Deploying analytics to predict and prevent customer churn is critical for any SaaS business looking to maximise revenue growth and retention. When done effectively across a large customer base, that analytics approach provides two major upsides. First, there's the direct financial benefit of capturing revenue that would otherwise be lost through un-renewed contracts. Second, and just as importantly, a model can identify not just which customers are most at risk of churning, but the factors driving that risk.
Our client, a $5bn cyber security software business, sought those insights to take proactive action across their customer success, sales, and product teams to improve retention.
The business had delivered impressive and consistent growth over the past five years but at the same time their churn rates had crept up in tandem to nearly 8%, four points higher than QuantSpark’s SaaS industry benchmark of 4%-5%, or a Gross Retention Rate (GRR) of 95-96.
While GRR is a useful comparative, it runs the risk of obscuring the true financial value of reducing churn levels. Increasing GRR from 92 to 95 doesn’t sound like much, but for a business turning over $500m+ annually, each percentage point is worth millions of dollars to the company’s top line.
Alongside the lost revenue pain point, those high churn rates belied an operational issue: the company realised there was no standardised process for tracking renewal actions taken by their 300-strong Customer Success (CS) team, or for assessing the effectiveness of those actions.
They needed a scalable, predictive approach to proactively identify at-risk customers, a methodology to systematically introduce effective interventions as part of their CS playbook strategy, and an easy BI interface to monitor churn levels across the business.
A project of this size and potential complexity required buy-in from multiple stakeholder groups, from business & tech teams to C-suite sponsors. To be successful, the partnership required 3 crucial elements:
When introducing new modelling approaches, the first hurdle is always to engender trust in the model itself. Data science teams need to have confidence the the data we’re outputting, that the business logic used is correct. CS teams meanwhile are naturally focussed on the ‘so what?’ They need to believe the risk scores that are generated by the model and in turn rely on the efficacy of the suggested playbook action.
The most straightforward way of achieving that trust is simple: while we rely on client analytics team to provide cleaned, quality data, we format and structure in a manner the business is most used to, rather than trying to reinvent the wheel. For CS teams, it’s about showing the model output can achieve the same level of accuracy as their existing method - whether that’s human instinct or a simple heuristic - but much faster. Once that trust is built, attention can be turned to delivering great quality of insight that will augment and enhance customer success strategies.
Building an accurate and effective model is just one half of the project: to be commercially successful, there needs to be widespread uptake of the new approach by the Customer Success reps. For this project the client assigned a working group of 10-15 CS managers who worked with our team to sense check the model’s output. They would use their own experience and deep understanding of their book of business to validate our prioritisation of customer risk scores. We also worked with them to understand exactly how CS reps would interact with a BI dashboard, creating early wireframes that would boost end user adoption.
Once the working group were assured of the value of our approach, they acted as evangelists for the new model, championing its effectiveness within their own regions and teams.
The company’s exec team shared the vision for a holistic proactive renewal strategy across the business, seeing this as not only an opportunity to predict churn risk, but to gather granular information on their customer base to enable CS teams to have much more useful conversations with their customers to maximise renewal chances.
Of course, any senior executive worth their salt wants to see actionable insight as soon as possible, so the QuantSpark team worked to demonstrate how our model’s findings would help mitigate churn. For example, we spotted that instances where customers hadn’t booked a CS meeting in 3 months directly contributed to increased risk scores. The executive team could immediately calibrate a strategy for Customer Success Managers (CSMs) to book monthly check-ins with their customers across the business.
By building the relationship at a senior level, we could deliver up front impact whilst developing longer-term value.
This phase focused on identifying and optimizing the key data inputs, known as features, that are most predictive of customer churn risk.
We examined numerous data sources provided by the client, including customer profile attributes, product usage metrics, support cases, and service interactions. Our analytics team performed extensive exploratory analysis on this data to determine which specific features showed the strongest statistical correlation with customer renewal outcomes.
Key outputs of this phase included:
Engineered features customised to this client's unique product and customer landscape. For example, features related to product configuration were prioritised based on the "set and forget" nature of their SaaS offering.
Analysis identifying the most significant drivers of churn risk based on the data. For example, lack of customer engagement over a period of 3+ months was found to be a top predictor of renewal risk.
A refined "renewals dataset" containing optimised features engineered from multiple sources, ready for modelling. Reducing noisy, irrelevant features is crucial so the model can find meaningful signals related to churn.
This detailed, iterative feature engineering process matters because feeding clean, tailored data into models is vital for accurate predictions. It provides a solid data foundation before embarking on modelling.
Attributes like customer size, location, and industry that indicate baseline propensity to churn. They cannot be directly targeted but companies should factor mitigation into their CS strategies.
Behaviours like low product usage and lack of engagement that signal churn risk but require investigation to diagnose the root cause.
Specific incidents like configuration issues and service problems that can be clearly linked to tailored follow-up actions.
With our refined features dataset ready, we tested different machine learning algorithms to predict customer renewal risk, settling on a Long Short-Term Memory (LSTM) neural network model.
The LSTM excelled at finding complex patterns in how risk evolves over time for each customer. It takes a historical time-series view of the engineered features as inputs to make personalised risk predictions.
Key outputs included:
A production-ready LSTM model capable of ingesting live data to generate updated customer risk scores.
Quantitative metrics showed the model performed well compared to benchmarks, with room for improvement as more data comes in.
Code and infrastructure to run the model daily and serve results to frontline users.
Choosing the right modelling technique matters because predicting churn is a nuanced problem. The LSTM neural network provides the sophistication required to uncover signals across thousands of intricate customer histories. This depth of insight simply isn't possible using simpler off-the-shelf models.
Choosing the right evaluation metric depends on the business context. It’s very easy to obsess over the model’s
Accuracy simply the overall percentage of correct predictions - but for predicting churn, precision and recall are often more informative than pure accuracy.
Precision reflects how reliable the model's high-risk flags are - low precision means lots of false alarms. Maximizing precision minimises wasted retention efforts on customers not actually at risk.
Recall indicates how many real churners the model catches - low recall means missed opportunities. Maximizing recall helps ensure proactive outreach to all potential defectors.
For reducing churn, balance is ideal - consistently flagging most at-risk customers without too many false positives. When in doubt, err on the side of higher recall to avoid missed saves.
Investing in a data-driven churn prevention strategy is an investment in the future, delivering benefits that will compound over time. As our client’s executive team realised, a holistic renewals strategy can enhance their customer experience far beyond the direct financial benefit of pro-actively saving at-risk customers.
By understanding the underlying drivers of churn risk and adapting customer success plays accordingly, CS teams can take that insight and have far more engaging conversations about customer needs. Happy customers that see you address their needs are much more likely to increase spend and adopt new products. Essentially, you move from reactive firefighting to proactive nurturing, putting the business in a much better position to increase expansion revenue.
By turning customer insight into customer loyalty, our client can expect to increase their GRR by 2-3 points, adding up to $30m to their bottom line annually, whilst deepening customer experience over the long term.