Recency

Let us begin with recency. Earlier, we defined it as the number of days since the last transaction of the customer. How do we calculate this metric? Apart from the date of the last transaction of the customer, what other information do we need? In all the discussions till now, we have missed out on a key point i.e. the time frame of the analysis.

The most crucial step in RFM analysis is to select a time frame from which we use the transaction data. How do we decide on this time frame? It depends on the domain to which we are applying this analysis. Customers visit a grocery store more often than they visit a consumer durables store. Similarly, people consume content from news & blogs more frequently while they may visit an e-commerce website only when they have to purchase something. Keeping in mind the domain to which the analysis is being applied, select an appropriate time frame. To calculate recency, compute the difference between the last transaction date and the analysis date i.e. the last date of the selected time frame.

In the above example, the analysis date is 2016-12-31. To compute recency, we first extract all the transaction date of customer Taj and then select the last transaction date, 2015-04-21, and subtract if from the analysis date to get the number of days since the last transaction date, 620.

Frequency

Frequency is the count of transactions. In the online/digital world, we need to decide whether we will consider all the visits to a website or app as the frequency or only those which resulted in a transaction/conversion. In the below example, we count the transactions for each customer and use it as frequency. Lionel has 6 transactions, Jaineel has 9 transactions and Taj has 4 transactions.

Monetary Value

Monetary value is the total revenue from each customer in the selected time frame. It is computed by summing up the transaction amount. In our case study, Jaineel has spent the highest amount of $843, followed by Lionel who has spent $472 and Taj has spent the lowest, $196. As you can observe, we have arrived at these figures by summing the values in the third column of the second table, Order Amount.

If we are applying this analysis to the digital world, we may think of using a metric such as time spent on the website/app instead of transaction/order amount.

Monetary Score

Let us generate the monetary score in our case study. The first step is to compute the quantiles using the quantile() function. We use the revenue column from the RFM table to compute the quantiles. If you look at the example, it gives us the cut off below which a certain percentage of customers are present.

• the bottom 20% of customers spend below $254.8.
• the next 20% of customers spend between $254.8 and $381.0.
• the top 20% of customers spend above $665.0.

Using these cut off points we have created intervals which can then be converted to if else statements. The intervals are then used to assign scores. For example, Lionel falls in the interval > 381.0 & <= 505.4 and hence is assigned the score 3. Similarly, Jaineel and Taj are assigned the scores 5 and 1. How do we interpret the scores? The score is more like a rank. A customer with a score of 3 is ranked higher than a customer with score of 1 as his transaction amount is higher. In the rfm package, we use the above method to assign the scores.

Some users reverse the order of the score i.e. top 20% customers by transaction amountare assigned the score 1 and the bottom 20% are assigned the score 5.

Frequency Score

The frequency score is computed in the same way as the monetary score. Instead of using the revenue column from the RFM table, we use the frequency column. Using quantiles, we arrive at the cut off points below which a certain percentage of customer are present. If you observe the example, the first table shows the quantiles and the associated cut off points.

• the bottom 20% of customers visit/transact less than 3 times.
• the next 20% of customers visit/transact around 4 times.
• the top 20% of customers visit/transact more than 7 times.

The cut off points are then used to create the intervals and assign the scores as shown in the second table. We assign a higher score to those who visit more frequently and a lower score to those who visit less frequently.

In our case study, Jaineel has visited 9 times and hence assigned the score 5 where as Taj has visited only 4 times and hence the score 2.

Recency Score

The recency score follows the same methodology but uses a slightly different concept while assigning the score. If you look at the metrics, the higher the values of frequency and monetary, the better as we want customers to transact frequently and spend higher amount but it is not the case with recency. Since recency represents the number of days since the last transaction, the lower its value the better i.e. customers who visited in the recent past are more likely to visit again whereas customers who visited long back may be as good as lost. Hence in the case of recency, higher score is assigned to those with lower recency value and vice versa.

In the above example, we have used quantiles to compute the cut off point for recency. The first table shows the quantiles and the associated cut off points:

• the bottom 20% of customers visited more than 481 days back.
• the next 20% of customers visited between 296.4 and 481 days back.
• the top 20% of customers visited less than 114 days back.

The above statements will become clear if you study the second table which includes the interval and the score. We have assigned a higher score to those who visited in the recent past (< 114 days) compared to those who visited way back (> 481 days). In our case study, Jaineel visited in the past 3 months and hence the score of 5 where as Taj visited almost 20 months back and has been assigned the score 1.

RFM Score

Now that we have calculated the individual scores, let us compute the RFM score using the below formula:

RFM Score = Recency Score * 100 + Frequency Score * 10 + Monetary Score

The below table shows the individal scores of recency, frequency and monetary as well as the RFM score. All of them are computed from the RFM table which in itself is based on the transaction data.

Data

To calculate the RFM score for each customer we need transaction data which should include the following:

• a unique customer id
• date of transaction/order
• transaction/order amount

rfm includes a sample data set rfm_data_orders which includes the above details:

rfm_data_orders
## # A tibble: 4,906 x 3
##    customer_id         order_date revenue
##    <chr>               <date>       <dbl>
##  1 Mr. Brion Stark Sr. 2004-12-20      32
##  2 Ethyl Botsford      2005-05-02      36
##  3 Hosteen Jacobi      2004-03-06     116
##  4 Mr. Edw Frami       2006-03-15      99
##  5 Josef Lemke         2006-08-14      76
##  6 Julisa Halvorson    2005-05-28      56
##  7 Judyth Lueilwitz    2005-03-09     108
##  8 Mr. Mekhi Goyette   2005-09-23     183
##  9 Hansford Moen PhD   2005-09-07      30
## 10 Fount Flatley       2006-04-12      13
## # ... with 4,896 more rows

RFM Score

Use rfm_table_order() to generate the score for each customer from the sample data set rfm_data_orders.

rfm_table_order() takes 8 inputs:

• data: a data set with
  • unique customer id
  • date of transaction
  • and amount
• customer_id: name of the customer id column
• order_date: name of the transaction date column
• revenue: name of the transaction amount column
• analysis_date: date of analysis
• recency_bins: number of rankings for recency score (default is 5)
• frequency_bins: number of rankings for frequency score (default is 5)
• monetary_bins: number of rankings for monetary score (default is 5)

RFM Table

analysis_date <- lubridate::as_date("2006-12-31", tz = "UTC")
rfm_result <- rfm_table_order(rfm_data_orders, customer_id, order_date, revenue, analysis_date)
rfm_result

rfm_table_order() will return the following columns as seen in the above table:

• customer_id: unique customer id
• date_most_recent: date of most recent visit
• recency_days: days since the most recent visit
• transaction_count: number of transactions of the customer
• amount: total revenue generated by the customer
• recency_score: recency score of the customer
• frequency_score: frequency score of the customer
• monetary_score: monetary score of the customer
• rfm_score: RFM score of the customer

Segments

Let us classify our customers based on the individual recency, frequency and monetary scores.

Segmented Customer Data

We can use the segmented data to identify

• champion customers
• loyal customers
• at risk customers
• and lost customers

Once we have classified a customer into a particular segment, we can take appropriate action to increase his/her lifetime value.

Segment Size

Now that we have defined and segmented our customers, let us examine the distribution of customers across the segments. If our segmentation logic is good, few or no customer should be categorized as Others.

segments %>%
  count(segment) %>%
  arrange(desc(n)) %>%
  rename(Segment = segment, Count = n)
## # A tibble: 12 x 2
##    Segment            Count
##    <chr>              <int>
##  1 At Risk              157
##  2 Potential Loyalist   132
##  3 Others               128
##  4 Champions            116
##  5 Need Attention       100
##  6 Hibernating           97
##  7 About To Sleep        92
##  8 Lost                  75
##  9 Loyal Customers       43
## 10 Promising             21
## 11 Can't Lose Them       17
## 12 New Customers         17

We can also examine the median recency, frequency and monetary value across segments to ensure that the logic used for customer classification is sound and practical.

rfm_plot_median_recency(segments)

rfm_plot_median_frequency(segments)

rfm_plot_median_monetary(segments)

Heat Map

The heat map shows the average monetary value for different categories of recency and frequency scores. Higher scores of frequency and recency are characterized by higher average monetary value as indicated by the darker areas in the heatmap.

rfm_heatmap(rfm_result)

Bar Chart

Use rfm_bar_chart() to generate the distribution of monetary scores for the different combinations of frequency and recency scores.

rfm_bar_chart(rfm_result)

Histogram

Use rfm_histograms() to examine the relative distribution of

• monetary value (total revenue generated by each customer)
• recency days (days since the most recent visit for each customer)
• frequency (transaction count for each customer)

rfm_histograms(rfm_result)

Customers by Orders

Visualize the distribution of customers across orders.

rfm_order_dist(rfm_result)

Scatter Plots

The best customers are those who:

• bought most recently
• most often
• and spend the most

Now let us examine the relationship between the above.

rfm_rm_plot(rfm_result)

rfm_fm_plot(rfm_result)

rfm_rf_plot(rfm_result)

Your Turn…

• if you look at the distribution of segments, around 13% of the customers are in the Others segment For segmentation to be effective and optimal, the Others segment should be eliminated or should have few customers only. Redefine the segments and try to reduce the number of customers in the Others segment.

• we have defined 11 segments. Try to combine some of the existing segments and bring down the total segments to around 6 or 8.

• the RFM score we generated uses score between 1 and 5. Try to create segments by using a score between 1 and 3 i.e. the lowest RFM should be 111 and the highest should be 333 instead of 555.

• reverse the scores i.e. so far we have assigned a score of 5 to customers who visited recently, frequently and had higher transaction amount and 1 to customers who visited way back, rarely and have low transaction amount. Reverse this score pattern and create the segments.