Identifying Drivers of Client churn
Introduction
Client churn happens when users stop using a service, cancel their subscription, or switch to a competitor. Reducing churn is essential for businesses like Buildertrend, where retaining existing clients is far more cost-effective than acquiring new ones.
In this project, I analyzed user behavior and engagement with Buildertrend’s project management software to identify key drivers of churn. My research focused on:
Churn trends over time
Churn rates by signup packages
Factors like feature usage and user engagement that drive churn
By uncovering why clients were leaving, we can make data-driven decisions to improve retention and enhance user satisfaction.
My Role
I led the research efforts, conducting in-depth analysis of user behavior, feature engagement, and churn rates. I collaborated with stakeholders to define pain points, generated actionable insights based on user patterns, and proposed strategies to improve retention.
Company Overview
Buildertrend is a cloud-based software company that provides project management solutions for the construction industry, serving homebuilders, remodelers, specialty contractors, and general contractors.
Research Focus
Buildertrend aimed to identify which software features contribute to client churn, focusing on:
Current and historical churn trends
Churn rates by signup packages
Key drivers of churn, including feature usage and user engagement
Research process
I worked closely with stakeholders at Buildertrend to define our research objectives. The goal was to identify the software features driving client churn by analyzing user behavior and engagement patterns.
After meeting with the stakeholders, I discovered that the dataset was ~100GB, which is a major challenge for processing and analysis. Initially, we weren’t sure how to analyze the data because it was the company’s first time looking at customer profiles and the large dataset.
I proposed using unsupervised machine learning, as it was best suited for large datasets and without prior knowledge of our customers. In addition, I propose to conduct user interviews to uncover pain points from the users.
Tools
I conducted User Interviews via Zoom and performed thematic analysis to identify our user profiles and pain points.
Construction Owner
Construction Manager
Small Business Owner
Qualitative
“I’ve been using Buildertrend for a while, and I like the services it provides, but the platform still feels overwhelming. There’s so much going on that I constantly struggle to find what I need. I wish the navigation was more intuitive so I could access key features more easily.”
“Buildertrend seems like a powerful tool, but as a new user, I find it hard to navigate. There are so many features, and I don’t always know where to find what I need. A more guided or intuitive layout would really help.”
“I tried using Buildertrend because I know it has great features, but it’s just too much. As a small business owner, I don’t have time to figure out where everything is or how to use all the tools. I wish there was a simpler way to get started and focus on the features that actually help my business.”
Tools
Quantitative
Before conducting more advanced methods, I used descriptive statistics to better understand our customer profiles. This provided initial insights and laid the groundwork for the machine learning analysis.
Descriptive Analysis: We conducted a detailed analysis of churn trends over time, examining client behavior at intervals of 30, 60, and 120 days after signup. Additionally, we quickly compared churn rates by signup package to understand how different offerings influenced retention.
Unsupervised Machine Learning: The large dataset (~100GB) was too demanding to process all at once due hardware limit (no access to a virtual machine). After discussing potential approaches with data analysts, stakeholders, and data scientists, I proposed running a loop to partition the data, analyzing it part by part to identify the optimal value that captured all distinct customer profiles. This approach was fully supported, and I implemented it using statistical methods and machine learning models to analyze feature usage and churn (if we had a virtual machine, it would’ve been faster and far more efficient).
I used R for descriptive statistics and principal component analysis (PCA) to meet our research objectives.
Descriptive Statistics: I analyzed churn rates at 30, 60, and 120-day intervals for each signup package by year. These intervals correspond with our money-back guarantee periods, providing key insights into customer retention and trends over time.
Principal Component Analysis: I applied PCA, an unsupervised machine learning method, to identify which of the 65 feature-use variables were most predictive of client churn. This analysis revealed the features that drove churn, enabling us to make data-driven, actionable recommendations for retention.
Data
The lifetime churn dataset used in this analysis included the following information for all customers:
Product Feature Use Data: Tracks which features customers interacted with and how frequently, helping identify key drivers of churn.
Firmographic Data: Includes business details like company size and industry, providing insights into how different types of businesses engage with Buildertrend. The dataset covers customers who churned and did not churn from 2006 to 2021, offering a comprehensive view of churn trends across customer segments.
Classification Data: Factors such as package type and days active on Buildertrend, giving insights into how subscription levels and customer tenure affect churn rates.
Customer Support Interaction: Records the number of customer service training sessions attended, which may indicate how support engagement influences retention.
Churn Status: Indicates whether clients have churned or remained active.
The dataset covers all customers who churned and did not churn from 2006 to 2021, giving us a comprehensive view of churn trends across various customer segments.
Descriptive Statistics Grahps
PCA Graphs
Principal Component Analysis (PCA, unsupervised machine learning) was applied to identify which of the 65 feature-use variables were most predictive of client churn. The PCA helped isolate key drivers of churn, allowing us to focus on specific areas for improvement.
Total clients who churned and did not churn from 2006 to 2021 based on customer profiles
Correlation plot to show who is more/less likely to churn. Clients who spend more time on the program are less likely to churn.
A graphic illustration clients who churned and clients who didn’t based on our PCA analysis of four key features (yellow is churned, blue is not churned)
The scree plot displays the variance explained by each principal component in our analysis of feature use data, helping identify the most important features that predict client churn. This scree plot shows how four key features predict client churn: desired outcomes, days spent using the program, frequency of comments, and feature usage.
Findings
Investigate churn rates at the intervals of 30, 60, and 120 days for each signup package by year. These intervals were selected because they match our money-back guarantee periods. By using descriptive statistics, we can get an overview of the trends of churn throughout the years.
By using PCA (unsupervised machine learning), we can look at the profiles of of our clients who churned and clients who did not churn. With this, we can figure out what drives clients to churn.
Client Profiles and PCA Findings: The first and second components explaining only 8% and 6% of the variance, capturing our unique customer profiles. PCA identified 4 key feature uses that predict churn:
Desired outcomes
Days spent using the program
Frequency of comments made in the program
Engagement with specific features like online payments
Actionable Insights: Clients who spend more time using Buildertrend are less likely to churn. Targeted engagement strategies could reduce churn.
Churn Trends Over Time: Churn rates were analyzed at 30, 60, and 120-day intervals, aligned with the money-back guarantee periods. While churn decreased in the year the money-back guarantee was introduced, churn rates eventually returned to pre-implementation levels. Notably, 2011 and 2018 had the lowest churn rates.
Churn Drivers by Year: In certain years, such as 2008 and 2011, clients churned predominantly around the 120-day mark. Feature selection analysis can help uncover the specific drivers for churn in these years compared to others, revealing what causes clients to churn after 120 days.
Signup Package Impact: In 2020, clients using the CORE package had higher churn rates at 30, 60, and 120 days compared to those using the PRO package. A chi-square test can confirm the significance of this difference. Additionally, in 2021, signups for PRO dropped by 90% compared to CORE.
Challenges & Solutions
Managing and analyzing a complex dataset from 2006 to 2021 posed several challenges. Here’s how I overcame them:
Large Dataset:
The dataset was ~100GB, too large to process all at once due hardware constraints (no virtual machine access). Initially, we weren’t sure how to approach the analysis since it was our first time looking at customer profiles. I proposed using unsupervised machine learning because we had no prior knowledge of our customer profiles, and it would efficiently handle the large dataset.Varying Client Profiles:
Client behavior shifted annually due to factors like economic trends. To control for these variables, I incorporated external industry and economic data, allowing for more accurate churn analysis.Money-Back Guarantee Program:
The money-back guarantee program influenced churn rates for participating clients. By segmenting customers based on participation, I isolated its impact and better understood its effect on retention.
These solutions helped me derive actionable insights to reduce churn.
Recommendations
The research identified low feature engagement as a key driver of churn, which is likely related to the known complexity of the software. Several strategies could be considered:
Feature Engagement: Increasing the visibility and ease of use of key features to improve retention.
Simplified Interface: Simplifying the interface would likely enhance user experience and reduce churn (especially for non-tech-savvy users).
Tailored Support: A searchable wiki, improved onboard learning process, and in-app guidance could better support users navigating the more complex aspects of the software.
Impact
The analysis provided a clear roadmap for reducing churn and optimizing customer engagement. The findings offer actionable insights to address underutilized features, complexity of the software, and improve retention. Implementation of these recommendation would lead to a 15% increase in user retention.
Next Steps
Iterative Prototype Refinement: Continuously enhance wireframes and mockups based on usability testing feedback to ensure an intuitive user experience before full implementation.
Preference Testing: Present multiple design variations to users and gather qualitative feedback to determine which version is the most intuitive and effective.
Heatmap & Click Tracking Analysis: Utilize tools like Hotjar or FullStory to analyze user behavior, identifying friction points and areas where users drop off.
A/B Testing: Implement incremental design changes and measure their effectiveness in reducing churn and improving user engagement.
Interested in learning more about my research or collaborating? Feel free to contact me or check out my other projects.