Why Am I Losing Customers? Part 1 - Technical Understanding Oriented Towards End Customer Behaviour
Introduction
[1] Churn or "customer churn", defined as switching services between providers is an useful term for losing customers due technical/functional/commercial considerations. due direct implication to provider's benefits is a widely investigated telecom area, considering the proper management or [2] churn management, the process to retain profitable customers. A basic churn management is establish processes to deal with high level understanding of how is the behavior of the end customers/subscribers, this behavior includes technical components and actions that the subscriber performs before leaving the provider.
[2] The annual telecom provider customer churn rate initiate from 25% for Europe to 48% for Asia, [3] 2% per month for mobile providers, [4] 26% as an annual average and [6] 30-35% annually after COVID. Considering these numbers a highly relevant point to understand what is happening in the subscriber's network, even more when [5] figuring out how to deal with customer churn is the provider's survival key.
This understanding is precise considering that retaining a customer is significantly cheaper than acquiring a new one, [2][7] acquisition can be 5 to 10 times higher than retention and [6] the churn rate may continue to rise.
All these concerns lead to a single question: Why Am I losing customers?
Data
For the technical understanding oriented towards end customer behaviour we require valuable data, this was collected using Axiros' solutions, as these solutions enable historical and aggregated data per individual device and network hierarchy, ensuring the proper granularity for the churn understanding.
The total amount of data is a matrix of 8687 by 475 columns containing 410 features. These features are the result of complex aggregations mechanisms, the raw data was collected from real devices and different specifications, including device hardware, WAN interfaces (health and usage for optical or ethernet), WiFi quality, WiFi usage and occurrence of reboots and factory resets, considering an historical storage and relying on intermediate aggregations.
This intermediate representation or KPIs are part of Axiros' solutions, the resulting features are obtained (again) by aggregations, some based on normal behavior and others by log-normal behavior identified before any treatment, the aggregations are usually based on time series, comparison between devices, statistical analysis for anomaly detection, and the last daily aggregation or last value.
Procedure
The churn understanding is based on exploratory data analysis using:
Graphical analysis
Correlation analysis
PCA dimensionality reduction
Identification of hypothetical causes based on scientific and experience
Non parametric statistical tests
Considerations and limitations about this article
Some features are named FTTH in the charts but are optical features, regardless of the optical technology or market
Ethernet is the WAN interfaces, no LAN Ethernet was monitored
Results
Due the number of features the comparison between each element is performed by specification (Hardware, Events, Traffic, WiFi, etc), considering the possible relationship and implication between each specification:
Hardware is compared to all other specifications, hardware includes CPU and RAM usage
Events are compared to all the other specifications, events includes reboots and factory reset events
Traffics are compared between each traffic specification, traffic includes Ethernet, optical and WiFi traffic
WiFi is compared between every WiFi specification, WiFi includes hosts rates, coverage, invariability, traffic and interference
Correlation and hypothetical causes
The following hypothetical causes were identified after the correlation analysis between specifications, due the number of features we do not present graphics (only the knowledge obtained):
Hardware
Higher traffic demands more CPU and RAM, essential for stability
RAM is critical for buffering and maintaining device performance
Reboots (BOOT event) free up memory (RAM)
Hardware influences possible WiFi rates but not coverage or connection stability
WiFi 2.4GHz and 5GHz traffic require significant CPU resources
Continuous WiFi scanning and diagnostics demand additional RAM
WiFi 5GHz stability depends on RAM availability
Events
Devices with stable WAN Ethernet connections reboot less often
Heavy Ethernet and WiFi users tend to avoid reboots unless interference or poor coverage occurs
WIFI 2.4GHz and 5GHz interference/noise often trigger reboots
Users experiencing stable WiFi performance typically require higher speeds instead of rebooting
Ethernet and Optical Traffic
Increased uplink/downlink usage demands more network capacity
Most WAN traffic originates from WiFi usage (both 2.4GHz and 5GHz)
Ethernet health influences traffic stability and experience
WiFi 2.4GHz and 5GHz Hosts
Strong RSSI improves coverage, but too many devices per AP degrade rates
Mesh systems outperform high-gain single APs for better performance
Interference from other networks disrupts stability and rates
Band steering improves overall experience by shifting traffic to 5GHz
Coverage should be optimized for both 2.4GHz and 5GHz to balance stability
WIFI 2.4GHz and 5GHz Traffic and Traffic Health
Less interference increases traffic, improving streaming and performance
5GHz optimization reduces congestion on 2.4GHz
A strong relationship exists between interference in 2.4GHz and 5GHz—optimizing one benefits both
WiFi uplink and downlink performance must be monitored separately
It can be seen that many events are monitorable and automatable to improve the final experience, but some other points require analysis before a massive device purchase. Essentially, the behaviour of the device before it is deployed is highly important considering Hardware and WiFi behaviour, (probably) legacy or end of life devices are not the suggested option for a proper churn management.
The good news about all this is that Axiros has solutions for all of the above, including LAN/WLAN and WAN sides, enabling a full optimization of these components. This explains (significantly) the churn reduction using Axiros solutions.
Churn and End Customer Behaviour
After PCA a significant reduction from 410 features to 51 (variance ratio of 0.95967) was performed, the number was identified by the graphical analysis (Figure 1)
(Figure 1 - Number of PCA components vs Explained Variance Ratio)
(Table 1 - Churn/No churn proportion)
The churn distribution in the dataset was imbalanced and near to 1:9 (Table 1), these providers have an average of 2.5% churn per month considering the data is from the last 4 months.
(Figure 2 - Box plot of 51 PCA Components)
After graphical and normal test analysis (Figure 2) we identified the features are not part of a normal distribution enabling non parametric statistical tests.
We tested each 51 distributions considering the churn/no churn category using Mann–Whitney U identifying 37 components with a p-value < 0.05 (Table 2), and considering each component is significantly different, this means the behaviour of these components are significantly different when the devices is tagged as churn or no churn.
(Table 2 - Relevant PCA components by P-Value)
After the component identification we return to the original features using the TOP 15 relevant features by PCA component, obtaining 96 original features, from these features we identified 65 original features that are significantly different between the 2 groups (churn/no churn) with a p-value < 0.05 (48 features with a p-value of 0.000 and 17 features with p-value >= 0.001 and p-value <= 0.049).
According to the found features the churn probability is based on historical, last values, anomalies and severity behaviors:
WiFi Traffic, Health and Host Health → A different behaviour is detected for 2.4GHz and 5GHz for churn users vs no churn users
Events Health → A different behaviour related to TR069 events is detected for churn users vs no churn users
WAN Traffic and Health → A different behavior related to the WAN access and traffic usage is detected for churn users vs no churn users
Hardware Health → A different behaviour related to CPU and Memory usage is detected for churn users vs no churn users
(Figure 3 - Box plot of 65 Found Features)
Performing the revision of each feature (Figure 3) and the average comparison to determinate the churn/no churn behaviour we identified:
1. WiFi Traffic & Health Indicators
No-churn users show higher WiFi 2.4GHz and 5GHz downlink and uplink traffic usage, implying better network utilization
A strong correlation exists between WiFi health and user retention—good WIFI 2.4GHz and 5GHz health cases are significantly higher for no-churn users
No-churn users also show a higher number of WiFi host health good coverage cases, leading to better network experiences
Poor WiFi performance (higher anomalies, interference, or weak signals) is associated with higher churn probability
2. WAN Traffic & Health Contributions
No-churn users have higher Optical downlink traffic usage, showing a direct relationship between higher fiber traffic demand and retention
Stable Ethernet health and higher Ethernet traffic usage correlate with lower churn rates
A higher optical health variation is linked to better retention, while poor optical health contributes to churn, considering variable optical is a lower problem compared to all the time optical degraded users
3. Event & Anomaly Trends
No-churn users experience fewer critical event anomalies, indicating more stable device behavior or less manual reboots
They have a higher number of good event health cases, suggesting fewer unexpected connectivity issues with less manual reboots
Churn users tend to have higher event severity cases, likely causing frustration and dissatisfaction, considering an increment of reboots
4. Hardware & Resource Utilization
No-churn users show lower CPU and memory health anomalies, indicating more efficient resource utilization
Good memory health correlates with better user experience and lower churn probability
High CPU or memory anomalies indicate performance degradation, often triggering higher churn rates
5. Coverage & Interference Management
No-churn users have higher 5GHz WiFi usage, benefiting from less interference and higher throughput
Poor WiFi placement and coverage contribute to churn, as seen in lower min values of WiFi 5GHz host coverage for churn users
WiFi 2.4GHz bad coverage cases are more frequent in no-churn users, possibly due to more intelligent band-steering or proactive troubleshooting
6. Stability vs. Variability Indicators
No-churn users show higher invariability in WiFi host health and event health, indicating a stable experience over time
Churn users experience higher anomalies in Ethernet, WiFi, and event severity, signaling unpredictable network issues
Coming Soon: Part 2 – Churn Analysis and Correlation Insights
Written by Cesar Quirga, M.Sc
Cesar Quiroga is a Project Definition Engineer at Axiros, specializing in defining projects scopes based on ISP necessities and ensuring that technologies align with evolving industry demands. He combines technical expertise with analytical skills to drive impactful solutions. Currently, his focus is on machine learning, statistical analysis, and data-driven decision-making, enabling him to design innovative solutions that enhance the end-user experience.
References
[1] Berson, A., & Thearling, K. (1999). Building data mining applications for CRM. McGraw-Hill, Inc..
[2] Hung, S. Y., Yen, D. C., & Wang, H. Y. (2006). Applying data mining to telecom churn management. Expert Systems with Applications, 31(3), 515-524.
[3] Berson, A., Smith, S., & Thearling, K. (2000). Customer retention. Building data mining applications for CRM.
[4] Oghojafor, B., Mesike, G., Bakarea, R., Omoera, C., & Adeleke, I. (2012). Discriminant analysis of factors affecting telecoms customer churn. International Journal of Business Administration, 3(2), 59.
[5] Mattersion, R. (2001). Telecom churn management. Fuquay-Varina, NC: APDG Publishing.
[6] Chang, V., Hall, K., Xu, Q. A., Amao, F. O., Ganatra, M. A., & Benson, V. (2024). Prediction of Customer Churn Behavior in the Telecommunication Industry Using Machine Learning Models. Algorithms, 17(6), 231.
[7] Yulianti, Y., & Saifudin, A. (2020, July). Sequential feature selection in customer churn prediction based on Naive Bayes. In IOP conference series: materials science and engineering (Vol. 879, No. 1, p. 012090). IOP Publishing.