• Change Language

Initial insights on the impact of COVID-19 in call centres

This page is designed to provide data on trends we are seeing across our ecosystem. We will continue to update this page as new data become available.

We are happy for this data to be used to support research, and we ask that data is accredited to Afiniti.

Please contact us if we can provide further support at letushelp@afiniti.com or press@afiniti.com

Executive Summary

In this analysis, we examine how churn models have been impacted by the changes caused by COVID-19. Our hypothesis was that predictive models may have been impaired by changes in customer behaviour.

Our analysis shows that churn models have in fact become more predictive post-pandemic than they were previously; the existing models even more accurately predict which customers are most likely to churn. This means that businesses can develop highly targeted retention strategies focused on the highest churn deciles, which tend to include younger, more nimble clients (single, low-tenure, single-product, apartment-dwelling).

Upon closer inspection, however, a potentially more important trend emerges: churn rates have been suppressed across almost all churn deciles, disproportionately affecting the lowest churn deciles (up to 80% drop in relative churn). These deciles tend to be older, risk-averse customers (multiple lines and products, high-tenure, living in houses).

While this suppressed churn is a positive in the short-term as it reduces the total number of cancellations, in the medium-term it could unwind, leading to significantly higher than average churn from the more valuable customers in the lowest churn deciles. Given the recession we will face as we emerge from the COVID-19 crisis, it is likely that the actual churn distribution for 2020 will be at least as high as 2019 (assuming that shelter in place policies are relaxed at some point). This would imply that the lowest churn deciles will churn over 80% more than 2019 averages towards the end of the year.

If this “Latency” hypothesis is correct, businesses should view this period as an extra opportunity to preemptively retain these higher-value customers who would have otherwise churned, and generate new offers and improved interactions to minimize the impacts of the latent churn unwinding.

Download PDF download.svg

Churn Model Predictivity Data

In order to assess the impacts of COVID-19 on churn model predictiveness, we looked at the observed churn per churn decile of Telcos in the US, UK, and EU. The graphs below compare churn deciles (x-axis), which segment customers into deciles according to their individual churn score (10 = customers with highest predicted churn probability, 1 = customers with lowest predicted churn probability), with observed churn (y-axis). Good models have a much higher observed churn for decile 10 relative to decile 1.

We normalized the pre-COVID-19 and March curves (both to an average of 1) to compare the relative model performance more clearly.



Monthly Churn Rates By Churn Risk Decile. Normalized to clients’ average churn for the period (each curve is normalized to an average of 1). “Deciles”: churn propensity scores are sorted by descending order and bucketed into deciles.

In these graphs, we can clearly see that the models performed better in March than they did in the period before the pandemic; the highest risk deciles show an even higher observed churn than they did pre-COVID-19. This suggests that COVID-19 has amplified customer behaviours rather than changed them, suggesting that businesses are now able to better identify high risk customers and build retention strategies accordingly.

To better understand this effect, we examined the average customer churn profile to see what the differences were between pre-pandemic churners and churners in March.

Churn Profiles

Customers who churned in March 2020 tended to be:


Lower Tenure

More likely to live in an apartment

More likely to be single or without a family account

Less likely to be a premium customer (lower monthly profitability)

More likely to have month-to-month pricing plans (mobile products)

More likely to have single products (less likely to have bundles)

More likely to have recent disconnects or port-outs from account

More likely to have low usage of products

Less likely delinquent or in collections

This profile has a significant overlap with the highest churn deciles (single products, low tenure, low usage), and explains the model outperformance. It also gives a hint as to why behaviours have been amplified rather than changed; this customer profile is most likely to be affected by the closures due to COVID-19, and therefore want to reduce their costs. Conversely, older customers with premium pricing plans (and generally in lower churn deciles) are less likely to be impacted and more likely to be risk-averse due to dependencies on mobile and fixed-line connections under work from-home policies. This effect is confirmed by comparing absolute churn by deciles.

Absolute Churn Rates by Decile



Monthly Churn Rates By Churn Risk Decile. Actual Churn Rates have been normalized to clients’ average Churn Rate in the pre-COVID-19 period. “Deciles”: Churn propensity scores are sorted by descending order and bucketed into propensity deciles.

The graphs clearly show that absolute churn has dropped across almost all churn deciles. This is in line with what we observed in our call volume analysis, where we saw a dramatic drop in retention volumes and an increase in abandon rates.

The graphs also confirm that the drop in churn from high churn decile customers is significantly lower than that of lower churn decile customers. This makes sense given the outperformance of the models and again reveals an uneven suppression of churn across the customer base.

Latent Churn

In order to investigate further, we examined the percentage difference in observed March churn versus pre-COVID-19 levels for each churn decile:



The graph clearly shows that the lowest decile customer churn reduced drastically as compared to higher deciles, showing up to a 75% drop in observed churn. These lower churn deciles generally correlate to higher value customers, with long tenures, families, and multiple products. This offers a clearer view of the magnitude of difference in behaviour between the generally younger, less valuable customers affected by COVID-19 in the higher deciles, and the older, more valuable customers in the lower deciles.

Given the economic impact that shelter-in-place policies have had on medium-term growth and employment, it is unlikely that these suppressed churn rates will hold. Assuming that churn rates for 2020 are at least the same as 2019, this would imply that the back end of the year will see a significant rise in churn to compensate for the current reduction. The lowest churn deciles would therefore see the highest increase in churn from these latent churners who have not been active in the initial stages of the pandemic.

This should be seen as a second chance for businesses to retain these customers that would have otherwise already churned. Predictive models and customer engagement strategies will need to be refined to preemptively target these customers, and prepare for this latent churn to unwind.

Closing Notes

Our analysis shows that there is significant churn suppression during shelter-in-place policies, primarily driven by a combination of structural changes (closure of retail stores, cancellation of truck-rolls, work-from-home policies) and risk-aversion (evidenced by the customer profiles of churners and non-churners).

Our clients will need to develop strategies to manage their churn risk post-pandemic now in order to weather the subsequent crisis. Our analysis provides strong evidence of considerable latent churn amongst the lower decile customers who tend to be families with high usage, multiple products, and long tenures.

Identifying and retaining these latent churners will require new targeted models and personalized interactions which adapt to the evolving conditions, including dynamic offers and optimized customer experiences during both reactive and proactive stages.

The Afiniti Advanced Analytics team is available to support our clients in building models to identify influenceable latent churners and implement decisioning (NBO) strategies to maximize their retention. We are also here to support ad-hoc analyses or best-practices on hypothesis testing, developing new offer strategies and AI models, and implementation and performance measurement.

This is the second in a series of COVID-19 Impact Insights that Afiniti’s Advanced Analytics team will be releasing to help our clients and global community navigate the changing pandemic environment and make data-driven decisions using emerging trends. We will continue to monitor latent churners as future trends emerge. Future reporting may dive deeper into product-level churn to enable customers to start developing decisioning strategies.

Annex: A Note on Churn Prediction Models

Churn Prediction Models are an essential part of our clients’ churn management. These models work by scoring the customer base based on the likelihood of churn, and then ranking and segmenting them into propensity deciles, such that that the top decile, 10, represents the 10% of customers with the highest predicted probabilities of churn, and the bottom decile, 1, represents the 10% of customers with lowest risk of churn.


Monthly Churn Rates By Churn Risk Decile. Actual Churn Rates have been normalized to the clients’ average Churn Rate in October 2019. “Deciles”: Churn propensity scores are sorted by descending order and bucketed into propensity deciles.

The accuracy of a churn prediction model is evaluated by measuring actual churn and predicted churn. In a decile analysis, a well-performing churn model is one that is stable through time and where the observed churn is highest in the high churn deciles and monotonically decreases along with the churn deciles in a “Staircase Effect”.

Shown here is an example of a churn decile analysis for one of our US Telecommunications Clients. As can be seen, this is a very well performing churn prediction model, since the 10th decile has a significantly higher observed churn (3.8x average), which monotonically decreases down to decile 1 (0.04x average). Furthermore, the performance is very stable between October ‘19 and January ’20.

COVID-19 Best Practices arrow-red.svg

How We Can Help arrow-red.svg