{"id":5435,"date":"2021-01-20T13:42:15","date_gmt":"2021-01-20T08:12:15","guid":{"rendered":"https:\/\/razorpay.com\/blog\/?p=5435"},"modified":"2021-02-23T11:10:55","modified_gmt":"2021-02-23T05:40:55","slug":"prioritzing-leads-for-sales-ml","status":"publish","type":"post","link":"https:\/\/razorpay.com\/blog\/prioritzing-leads-for-sales-ml\/","title":{"rendered":"How We Use ML to Prioritise Leads for Our Sales Team"},"content":{"rendered":"

\u201cHow can we improve the conversion rates of incoming leads?\u201d<\/span><\/i><\/p>\n

\u201cHow do we improve sales team’s efficiency?\u201d<\/span><\/i><\/p>\n

\u201cWhich is the best medium to target leads?\u201d<\/span><\/i><\/p>\n

These are some questions that sales managers are commonly asked. The best answers to such questions lies with data. This is why scoring incoming leads based on various data points can help in effective targeting.\u00a0<\/span><\/p>\n

At Razorpay, we have tried two common approaches to arrive at a lead score.<\/span><\/p>\n

1) Heuristics based scoring<\/b><\/h3>\n

This approach involves providing weightage to various important factors based on intuition and on the lead score. The leads in the high scoring bucket are reached out using calls, the ones with medium score are reached out using SMS, and the remaining in low scoring buckets could be reached out via emails. The thresholds of these buckets could be decided based upon the team’s capacity for handling calls.<\/span><\/p>\n

For instance, let\u2019s suppose an inbound sales team of 5 members can call a maximum of 100 leads a day out of 500. Then the top 20% leads could be classified as high quality buckets and the remaining could be placed under medium and low-quality buckets. This method is used when an organisation is in its early stages and there is insufficient historical data to learn from.<\/span><\/p>\n

2) Machine Learning based leads scoring<\/b><\/h3>\n

This approach requires leveraging historical data to identify the most significant features that result in lead conversion, computing features weightage, and building a predictive model to score new leads.<\/span><\/p>\n

This type of method is applied when there is sufficient sample size available. As a thumb rule, the minimum sample size for the binary<\/span> logistic regression<\/span><\/a> model could be determined using the basic formula, n = 100 + 50i, where i is the number of features used in the statistical model. For instance, if we have 5 features, then the minimum sample size required to create the model=350.<\/span><\/p>\n

The first approach is pretty simple and straightforward, so a detailed explanation of the same is skipped. Here, the implementation of the second approach to create a leads scorecard is discussed in detail. The scorecard is very easy to interpret and could be easily understood by people with no technical background. A sample leads scorecard with two features is shown below.<\/span><\/p>\n

\"\"<\/a>Figure 1: Sample scorecard with two features<\/span><\/p>\n

The <\/span>problem statement<\/span><\/h2>\n

There are a lot of customers who sign up on the<\/a><\/span> Razorpay<\/span> website expressing their interest to become partners. However, the volumes of signups and limited team size makes it difficult for the sales team to get in touch with all the leads. This is where the lead scorecard comes to the rescue. There are two main parts in building the scorecard:<\/span><\/p>\n

    \n
  1. Developing a statistical model<\/span><\/li>\n
  2. Applying the statistical model to compute the lead score for a partner merchants <\/span>
    \n<\/span><\/li>\n<\/ol>\n

    Statistical model<\/span><\/h2>\n

    Building a statistical model involves data collection, exploration, transformation, feature selection, model training, and model testing.<\/span><\/p>\n

    Step 1: Data Collection<\/span><\/h3>\n

    We gathered historical data of 10,000+ partners to study the influence of various factors on their conversion. Shown below are the first set of all the features, called independent variables, and the outcome variables.<\/span><\/p>\n

    \"\"<\/a>Figure 2: Independent features for the data model<\/p>\n

    \"\"<\/a>Figure 3: Target feature for the data model<\/span><\/p>\n

    In the blog, the partner merchants who have referred and on-boarded at least one sub-merchant are referred to as “Good partners” and the rest are referred as “Bad partners”.<\/span><\/p>\n

    The dataset is divided into training and test – training for learning from the historical data, and testing for evaluating model accuracy metrics.<\/span><\/p>\n

    Step 2: Data Exploration<\/span><\/h3>\n

    Data exploration involves a univariate and bivariate analysis of various independent features to understand basic trends in data.<\/span><\/p>\n

    The univariate analysis involves only one variable for summarising the data. Histograms, box plots, etc are some of the most common graphs for performing univariate analysis of continuous features.<\/span><\/p>\n

    The bivariate analysis involves more than one variable for determining empirical relationships between them. It could be descriptive as well as inferential. The bivariate analysis could be done between two continuous or categorical features or one continuous and one categorical feature. Stacked bar charts, correlation plots, etc are some of the generally used graphs for bivariate analysis.<\/span><\/p>\n

    Step 3: Data Transformation<\/span><\/h3>\n

    This step involves transforming all the features using the weights of evidence(WOE) method. All the categorical variables can be directly transformed, whereas continuous features (like Delta in months, Months since signup) are bucketed before WOE imputation. The buckets could be created based on the percentile approach or business intuition.<\/span><\/p>\n

    WOE provides the weights for each of the categories of features, based on the distribution of good partners and bad partners.<\/span><\/p>\n

    Information value(IV) is a useful method for measuring the predictive strength of features and for selecting significant features for the model.<\/span>\"\"<\/a>Figure 4: Mathematical formula for WOE and IV<\/span><\/p>\n

    The WOE and IV calculations for each category of one of the continuous features (Month since signup) is shown below.<\/span><\/p>\n

    \"\"<\/a>Figure 5: WOE and IV computation for a continuous feature (Features analysis reports)<\/span><\/p>\n

    Where:<\/span><\/p>\n