Effectively reaching out to potential customers is a key for insurance professionals to reach the better insurance renewals rates and, consequently, better sales. Nobody wants to be bothered by calls suggesting a change just after they’ve renewed their existing policy, and customers looking for a change may not be aware of the products available that better suit their needs.
Our Data Scientist, Sarah Davies worked on the challenge to figure out when is the best time to reach out to insurance customers. She worked with a dataset of records, including company names and when they first joined their current insurer. Sarah was given the freedom to use any other data she could get her hands on, and here are her findings.
“Companies House turned out to be a treasure trove, as every registered company must submit information to Companies House and ensure it is up to date. These records are then easily accessed by company number.
Finding company numbers, however, was non-trivial. Many companies’ insurance policies weren’t held under their registered name, for reasons ranging from dropping the ‘ltd’ at the end to the policy being under the name of (presumably) a director. As the model is targeted at SMEs who generally won’t have large accounting departments to sort such issues, the initial match rate was quite low (<20%). However, this still left a large and fairly representative sample (28k) to perform initial analysis on, so I decided not to invest time into trying to increase the match rate.”
At the early stage of her project Sarah found that what leapt out initially from the data was that most companies in the dataset bought their first policy shortly after incorporation. However, for companies that joined their insurer over a year after incorporation the renewal dates spread out. This is due companies switching policy before their previous term had ended, probably for reasons such as moving into a bigger office or buying more expensive equipment. To dig into this trend further, Sarah defined ‘drift’ to be the number of days between incorporation anniversary and insurance renewal date. She then examined how the distribution of drift varied according to various factors.
Contrary to anecdotes about New Year being an incredibly hectic time in the insurance industry, renewals in this dataset were heavily skewed towards the summer. There were also unexpected trends when plotting incorporation month against drift. Companies incorporated in summer were not only faster on average to purchase insurance after incorporation, they also were less likely to switch midyear than winter companies (whose renewal dates are approximately normally distributed for ‘old’ companies).
There was also a trend that companies incorporated during December tend to wait until January to purchase insurance, presumably as founders are too busy being festive to want to think about insurance.
Industry also had an effect on drift, especially for younger companies – median renewal date for young management consultancies was 32 days, whereas 50% of recruitment consultancies are still uninsured after 2.5 months (median 77 days). This is of great importance when planning how to prioritise contacting new companies, as for some industries there’s a very small window where contact is effective.
Sarah explains, “initially I went for off-the-shelf scikit learn regression models to tackle the problem. Whilst these performed fairly well in terms of the standard performance metrics, they had a problem: the estimates tended to be over-predictions, which in context is unacceptable — the window for contact closes as soon as the SME buys a new policy. With that in mind, I switched to using an asymmetric loss function that penalised overestimates more heavily than underestimates. This gave far more useful predictions — after some initial tuning, only 10% of predictions were overestimates (and even these were generally <10 days over) and 50% of predictions were within the two months prior to the actual renewal.
Initially, this model was just an exploratory exercise into what additional information we could predict about a company’s insurance needs using open data. The final model actually yields some really useful insights. The distribution of residuals has really exciting implications, as brokers can narrow the time when each company is on their ‘to contact’ list down to 70 days (20% of the year) and still make contact for over 50% of renewals. This would enable better prioritisation of clients, as they could be categorised into levels based on expected time until they’re next due to renew their policy.
For new companies being able to predict when they’ll purchase their first insurance is key, as SMEs will often stay with their initial insurers for long periods. Selling to newly incorporated companies can be tricky — they need to be adequately established to need insurance, but not have already bought their insurance. The model’s predictions of drift can be used to identify where this narrow window is likely to fall, so insurers can anticipate the business’ needs and reach out to them at the right time.
In terms of brokers’ relations with clients there’s another important use — 85% of predictions are within half a year of the true renewal date, so my model could be reliably used to identify companies with more than 6 months until their next renewal. Applying sales pressure to these companies is likely to harm their relationship with the insurer, but they do present opportunities to build relationships by checking whether they’re happy with their current cover. Alternatively, these SMEs may need a higher level of cover but be avoiding looking at alternative policies until their renewal dates, so would appreciate guidance on which policies might better fit their needs. Predicted renewal dates allow insurers to anticipate a potential client’s current needs and use a nuanced approach to sales, which is the key to building effective relationships.”
Written by Anna Kurmanbaeva