Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
Ralph Clayton on the Price Writer Podcast Ep 18
12 Plays1 month ago
Join me as I interview Ralph Clayton who took the most unusual route into insurance of any of our guests so far... He choose a career in insurance at the outset. Ralph is a true expert on pricing and data science. Discover Ralph's journey from starting in insurance pricing to embracing cutting-edge techniques in data science and cloud computing. He shares invaluable insights on how modern tools like machine learning, Databricks, and cloud services can revolutionize pricing strategies, enhance efficiency, and can lead to significant cost savings. In this detailed discussion, Ralph explains:
- Why the traditional actuarial route was not for him because of too much theory and not enough practical value.
- The advantages of using serverless computing with AWS Lambda for scalable, cost-effective solutions.
- Implementing a modern pricing architecture using Databricks, focusing on automation and efficient data practices.
Ideal for insurance pricing directors and professionals, this episode provides a roadmap to adopting sophisticated analytics and programming approaches in the insurance industry. Learn how integrating these technologies can help manage revenue, margins, and volumes effectively, all while navigating the complexities of anti-selection and moral hazards. Tune in to gain a competitive edge in general insurance pricing and lead your team towards more innovative and profitable practices.
Recommended
Transcript
Introduction and Ralph's Background
00:00:00
Speaker
Hello Ralph, welcome to the show. Hi Jeremy, thank you for having me. You're welcome, it's great to see you. So the first thing I want to ask you is, how did you get where you are today?
00:00:10
Speaker
So I think different to how most people I've seen answer this question, which is I did actually and intend to join the insurance industry. I studied economics at university and there was a couple of modules that they covered it.
00:00:24
Speaker
know, a lot on the consumer side, so measuring expected utility from having some certainty in different scenarios and whether that was worth them taking premiums. And then lot of customer behavior stuff. So covering naturally elasticity and economics, but also moral hazard and and fraud.
00:00:46
Speaker
And ah generally like the concept of risk as well, just making decisions based on uncertain outcomes. So insurance can seem to fit quite well.
Graduate Scheme and Interest in Machine Learning
00:00:57
Speaker
I started out as a yeah ah graduate on a grad scheme on working on a bike product, which suited me quite well as a keen motorcyclist at the time.
00:01:06
Speaker
The underwriters couldn't believe the premiums was paying on my bike in terms of how cheap they were. ah They were quite a risk averse book. And there's me on a 20 year old on a super bike paying a hundred and thirty quid or so.
00:01:20
Speaker
And yeah, i was very keen on spending time on upskilling and becoming the kind of the best pricing analysis that like I could be. So think like many others that originally looks like taking or looks like taking the actuarial exams and becoming qualified and and studying everything there is on on the syllabus.
00:01:42
Speaker
But after sitting and passing the first one, I did find that there wasn't really much I could apply to my job straight away. it was, granted it was financial mathematics and covered a lot of that at university as well.
00:01:57
Speaker
But looking through the syllabus, it was similar sort of story. I was and quite keen on staying in insurance pricing. I couldn't see myself moving to other parts of insurance. And I figured that if there was something on the syllabus that I could use, it'd be more efficient to just go and learn that as opposed to learn lots of other topics and and focus on exams rather than just learning what I was actually need to learn and and applying it.
00:02:25
Speaker
So that's where my focus went. more to potentially other things that like could bring value. So I'd been to a few talks and and machine learning was a topic that came up quite a lot and um generally got the impression that not many people knew what it was or or what I could do too much about it in general.
00:02:45
Speaker
So that's where I figured I would put some focus and that became useful at
Applying Machine Learning in Insurance
00:02:52
Speaker
my next role. So was at the, my first job for a little over two years and then I joined a a non-standard home insurance broker, very small, and they didn't have any analytics or budget for expensive software or anything like that. So that's where building ah think things from scratch, that's where looking into open source programming languages and machine learning actually became quite useful because and straight away had to ah to but build the initial pipelines from the Bordero files, build out the reporting and the analysis, but then
00:03:28
Speaker
go and build all the yeah predictive modeling. Machine learning was a really good use case here because there was lots of models I needed to build. ah was more interested in the insight rather than deploying them.
00:03:40
Speaker
And of course it didn't have the expensive software that was used to. You know, you can build most machine learning, like machine learning techniques quite easily in, in open source languages.
00:03:52
Speaker
So yeah, these were a great use case and decided I'd focus pick one technique, kill a lot of machine learning software will build multiple different models using different techniques and pick the best one.
00:04:05
Speaker
And i was keen on learning the more in depth. So I figured I'd pick one and and learn that. So I started with GBMs just where they generally perform the best on on tabular data. So yeah, use those and and they worked out quite well. And there was a lot of insight that we could get from those, which helped on the pricing side, and especially identify identifying key drivers on, on poor performance with the most important plots and, and things like that.
00:04:34
Speaker
Yeah. but Also the, yeah, the broker didn't have much capability in terms of what they could do with prices. So built out the rating engine. And again, that was with code and that gave you a lot of flexibility to, to do whatever you please really.
00:04:53
Speaker
So, so was at that broker for yeah about about a year or so. And what I'd built in that time was quite bit different to what i built at the underwriter
Contracting and Cloud Computing
00:05:02
Speaker
previously. So I figured there was quite a lot of value in the things that I'd learned to the wider industry. And and that was the that was my kind of step into contracting where where the aim was to bring kind of machine learning and more programming approaches to pricing.
00:05:20
Speaker
ah So the and the the project was just comparing a ah je GBM to a GLM, which was the the one I was, had been curious about the whole of the previous year.
00:05:30
Speaker
And, you know, sure enough, couple of hours, that the GBM is built and it's performing better. But the drawback is that the rating engine there was closed source software and it didn't really right support models built with other vendors.
00:05:50
Speaker
There was some There was a way to implement the models, but it was incorrect and it it didn't work. So the next project was something that was actually implement. You could implement, which is building a postcode files.
00:06:04
Speaker
So the old methodology was use expensive software was time intensive. i think it was about took analysts about two weeks worth of time to build a model. And but then when you've got multiple perils, multiple products, that's a lot of time being spent. yeah And the, the GBM methodology.
00:06:22
Speaker
was fully automated, again, used free open source programming language and, and ended up outperforming the the old method as well. so So that, that was rolled out across and all all the products. And again, as well as postcode, and postcode classification, it could be used for motor vehicle classification as well.
00:06:43
Speaker
Yeah. So by this point, I'm pretty sold on what I've spent my time learning and and doing. It's showing to have a lot of value. So that's where I've committed and gone all in into learning about data science in more detail. So I've enrolled in my masters and I'd started doing these at the broker anyway, but personal projects started ramping up those and really and trying to learn more about the kind of data science field as a whole, and and just trying to see what techniques and things could be brought over to an insurance pricing.
00:07:18
Speaker
Going forward a little bit, I've gone part time my contracts and spending a lot of time on my masters doing all the shiny data science stuff. So image classification, text analysis, building reinforcement learning models.
00:07:33
Speaker
And again, spending a lot of time on the the personal projects as well. So almost identified more as a data scientist than a pricing analyst with the amount of time I'm spending on the former.
00:07:49
Speaker
And then ah think we were I can't remember when it was now, think it was about for three or four years ago, you were talking about generative algorithms before, before all the hype, before you could, before it was so common, yeah writing texts and building images.
00:08:07
Speaker
And I, ah just invested a kind of data science rig and went to install a Python library, which had a, which you you could do this with. And the requirements was 12, 12 gigabytes of video RAM.
00:08:21
Speaker
And my rig had 10 and I wasn't really prepared to go and replace my GPU with ah with a bigger one straight away. And even if I wanted to, ah ah couldn't because you can get GPUs back then. And I was very lucky to get mine.
00:08:36
Speaker
So that that was the introduction to cloud, just where if you needed something beefier, you could just go and rent rent that compute from from one of the providers.
00:08:48
Speaker
So that one was started off with AWS, yeah, built, built a lot of cool things with AWS. so think particular favorite service on those la Lambda functions, which are yeah little serverless instances that you can put a bit of Python code and it will spin up and run the Python code and then spin down and you charge by the millisecond.
00:09:11
Speaker
But you can actually do so much with those. If you want to break down big tasks. if you and run them parallel, you can go and spin up thousands of Lambda functions that will all execute on small parts of your task and finish it super quick.
00:09:24
Speaker
Ended up, you can build whole Python dashboard web apps in them as well. So host hosted a few of those. And again, because they're not running constantly, they're quite cheap and it's had some going for years and they've never really been charged because they haven't hit the thresholds.
00:09:42
Speaker
And so managing all all the kind of cloud infrastructure have gone down the route of, Infrastructure as code. So using things like Terraform and then as a Linux linux user already yeah using version control, but using Docker so you can take the environments on AWS, have a kind of a local version and build and test locally.
00:10:04
Speaker
And I think anyone listening to this, listening to the podcast at this point will think, what's this guy talking about? This has nothing to do with pricing. No, it totally is, though. And I remember using SageMaker and being like you, just, wow, this is really powerful. i can I can get all this stuff done without using my own computer with a lot of power.
00:10:29
Speaker
And yeah, and it just wasn't being used by anyone else. There is a yeah I think that's quite a steep learning curve when you first go on the cloud, when you first open up like AWS and it's like loads of names of all these things you can do. And I think that's quite daunt daunting for people. Yeah, it is. And then especially where AWS in particular, you can point and click on the browser, but it's not designed to be used that way. It is designed be used.
00:10:59
Speaker
to be interacted with a command line and interface through code. So that's the way they've intended. And it's a completely different skill set to to data analytics. It's people who are familiar running servers.
00:11:12
Speaker
They know the language. And it was harder back then as well because you didn't have ChatGPT. Nowadays, ChatGPT does solve a lot of the problems. You can't get it to set stuff up. But a big goal but big yeah thing that makes all of that easier was this is how I got started with Databricks because it's user interface over the top of the the cloud compute and it does make it a lot easier to manage all the infrastructure and you don't need to worry about getting the right things installed and having the right versions of stuff because yeah it's all in one on Databricks and there's everything else you need for kind of data analytics best practice so I started using that a lot just where it's
00:11:57
Speaker
just made everything much, much easier. And, you know, there's good data breaks, isn't it? It's very fast. I'd love to know how they get it to be so fast. So processing really big tables at lightning speeds, things that you would normally expect to be waiting quite a while for them to process just seem to be done immediately.
00:12:22
Speaker
that's, yeah, that's all to do with Spark. So it's the same The same people that built Databricks built, built Spark, which I'd actually started using Spark on AWS before. And this is how got into Databricks because it was Spark I was interested in just because you can run things, rings quickly.
00:12:39
Speaker
And then the the good thing about it is it it scales. You run a big job and so you, your compute will scale up and run everything in parallel. And then as soon as you finished, everything scales back down and then you're not being judged as much.
00:12:52
Speaker
Yeah. I think, yeah, I think my biggest cluster I've I provisioned on there was over a terabyte of RAM that I just needed the one time. Yeah. So, you you know, you can just ah go and have a terabyte of RAM for a bit and and then you don't need anymore. You don't need to pay for it anymore.
00:13:08
Speaker
Yeah.
Building Pricing Architecture and Future Visions
00:13:09
Speaker
Yeah. because i and That's where, that's where almost my two career tracks converge, which is where last year I got, got the opportunity to build out a pricing architecture within Databricks and through doing so,
00:13:24
Speaker
introduce kind of database practices, which but will save a a lot of time compared to how practices on the old architecture was set out. There's lots of tasks which will take months of time doing, which will just be zero time under the new architecture and then everything's version controlled and just make everything going simpler. But the main bottleneck is implementation with other software.
00:13:53
Speaker
But yeah, that that kind of brings brings me yeah where I am now, where I spent a lot of time doing something that almost felt completely different to pricing. And it does slowly feel like it's converged now. And the industry has started to move in that direction now as well. So definitely seeing it a lot more common that roles will be looking for kind of Python skills and and yeah machine learning. and And maybe one day that you'll see cloud on on job adverts, but I've not seen that yet where AWS. I've not seen it yet.
00:14:27
Speaker
Yeah. do Do you think pricing teams should be moving on to Databricks? I think so because ah think the one, the trade-off is is ease versus cost. And I think there's so much value that will come from the ease of and using a platform like that where and installing Python on on a work computer is is definitely a pain with all the permissions and then having rights and getting things working properly. And even if you get everything installed for some reason, it'll be slow because it'll going through a firewall.
00:15:00
Speaker
So if you're wanting to use kind of more modern techniques installing Python on everyone's machine and having the right versions of everything is is something that would be quite difficult to do. Whereas Databricks, you're all sharing the same, same compute. So straight away, it's all on the same versions and you don't really have those issues.
00:15:21
Speaker
So then it it depends on what the team builds out, but if they're using it to build out streamlined architectures and they use just using proper data best practices, they'll so see huge savings in time and big increase in efficiency. So the costs of the platform should outweigh those. And yeah think a good thing about the cloud is that it's, it's scalable. So if you build things sensibly and ah you don't,
00:15:50
Speaker
have big jobs running 24 seven and you do batch processing. a lot of stuff in pricing it doesn't really need real time streaming, streaming architecture. So you can build things super cheap.
00:16:03
Speaker
I've been using cloud and Databricks for several years now. And yeah other than the instances of a terabyte RAM and 3000 Lambda functions, it's generally been pretty cheap and usually I'm coming underneath the free tiers.
00:16:20
Speaker
yeah So yeah, and insurance data isn't that big compared to what other users of Databricks will be dealing with as well. So it's certainly possible to stay on the the cheaper end.
00:16:34
Speaker
know projects ah in Databricks that will use trillions of of data points in a dataset and when scoring models will be doing so on billions of requests daily. So That's the kind of scales that Databricks is designed to handle. And then we've got insurance data sets for it.
00:16:55
Speaker
You have quote data, which is a few million rows, but then ah other data sets, Richard, just a a couple of hundred thousand. If that's you can definitely build things to be cheaper on there, but yeah, it's just a case of of being mindful because it's certainly possible to to have unexpected bills on there if things are yeah i must admit that people are running sas servers with multiple licenses and an insurer that's an expensive option in itself and it's still fairly common yeah yeah and yeah i think if yeah there's definitely the the use case in terms of just making things more efficient but yeah the ultimately you have
00:17:40
Speaker
if things cost money and it's more just yeah working out what has a good return and yeah i think legacy legacy ways of doing things if they're expensive and also not that efficient then it's worth evaluating this.
00:17:57
Speaker
yeah And Rav, what would you say is your mission for general insurance pricing? yeah I don't think that the mission has changed a ah great deal. like Essentially it's to bring more modern approaches and and tools and and techniques to, to the industry that, you know quite common in, in and other industries that use data and analytics and and data science.
00:18:23
Speaker
But yeah it's changed a little bit in terms of originally it was about, it was pretty much just machine learning techniques, but machine learning isn't the, isn't the kind of main value add from data science because a lot of pricing teams can already build a ah pretty good predictive model.
00:18:40
Speaker
So yeah you might build a you might use machine learning to build a slightly better model, but it's not going to not going yeah change a huge amount. ah and But it has a lot of potential for all the other processes around that in terms of getting the data for for building those models and validating the models and deploying them and and monitoring them.
00:19:06
Speaker
yeahp There's tons of tools and approaches that can make all the rest of that fully automated if if you wanted and just just make everything much easier so that's where the value value i see is i think a lot of people focus on machine learning and yeah if they're not interested in in deploying it because they see it as riskier then the the rest gets neglected so and i think yeah that's the oversight so the mission is still to bring more modern approaches and techniques but less of a focus on the machine learning but still obviously have good understanding and
00:19:40
Speaker
still opt to use it myself because I prefer it. yeah Yeah, I could imagine extraterrestrial observer being like finding it rather strange that when there's all this that could be done in infrastructure and data and the higher level, the strategy and the capability of organisations, people seem to push the modelling side.
00:20:00
Speaker
If you took all the different skill sets in pricing teams, modelling is probably the the strongest when you compare to other industries. Yeah, it is. There's such a big focus on that area. And actually, i in my opinion, the returns on improving that can be quite diminishing with there's so much uncertainty. And often we're dealing with inadequate, just like you said, it may be millions of lines of data, but some claims are really rare and things like that. So we've got massive amounts of uncertainty and modelling is not really going to help us with that.
00:20:39
Speaker
Yes. So yeah, the, yeah yeah, in summary, so yeah, that's the mission really just to try and bring better, better ways of working with data to the industry mostly from data science fields, but also data engineering and and software engineering if it's applicable.
00:20:59
Speaker
And what would you say is your vision? Where do you think we should be ending up? So very, very similar to to the previous answer, but just But working with data best practices that are common and and in data science, data data science teams operate very differently to pricing teams.
00:21:21
Speaker
Yeah. But at the core, they're doing the same, same thing. They're taking data and using that to build models or analysis to either facilitate business decisions or that is the actual product if it's a data science product, but they will work in ways which are more akin to software engineering and yeah it's a lot more efficient and essentially working on a code base and building a system.
00:21:48
Speaker
Whereas think pricing teams is generally a lot of sporadic ad hoc analysis and processes and tasks. So I think the my vision would be that you adopt more of those principles and build a whole kind of price pricing architecture or system that serves all of these and you you work as a code based on improving the system rather than spending time on processes and ad hoc analysis.
00:22:15
Speaker
And I think the, what that enables then is ah freeing up time and focus for pricing specialists, rather than focusing on more mundane, repetitive tasks and yeah processes and refreshing reports and and so on. They can actually focus on pricing challenges and then that enables we'll focus on those. And so you have more innovation and people actually working on the harder stuff and and utilizing more of the their expertise.
00:22:46
Speaker
Yeah, I think it's a really interesting point because prices are actually software. that That's really what we're making. And that kind of isn't always realised. I think they where we've come from means that's often a very scientific, kind of academic, researchy backbone to pricing teams.
00:23:09
Speaker
And the reality today is that they're operational teams and they are deploying software into already complicated IT systems to generate people's prices.
00:23:22
Speaker
And that's what the teams are about. Most conversations aren't, oh, what should we analyze today? It's when are we deploying X, Y, Z? Here's the deployment plan. We need to line up with these sprints and all of this. It's very operational role now. Yeah.
00:23:38
Speaker
ah But then at the same time, a lot of the skill sets, which are key in software teams, just don't don't exist in in pricing teams. so No, they don't.
00:23:51
Speaker
And I think obviously the the main one is programming skills. So there's the reliance on software, which then doesn't, you know, a lot of the time doesn't have the the capabilities to do things in the best way as well. so Software, which doesn't interact with other software, yeah doesn't allow for automation, doesn't have version control, then suddenly you don't have three key points of building software, which is maintenance and implementation with other parts, version control and automations.
00:24:26
Speaker
Yeah, there's definitely there that soft software element of pricing, but then without a lot of the tools and approaches and and techniques that you'd see elsewhere. Yeah, absolutely.
00:24:38
Speaker
Absolutely. It's fascinating area. and I think it's something that's really changing. And I think people are waking up to the fact that there are sir better software options out there and they really need to be doing them. they shouldn't be having people work on very manual, old school, old style software solutions when there are such better options available.
00:25:02
Speaker
And as we talked about the cloud as well, it makes a big difference. Thank you very much, Ralph. It's been excellent talking this afternoon. And thanks so much for taking the time. Cool. Thank you very much for having me.
00:25:13
Speaker
Cool, Brent. You have a good rest of your day. Thanks. See you too. Bye. like