FINAL ASSIGNMENT ON MIS
1. Discuss whether the churn problem can be modelled as a Markov chain. What are the
assumptions made while modelling customer churn as a Markov Chain?
Yes, churn can be modelled as a Markov chain under specific conditions. A Markov chain is a
stochastic process where the future state depends only on the present state and not on the past states. In
the context of churn, this implies that a customer's decision to stay or churn in the next period depends
solely on their current state (e.g., active, inactive, churned) and not on their previous behavior.
However, this is a simplification. Customer behavior is often influenced by past actions, such as
purchase history, service interactions, and marketing campaigns. Therefore, while Markov chains can
provide insights, they may not capture the full complexity of churn behavior.
ASSUMPTIONS OF MODELLING CHURN AS A MARKOV CHAIN:
Markov Property: The future state of a customer depends solely on the current state and not on
the past. This assumption is often violated in practice, as customer behavior can be influenced by
past experiences.
Stationarity: The transition probabilities between states remain constant over time. This
assumption may not hold true in dynamic markets with changing customer preferences and
competitive landscapes.
Discrete State Space: The number of possible customer states is finite and discrete. This
assumption can be restrictive, as customer behavior can exhibit a wide range of nuances.
Time Homogeneity: The transition probabilities are independent of time. This assumption might
not be accurate for businesses with seasonal fluctuations or product lifecycles.
LIMITATIONS OF MARKOV CHAIN MODELS FOR CHURN:
Oversimplification: Markov chains often oversimplify customer behavior by neglecting the impact
of past actions and external factors.
Limited Predictive Power: While Markov chains can provide insights into churn patterns, their
predictive accuracy may be limited due to the aforementioned assumptions.
Difficulty in Incorporating Customer Attributes: Incorporating customer-specific attributes
(e.g., demographics, usage patterns) into a Markov chain model can be challenging.
ALTERNATIVE APPROACHES:
To address the limitations of Markov chain models, more sophisticated techniques such as:
Hidden Markov Models (HMMs): Allow for unobserved states, making them better suited for
capturing underlying customer behaviors.
Survival Analysis: Focuses on predicting the time until an event (churn) occurs, considering
various customer-specific factors.
Machine Learning: Offers a wide range of algorithms (e.g., logistic regression, decision trees,
random forests) to model churn based on various predictors
In conclusion, while Markov chains can provide a basic framework for understanding churn behavior,
they should be used in conjunction with other techniques to obtain a more accurate and comprehensive
picture of customer churn.
2. How do you define churn for the case study given? Consider Exhibit 6. Can the recency state
13 is considered as churn state. Why? Give valid reasons.
In the case of Exhibit 6, churn is not defined based on a recency state of 13. A recency state of 13
likely indicates a customer who has not interacted for a period, but the specific timeframe that
constitutes churn depends on the business context.
Here are some reasons why a recency state of 13 might not be considered churn:
The threshold for churn might be higher than 13. Businesses may define churn based on a
longer period of inactivity, such as 30, 60, or 90 days.
The business might have a subscription model. For subscription services, churn might be defined
as a customer canceling their subscription, rather than simply not interacting for a period.
Overall, it's important to consider the business context when defining churn. The appropriate recency
threshold for churn will vary depending on the industry, business model, and customer behavior.
Can recency state 13 be considered churn?
No, based on the current threshold of 6, a recency state of 13 is not considered churn. However, it's
important to note that the threshold can be adjusted based on business needs.
Why?
The threshold of 6 is defined arbitrarily in this example. A business analyst should consider factors like
the typical customer purchase cycle, subscription billing cycle, or marketing campaign cadence to
determine an appropriate churn window. For instance, if a company typically runs marketing
campaigns every month, a recency state of 2 (indicating a customer hasn't interacted in the last 2
months) might be considered churn.
IN CONCLUSION:
Churn is typically defined as a customer stopping using a product or service.
The specific recency state that constitutes churn depends on the business context.
A recency state of 13 may or may not be considered churn depending on the defined churn
threshold.
Businesses should carefully consider factors like customer behavior and marketing cycles when
defining churn.
3. Use Exhibit 7, find out how many months customers in non-absorbing states (1-12) reach
Churn state 13.
I've analyzed Exhibit 7 and found that customers in non-absorbing states (1-12) reach Churn state (13)
after 3 months
[Link] is a hypothetical state of 1000 customers in State 1. What would be the distribution of
1000 customers over a period of 4 months.
Distribution of 1000 Customers Over 4 Months
Problem: Given a hypothetical state with 1000 customers, determine a potential distribution of these
customers over a 4-month period.
ASSUMPTIONS:
We don't have specific data about customer behavior or industry.
The goal is to provide a general, hypothetical distribution.
Factors like seasonality, product lifecycle, marketing campaigns, and economic conditions can
significantly influence the actual distribution.
Potential Distribution Models
Without specific data, we can explore some common distribution patterns:
1. Uniform Distribution:
Equal number of customers in each month.
Distribution: 250 customers/month.
2. Linear Growth:
Steady increase in customers over time.
Example:
Month 1: 200 customers
Month 2: 300 customers
Month 3: 400 customers
Month 4: 500 customers
3. Exponential Growth:
Rapid increase in customers over time.
Example:
Month 1: 100 customers
Month 2: 300 customers
Month 3: 600 customers
Month 4: 800 customers
4. Seasonal Variation:
Fluctuations in customer numbers based on seasons.
Example:
Month 1 (Winter): 200 customers
Month 2 (Spring): 350 customers
Month 3 (Summer): 300 customers
Month 4 (Autumn): 150 customers
Factors Affecting Distribution
To provide a more accurate distribution, consider the following factors:
Industry: Different industries have varying customer acquisition and retention patterns.
Product Lifecycle: New products typically have a different customer acquisition curve than
established ones.
Marketing Campaigns: Promotional activities can significantly impact customer acquisition.
Economic Conditions: Economic factors influence consumer spending and business decisions.
Customer Behavior: Understanding customer buying patterns and preferences is crucial.
Data-Driven Approach
For a precise distribution, gather and analyze relevant data:
Historical customer data: Analyze past customer behavior to identify trends.
Market research: Understand customer preferences, competition, and industry trends.
Sales data: Track sales performance to correlate with customer acquisition.
By combining these factors and analyzing data, you can create a more accurate distribution model for
your specific situation.
[Link] friends Karan and Arjun purchase items from Flipkart from April 2013. Karan purchased
products every month since then, except in August 2013, whereas purchased his next product in
September 2013. From months to churn calculated in Q2, calculate the estimated remaining
lifetime for both Karan and Arjun at the end of September 2013
Understanding the Problem
Before we dive into calculations, let's clarify the problem:
Data points: We have purchase data for Karan and Arjun from April 2013 to September 2013.
Goal: Calculate the estimated remaining lifetime for both at the end of September 2013 based on
their purchase behavior in Q2 (April, May, June).
Assumptions: We'll assume that the purchase behavior in Q2 is representative of their future
behavior.
Calculating Remaining Lifetime
Key Metrics
To estimate remaining lifetime, we need to calculate the average purchase frequency for each user in
Q2.
Karan: Purchased in April, May, and June.
Arjun: Purchased only in April.
CHALLENGES
Short timeframe: Using data from only one quarter might not be representative of long-term
behavior.
Missing data: Arjun's data is limited, making it difficult to estimate his purchase frequency.
Definition of "remaining lifetime": It's unclear if we're calculating the number of months until
their next purchase or a longer-term estimate.
Estimations
Assuming "remaining lifetime" means the number of months until their next purchase:
1. Karan:
Purchased 3 times in Q2.
Average purchase frequency: 1 purchase every month.
Estimated remaining lifetime: 1 month (assuming he'll purchase again in October).
2. Arjun:
Purchased 1 time in Q2.
Average purchase frequency: 1 purchase every 3 months.
Estimated remaining lifetime: 2 months (assuming he'll purchase again in December).
Important note: These are very rough estimates based on limited data. A more accurate prediction
would require a longer data history and potentially more sophisticated statistical models.
ADDITIONAL CONSIDERATIONS:
Customer Lifetime Value (CLTV): If the goal is to estimate the total value a customer will bring
over their lifetime, more complex models are needed.
Churn Prediction: Predicting when a customer will stop purchasing altogether requires different
techniques and data.
