2) and USD 55000/month(27500 × 2) in marketing budget. This reduction in marketing budget can be utilized for launching customized campaign to increase conversion rate. Now the marketing team has to optimize the cutoff propensity score to target more/less customer to make a trade of between reduction in marketing budget and increase in sales figure.

Also these models would help the marketing SBU to understand the relationship (linear/non-linear) between different independent variables(customer features and activities) and vehicle sales rate, which in turn brings the insight about driver of sales or the important features, which differentiate probable buyers from non-buyers. Understanding these drivers and the exact relationship the customer team can come with different campaign strategies for different customer segments and identify combination of features (decision rule) for buyers.

Fig 7: Propensity Model Insight- Driver of Sales and Buying Decision Rule

The multinomial classification algorithm has helped the promotion team to build automated recommendation system with right vehicle recommendation suited for each customer.

Fig 8: Automated Product Recommendation System

Additionally it provides insight about the key buyer features(behavioral attributes) for each product class, so that the promotion team can understand respective target audiences better.

Fig 9: Product wise Customer attributes Matrix

Road Ahead:The case study demonstrated the use of advanced analytics framework to solve a very common business problem for marketer using internally available data(Customer database, Presales and Sales data etc.) and market research data(Questionnaire and Survey Data).Today with the evolution of big data technology and explosion of terabytes of data, there is a potential to use other data sources, especially to leverage the social media space, like Facebook, Twitter etc. to have to get real time information about the customers.

There is a myriad of opportunities to integrating these unstructured data elements and performing latest techniques, like advanced text mining, NLP etc. to discover individual customer’s opinion about products/services to take more informed decision. This will make the campaign strategy more effective and precise.