Name: Lahiru Somaratne
Years in the field: 6 Years in Insurance
Current Company: Ameritas Life Insurance Corporation
1. Could you please briefly introduce yourself?
I’m Lahiru Somaratne, a Lead AI Specialist at Ameritas with a background in actuarial science (B.S. and M.S. from the University of Nebraska–Lincoln). My passion lies at the intersection of actuarial science and artificial intelligence—exploring how AI can uncover patterns, improve decision-making, and transform the way insurers understand risk at both the macro and micro level.
2. What first inspired you to explore deep learning methods, like LSTMs, for actuarial loss reserving?
When I first started exploring traditional machine learning, I was excited by its ability to handle large amounts of data. However, I quickly noticed how much the results still depended on feature engineering—choosing the right inputs and transformations was often as important as the model itself. What ultimately drew me to deep learning, and LSTMs in particular, was how these models could automatically learn increasingly abstract patterns from the data, layer by layer. Instead of relying so heavily on manually designed features, the model could uncover subtle, nuanced relationships that might otherwise be missed. For actuarial loss reserving, where patterns evolve over time and can be quite complex, that capability felt like a game-changer.
3. For readers who aren’t technical, how would you explain what an LSTM model is and why it’s useful for predicting insurance losses?
An LSTM (Long Short-Term Memory) model is a type of AI designed to learn from patterns that unfold over time—kind of like how we remember events in a story. What makes it special is that it can keep track of both short-term details and long-term themes at the same time. Imagine trying to predict the ending of a mystery novel: you need to remember the clue from the last page you just read, but also something important that happened ten chapters earlier. LSTMs work the same way with data. In insurance, this is especially useful because losses develop over time. Some patterns show up right away, while others only emerge years later. By remembering both kinds, LSTMs can make better predictions about how claims will ultimately play out.
4. Your project compared deep learning with the traditional Chain Ladder method. What stood out most in that comparison?
The Chain Ladder method is elegant and practical, and actuaries often strengthen it by carefully selecting LDFs—including from external benchmarks when appropriate. This process draws heavily on professional judgment, which is one of the method’s real strengths. Deep learning, by contrast, doesn’t rely on manual selection. Instead, it learns directly from the data, automatically identifying which patterns to emphasize and how they generalize across companies and triangles.
Of course, this cross-learning only works when the triangles come from a similar environment—without that consistency, the patterns may not transfer well. In that way, deep learning complements actuarial approaches by offering a systematic, data-driven way to uncover development patterns that might be difficult to select by hand.
5. What were the biggest challenges you faced while working with this kind of data, and how did you overcome them?
Perhaps the biggest challenge was simply the structure of the data. In most fields, data comes in rectangular form—think rows and columns in a spreadsheet. But actuarial reserving data is triangular, because newer accident years don’t have as many development periods as older ones. That irregular shape makes it harder to feed directly into machine learning models. Fortunately, this isn’t a new problem in AI. In Natural Language Processing (NLP), for example, sentences come in all different lengths:
- “Insurance is complex.” → 3 words
- “Insurance losses develop over many years.” → 6 words
Since models expect uniform input sizes, NLP practitioners use padding—they add blank tokens so both sentences become the same length:
- “Insurance is complex. [PAD] [PAD] [PAD]”
- “Insurance losses develop over many years.”
We used a similar idea for loss triangles: padding newer accident years with “blank” development periods so the data formed a neat rectangle. This preserved the triangular structure while making it compatible with deep learning architectures.
6. How often do you present your research at actuarial conferences, and what do you enjoy most about sharing your work with that audience?
I’ve had the privilege of presenting my research at actuarial conferences every couple of years. One aspect I deeply respect about the profession is its long history and careful approach—new ideas are expected to be tested and proven before they gain acceptance. At the same time, technology is advancing rapidly, and it’s exciting to see actuaries’ growing curiosity about methods like machine learning and AI.
What I enjoy most is the reception from the audience: even when the ideas are novel, there’s genuine interest and thoughtful discussion. This balance between tradition and openness makes sharing my work rewarding, and I believe it’s a dynamic that will help the profession continue to evolve in the long run.
7. Where do you see the greatest opportunities for artificial intelligence in actuarial science going forward?
I believe some of the most exciting opportunities in actuarial science come from the rise of multimodal data and multimodal AI models. Traditionally, actuaries worked with structured, regularly reported numbers—claim counts, paid losses, exposures. But today, risks can be described by many different kinds of data: claim notes written in text, medical images, telematics from cars, or climate measurements, often recorded at different intervals.
The exciting development is that multimodal large language models (LLMs) are starting to handle exactly this kind of variety. They can combine text, numbers, images, and even sensor data to uncover relationships that would be difficult to capture with a single data source.
If actuarial science can leverage these advances, we’ll be able to build models that don’t just predict losses more accurately, but also help us understand the drivers of risk in a richer, more holistic way. This opens up opportunities across pricing, reserving, underwriting, and risk management.
8. Finally, what advice would you give to students or young professionals who want to combine actuarial science with machine learning?
Actuarial science and machine learning may seem like very different worlds, but in many ways they’re cousins. Both involve analyzing data, identifying patterns, and making predictions. The main surface-level difference is that machine learning relies more heavily on coding.
Learning to code isn’t just about writing programs—it’s a tool that lets you experiment with problems you already understand. Equally important is learning to measure trade-offs when framing an actuarial problem as a machine learning task. For example, the assumptions we make in actuarial models—about how losses develop over time, or the degree of variability we expect between policies—can guide which ML models to choose and how to configure them. Should the model focus on sequential patterns, like an LSTM for claims over time? Should it emphasize robustness to outliers? Understanding these trade-offs helps you select the right approach and ensures that your machine learning solution is grounded in the realities of actuarial reasoning.
When approached with care and understanding, actuarial science and machine learning complement each other beautifully. Using ML can give you a deeper intuition for what the data is telling you about risk, while your actuarial expertise helps you guide the model effectively. Together, they create a virtuous cycle: each field strengthens your understanding of the other, allowing you to build smarter, more insightful solutions.