6. Learning ML
I want to learn ML - what do you suggest?
Overview
A lot of websites (us included!) will say things like it’s never been easier to learn ML and that’s true in the sense that there is an immense treasure trove of resources on the internet and in the open source coding community meaning that you can essentially learn whatever you want. However that brings challenges all this material can be overwhelming so it may be difficult to know where to start.
The learning path
For many people, successfully starting your learning journey in ML and AI becomes a mindset thing.
If you are looking for reasons to avoid learning about ML and AI then there are plenty… but if you are committed to learning then it’s easier now that it ever has been. We consider that you can become competent in a narrow field in 3-6 months if you set aside 7 hours a week to learn, so a similar time commitment to an actuarial exam say.
Tailor your learning journey to your current skillset, learning style and interests. There’s a reasonable natural evolution of analytic techniques: GLM’s -> elastic Nets -> GBMs -> Neural Nets
So if you’ve long forgotten about GLMs start with them and set your 3 month target to be competence in elastic nets and 6 months to be GBMs. Recognise too that you will need to program a computer either in Python or R to apply ML and AI, so you may need to spend some time building confidence in one of these first.
Don’t jump into Neural Nets and complex natural language models, that’s a less developed domain and far more complex. GBMs are good first target as they remain the most powerful and widely used regression technique.
Then understand your preferred learning style, books, research papers, guided online courses all exist often at low to no cost, pick what works for you. And lastly align your journey to a target you have a genuine interest.
One useful resource is Kaggle (or other similar platforms), particularly if you’re the type of person who enjoys a bit of friendly competition (though you can still learn a lot without entering too so don’t let a dislike of competitions make you avoid these websites), with the added benefit of learning from what others do. Participants often share their code and approach during and at the end of competitions. From time to time there are insurance competitions where you can hone your skills, but equally, you can learn and apply ML techniques in other areas too.
Practical tips
We also offer some practical tips in an article published in the Actuary in March 2022. One of our key suggestions there from Greg Taylor is to gain familiarity with new processes through baby steps so don’t try to take everything on at once. So if you want to use ML in practice, start with some data that you are already familiar with so that you will know when models are performing well. Start with the simpler techniques first before jumping into the more complex ones. Learn incrementally rather than going for the moonshot and risking becoming disheartened.
Don’t get bogged down by language wars (R vs Python vs newcomers like Julia) - just pick one, whichever one is more likely to be supported by your company, and start learning to code. There are a wealth of free and paid-for options out there that can get you started. You don’t have to start with ML courses either, if you don’t want to – a lot of the time, it’s better to start with what you know and build from there.
For example:
- You could try taking the SQL or SAS code behind a data process that you already know very well and try to replicate the output in your language of choice
- You could try building a chain ladder model from scratch
- You could check out the chainladder package in R or Python and try to get that working.
The most important thing is to keep what you’re doing interesting to you and, if possible, directly applicable to your work. Also, don’t be afraid to ask your manager to put in some time for you to develop these skills. Personal development time can definitely help with finding the time to do this. Reserving actuaries, unlike Pricing actuaries, often have the luxury of knowing when you are likely to be busy months in advance, so you can take advantage of this and build in time during “quiet months” to build your skills.
When it comes to learning these skills, remember that there is no rush. Go at your own pace, and do what works for you. You don’t need to be familiar with every single technique and model, build up your skills incrementally.
Don’t be put off by the jargon
One thing to remember is that there’s a lot of jargon used in ML which can make it seem more mysterious than it is. Terms like one-hot encoding, feature engineering or bagging may seem like exotic things, but there’s a good chance you already know the basics of what these are, particularly if you have done some statistical modelling in the past - if you know about creating dummy variables for a categorical term, defining a candidate set of independent variables or bootstrapping techniques then you’re well on the way to understanding those particular terms. If you haven’t done a lot of statistics, then this is more jargon, but it’s still likely that once you drill down to what’s going on, you’ll be familiar with a lot of the concepts.
Further reading
- A suggested path into machine learning
- Our Foundations workstream page which has a curated set of material for getting started.
- Foundations blog posts: a number of articles and tutorials on topics such as getting started in R, GLMs in R and Python and some machine learning tutorials, also in R and Python.