RFM is a classic Lifetime and Responsiveness segmentation model. It has been trialed and tested over the years and is a great starting point for any retailer including eCommerce companies looking to manage their customer base more proactively.
Recency (R) - Time since last purchase in days
Frequency (F) - Total number of purchases
Monetary value (M) - Total monetary value
- Python 3.7 - More information
pip install -r requirements.txt
or pip install
plotly
pandas
numpy
scipy
statsmodels
matplotlib
streamlit
- Find the Data here - Download the publik dataset from here
You can find the code in the model model.py file
After running the model and writing the csv file. You can also use a basic streamlit app included in the repository names app.py. The app will allow you to do analyse and review RFM scores in the output data. This is useful for model optimisation. You can run the app with the following command:
streamlit run app.py
Access the app via the localhost link.
##########################################################################################################################################
### RFM MODEL ###
##########################################################################################################################################
import plotly.express as px
import statsmodels.api as sm
import pandas as pd
import numpy as np
import warnings
warnings.filterwarnings('ignore')
data = pd.read_csv('transaction_data.csv', encoding='ISO-8859-1')
##########################################################################################################################################
### Create the Customer Table ###
##########################################################################################################################################
# Calculate Sales Value
data['sales_value'] = data['Quantity'] * data['UnitPrice']
# group columns by customer_id
rfmTable = data.groupby(
['CustomerID'], as_index=False
).agg(
{
'sales_value' :sum
, 'InvoiceNo': pd.Series.nunique
, 'InvoiceDate': max
}
)
# Calculate recency
rfmTable['InvoiceDate'] = pd.to_datetime(rfmTable['InvoiceDate'])
rfmTable['InvoiceDate'] = rfmTable.InvoiceDate.dt.date
today = rfmTable.InvoiceDate.max() #use the latest date in the dataset - in the real world this will be todays system date
rfmTable['Recency'] = (today - rfmTable['InvoiceDate']).dt.days #Days since last order
##########################################################################################################################################
### Stats Tests ###
##########################################################################################################################################
# first rename the columns to a more user friendly format
rfmTable = rfmTable.rename(columns={
'sales_value':'MonetaryValue', 'InvoiceNo':'Frequency', 'InvoiceDate':'LastOrderDate'
}
)
#show distribution of values
#recency
fig = px.histogram(rfmTable, x="Recency", y="CustomerID", marginal="box", # or violin, rug
hover_data=rfmTable.columns, title='Recency Plot')
fig.show()
#frequency
fig = px.histogram(rfmTable, x="Frequency", y="CustomerID", marginal="box", # or violin, rug
hover_data=rfmTable.columns, title='Frequency Plot')
fig.show()
#monetary value
fig = px.histogram(rfmTable, x="MonetaryValue", y="CustomerID", marginal="box", # or violin, rug
hover_data=rfmTable.columns, title='Monetary Value Plot')
fig.show()
#Q-Q plot of the quantiles of x versus the quantiles/ppf of a distribution.
# set up the plot figure
from statsmodels.graphics.gofplots import qqplot
from matplotlib import pyplot as plt
f, axes = plt.subplots(2, 2, figsize=(20,12))
#define distribution graphs
qqplot(rfmTable.Recency, line='r', ax=axes[0,0], label='Recency')
qqplot(rfmTable.Frequency, line='r', ax=axes[0,1], label='Frequency')
qqplot(rfmTable.MonetaryValue, line='r', ax=axes[1,0], label='MonetaryValue')
#plot all
plt.tight_layout()
##########################################################################################################################################
### RFM Score Function ###
##########################################################################################################################################
# Detemine the dataset quantiles
q = np.arange(0, 1, 0.10).tolist()
quantiles = rfmTable.quantile(q=np.around(q,decimals=2))
# Send the quantiles to the dictionary
quantiles = quantiles.to_dict()
# Start creating the RFM segmentation table
rfmSegmentation = rfmTable[['CustomerID','MonetaryValue','Frequency','Recency']]
# We created to classes where high recency is bad and high frequency/ money is good
# 1. Arguments (x = value, work on intervals of 90 days)
def RClass(x):
if x <= 60:
return 1
elif x <= 120:
return 2
elif x <= 180:
return 3
elif x <= 360:
return 4
elif x <= 540:
return 5
else:
return 6
# 2. Arguments (x = value, p = frequency)
def FClass(x,p,d):
if x <= d[p][0.3]:
return 6
elif x <= d[p][0.4]:
return 5
elif x <= d[p][0.6]:
return 4
elif x <= d[p][0.8]:
return 3
elif x <= d[p][0.9]:
return 2
else:
return 1
# 3. Arguments (x = value, p = monetary_value)
def MClass(x,p,d):
if x <= d[p][0.2]:
return 6
elif x <= d[p][0.4]:
return 5
elif x <= d[p][0.6]:
return 4
elif x <= d[p][0.8]:
return 3
elif x <= d[p][0.9]:
return 2
else:
return 1
# 4. Customer Segment Arguments (x = value, slice by value distribution in order to segment stage)
def CustomerSegment(x):
if x['R_Quartile'] ==1 and x['F_Quartile'] ==1 and x['M_Quartile'] ==1:
return "Champions"
elif x['R_Quartile'] <=2 and x['F_Quartile'] <=2 and x['M_Quartile'] <=2:
return "Loyal_Customers"
elif x['R_Quartile'] <=2 and x['F_Quartile'] <=3 and x['M_Quartile'] <=3:
return "Potential_Loyalists"
elif x['R_Quartile'] <=2 and x['F_Quartile'] <=4 and x['M_Quartile'] <=4:
return "Promising"
elif x['R_Quartile'] <=2 and x['F_Quartile'] <=6 and x['M_Quartile'] <=6:
return "Recent_Customers"
elif x['R_Quartile'] ==3 and x['F_Quartile'] <=3 and x['M_Quartile'] <=3:
return "Customer_Needs_Attention"
elif x['R_Quartile'] ==3 or x['R_Quartile'] ==4 and x['F_Quartile'] >=5 and x['M_Quartile'] >=5:
return "Hibernating"
elif x['R_Quartile'] ==4 and x['F_Quartile'] <=3 and x['M_Quartile'] <=3:
return "At_Risk"
elif x['R_Quartile'] ==4 and x['F_Quartile'] >=3 and x['M_Quartile'] >=3:
return "About_to_Sleep"
elif x['R_Quartile'] >=5 and x['F_Quartile'] >=3 and x['M_Quartile'] >=3:
return "Lost"
elif x['R_Quartile'] ==5 and x['F_Quartile'] <=3 and x['M_Quartile'] <=3:
return "Cant_Lose_Them"
elif x['R_Quartile'] ==6 and x['F_Quartile'] <=3 and x['M_Quartile'] <=3:
return "High_Value_Sleeping"
else:
return "Lost"
##########################################################################################################################################
### CALCULATE THE RFM SCORES ###
##########################################################################################################################################
# Scores
rfmSegmentation['R_Quartile'] = rfmSegmentation['Recency'].apply(RClass)
rfmSegmentation['F_Quartile'] = rfmSegmentation['Frequency'].apply(FClass, args=('Frequency',quantiles,))
rfmSegmentation['M_Quartile'] = rfmSegmentation['MonetaryValue'].apply(MClass, args=('MonetaryValue',quantiles,))
# Classify the RFM score for the customer base
rfmSegmentation['RFMClass'] = rfmSegmentation.R_Quartile.map(str) \
+ rfmSegmentation.F_Quartile.map(str) \
+ rfmSegmentation.M_Quartile.map(str)
# Classify customer segments based on RFM scores
rfmSegmentation['Customer Segment'] = rfmSegmentation.apply(lambda x: CustomerSegment(x), axis=1)
#scatter plot to display segments
rfm_scatter = rfmSegmentation[(rfmSegmentation['MonetaryValue'] > 0) & (rfmSegmentation['Recency'] <=360) & (rfmSegmentation['Frequency'] <= 50)]
fig = px.scatter(rfm_scatter, x="Recency", y="Frequency", color="Customer Segment",
size='MonetaryValue', hover_data=['R_Quartile', 'F_Quartile', 'M_Quartile'])
fig.show()
# Save the results to a csv file
output_table = rfmSegmentation.to_csv('rfm_segments.csv')
- Include the computations as a part of your ETL process (I use KNIME) - as the step before writing the final customer lifetime table to your Data Warehouse.
- Have the model run on a monthly cycle - then you have snapshots of how customers progress through the different segments over time. It is a great way to track if you are feeding enough customers into active segments to feed your MVP (most profitable customers).
- Trigger events based on scores or segments with your CRM or campaign management system (On-boarding, Retention and Reactivation campaigns).
- Francois van Heerden - Experience - LinkedIn Profile
- Found inspiration from multiple fellow Data Scientists in the open source community.
- But I would like to specifically highlight this post How to segment customers with RFM analysis