Analysis and retention planning on customers leaving an energy retailer

In the year the data covers, 1,595 of 16,096 customers left, a churn rate of 9.9% or roughly one in ten. However, counting customers is only half the picture. Customers differ significantly in what they are worth, so throughout this analysis I also track net margin (the profit a customer generates in a year after costs). Any retention budget has to pay for itself against this lost profit.

The two charts below size the problem both ways. The left chart counts customers; the right chart adds up profit.

The leavers took €399k of the €3.51m yearly profit pool with them. That is 11.4% of the profit from 9.9% of the customers, meaning the typical leaver is worth slightly more than the typical stayer. Thus, the company is not simply losing its smallest accounts.

Two more observations came out of the first pass over the data. First, churn is nearly flat across customer size, so large accounts leave about as often as small ones. Second, churn falls steeply with the number of years a customer has been with the company. This second observation becomes the central thread of the analysis and is tested properly in hypothesis 2.

The theory is intuitive: energy is a commodity, prices went up, and customers left. It is also testable, because the data includes every customer’s monthly prices for the whole year. If rising prices drive churn, the customers with the largest price increases should have left at the highest rates.

To verify this, I measured how each customer’s price changed across 2015, sorted all customers by that change, and divided them into ten equal groups: group 1 received the deepest price cuts and group 10 the steepest rises. The chart below shows the share of each group that left. If the theory were correct, the bars would climb from left to right.

The bars show the opposite. The statistical link between price change and leaving is 0.003, on a scale where 0 means no link at all. In fact, the group whose prices fell the most left at 11.9%, compared to 8.7% for the group with the largest rises.

Still, I do not believe price cuts cause churn. A more likely explanation is that the retention team had already been giving discounts to accounts it was worried about, and those accounts left anyway. In other words, the price data records the company’s own past rescue attempts, not the cause of the leaving. This matters for two reasons: (1) a naive prediction model would learn this pattern backwards, and (2) it is the first hint that discounts were not addressing whatever was actually driving customers out. The exit interviews in the companion article explain this further.

Section 1 showed churn falling with years on the books. A cleaner way to verify this is a cohort view: group customers by the year they joined, then compare the groups. Cohorts are useful because they separate when a customer joined from everything else that changes over time.

The pattern is clear. Customers who joined in 2013 left at 15.2%. The 2011 intake left at 10.3%, the 2010 intake at 7.2%, and the small 2007 group at 1.7%. As a rough rule, every additional year a group has been on the books halves its churn rate.

It should also be noted that this pattern is a projection. Each new intake starts at the highest churn rate, so if nothing changes about the first years of the relationship, the customer base keeps refilling with its riskiest customers and overall churn stays near 10% regardless of how loyal the older base is. Thus, growth alone will not fix the churn problem.

Energy contracts run for fixed terms, so I expected leaving to concentrate where leaving is easy: near the end of a contract. If true, this would be operationally valuable, since the renewal calendar would tell the retention team exactly who to call and when. The chart below groups customers by how much time their contract had left, and shows the share of each group that churned.

The spread across the groups is small, roughly 9.4% to 11.1%. Customers with two years left on their contract leave about as often as customers whose contract ends next quarter. In other words, customers do not wait for a convenient exit window; when they decide to leave, they leave, early-termination clauses included.

A retention budget spread evenly across 16,096 businesses would mostly be wasted, because the accounts range from very small businesses to industrial sites. To create a usable map, I used a standard grouping method (K-Means clustering) to sort customers into four groups based on how much energy they use, the size of their connection, the profit they generate, and how long they have been customers. The method finds accounts that resemble each other without being told what to look for. Whether a customer later left was deliberately kept out of the grouping, so the map can be used on day one, before anyone leaves.

The two big-business segments hold 30% of the customers but 65% of the profit. The most expensive attrition per account is in Industrial premium, where 1,178 accounts with a median profit of €715 each are leaving at 10.8% a year. The segment with the highest churn is Recent switchers at 11.8%, which is consistent with hypothesis 2: newer relationships churn more.

Everything so far describes groups. However, to act, the retention team needs a ranked list of the individual customers most likely to leave next. That is what a prediction model produces. It reads each customer’s attributes (tenure, size, profit, sales channel, prices and so on) and turns them into a single number: the predicted chance that this customer leaves.

To keep the evaluation fair, I trained three standard models on three quarters of the customers and graded them on the remaining quarter they had never seen. This held-out grading matters because any model can memorise the past; the only test that counts is ranking customers it has no answers for. Briefly, logistic regression draws the straightest possible line through the data and serves as the simple baseline, while random forest and gradient boosting are families of decision trees that can capture bends, such as the sharp drop in risk after the first few years.

The fairest plain-language comparison is to let each model flag the 10% of unseen customers it considers riskiest, and then count how many of those customers actually left.

The best model’s riskiest picks left at 35.8%, roughly 3.6 times the random-picking rate. It should be noted that the two tree-based models clearly beat the straight-line baseline, which indicates that this churn problem contains genuine non-linear patterns, most notably the drop in risk with tenure. The next chart shows what the winning model’s full ranking looks like in practice.

What is the model paying attention to? The chart below answers this directly, by hiding one signal at a time and measuring how much the ranking degrades. The answer is consistent with the hypothesis tests: profit margin, tenure, connection size and sales channel dominate. The price-change signals I engineered for hypothesis 1 do not appear in the top eight. The model was free to support the client’s theory, and it did not.

A retention discount has one benefit and one cost. The benefit is that a would-be leaver stays and keeps paying, now at 80% of their old profit. The cost is that most customers who receive the offer were staying anyway, and now they pay 20% less for nothing. Whether a campaign makes money depends on the balance between the two, and it hinges on a number nobody knows before running one: the save rate, the share of would-be leavers the offer actually convinces to stay. Retention programs typically achieve somewhere between 20% and 40%, so I simulated the whole range on the held-out customers, using their real outcomes.

The lines rarely rise above zero. Offered to the whole customer base, the discount loses an estimated €506k a year. Offered only to the model’s riskiest tenth, it still loses roughly €22k, because even at that depth roughly 64% of the discounts go to customers who were never leaving, and the profit given away outweighs the profit saved on the leavers who are caught. For the riskiest-tenth version just to break even, the offer would need to save 51% of would-be leavers, roughly double what retention programs usually achieve.

Still, one version does work. If customers are ranked by risk multiplied by account value instead of risk alone, and the 97 accounts that already lose money are excluded (there is no reason to retain unprofitable accounts with a discount), then a campaign restricted to the top 2% of the list, roughly 310 accounts, is estimated to make money.

The arithmetic is one reason to doubt the discount. The other is that this dataset records who left but not why, so a retention plan built on it alone is guessing at motive. The companion article takes up that question on a telecom dataset in which every one of the 1,869 leavers has a recorded reason. Telecom is not energy, so I treat its findings as supporting evidence rather than proof about energy customers. Still, the result is consistent with the simulation above: when leavers are asked directly, price ranks fourth, behind competitors, service quality and how customers were treated, and a discount answers only one of those.

Almost nothing above proves cause and effect on its own. Observational data can say that the newest customers leave most, but it cannot say that a better first year would have kept them. One approach to bridge that gap is to perform A/B testing. In an A/B test, a group of customers is split at random, one thing is changed for half of the group, and the other half is left alone; churn is then compared a few months later. The random split is what makes the comparison fair, because both halves contain the same mix of everything else. Each experiment below names its change, its audience and its measure.

First, the energy data is a single snapshot with one window of outcomes, so it says who is likely to leave rather than when. Second, the discount arithmetic assumes a save rate until the pilot measures one, and the €12k best case rests on roughly 80 held-out customers, so it should be treated as pilot-sizing guidance rather than a forecast. Third, the data records nothing about motive, which is why the question of motive is handled in the companion article, on a dataset that was collected to answer it. Finally, none of the patterns here are causal until the experiments run. The price-cut finding is exactly what a selection effect looks like, and it is why every recommendation above is phrased as a test rather than a rollout.