There are four major factors shaping today’s property and casualty (P&C) insurance industry: 1) increased risk from climate change, 2) increased competition from newer entrants, 3) time-intensive and costly manual processes, and 4) an explosion of data and imagery so vast that most don’t know how to take advantage of it. The fusion of these pressures has resulted in the insurtech boom, and new startups and legacy companies alike have seized the opportunity to bring a fresh perspective to fill these gaps. Many of them have pulled inspiration by applying tools like geospatial imaging or techniques like machine learning (ML) to triangulate and mitigate risk. At Arturo, we apply the power of machine learning to imagery, and this produces upwards of 70 property characteristics in mere seconds. These get ingested into an insurance carrier’s existing workflow, enabling them to make smarter decisions across the policy lifecycle—for cheaper. But how does this newfangled methodology really work? And how do we reliably know it can reduce risk?

‍Multiple sources of imagery

‍Just as a great meal comes together with great ingredients, truly impactful intelligence begins with harvesting the right sources of imagery. At Arturo, we partner with major imagery providers, across aerial, satellite, drone, stratospheric balloon, ground-level, and beyond. We’re always on the lookout for new, cutting edge providers to incorporate into our machine learning models.  Utilizing just one provider means there are often some tradeoffs to get something close to what makes most sense for your business. By contrast, pulling from multiple providers means we can maximize the three qualities that matter most: resolution, frequency, and coverage.

• Resolution: In simple terms, greater resolution means the image is more clear, enabling any model (or person, for that matter) to derive meaning from the imagery. For instance, it is impossible to build accurate roof condition models, taking into account granular details like chimneys, solar panels, skylights, and roof material with a low resolution model.
• Frequency: While resolution refers to the quality of the imagery, frequency or cadence describes how often the same location is imaged. This helps ensure currency, as a provider who takes a snapshot of a house every few years isn’t enabling as much actionable knowledge—and is perhaps missing major changes—as someone who revisits it every few months. To put this into context, it is best practice for insurance carriers to run an entire portfolio analysis four times a year, or once per quarter.
• Coverage: This refers to the geographic coverage of an imagery provider. Oftentimes, imagery providers will recapture the same areas multiple times, so they can grow their library of historical imagery, but fewer venture into new locations. It isn’t so easy to find a single provider whose imagery covers the entirety of a carrier’s portfolio.

All of this is to say the more imagery, the better an insurance carrier can understand the condition of a property, especially as it changes over time or even immediately after a natural disaster. That’s why at Arturo we take advantage of multiple sources of imagery, giving us the ability to attain the most accurate and the most comprehensive property analytics.In short, the ability to see homes without spending the time or money on a physical property inspection can be an incredible value-add across the policy lifecycle. Still, at the end of the day, the imagery is only as useful as the intelligence that can be derived from it, and that’s where machine learning comes in.

‍Machine learning explained

‍At its core, machine learning is the practice of training a computer to recognize patterns that we, as humans, can intuitively recognize. Pattern recognition is something people are exceedingly good at—so good that when you visit a website and it asks you to verify you’re not a robot, tasking you with picking out photos of cars from a series of grainy, choppy photos, most anyone can do it with ease. For a computer, that’s really, really hard. And it takes a broad team of expert engineers and scientists to do it well. When a machine learning engineer trains the computer to such an extent that it grows to handle tasks on its own, like IBM’s Watson, that creation is called artificial intelligence. There are many ways to train a computer to identify objects and features, and they are typically grouped into supervised and unsupervised learning algorithms. The difference between the two is the availability of ground truth, or labeled data. To the extent the labeled data has sufficient variety and variability, the deep learning algorithm can robustly extract discriminative information from it and use them to recognize the object or feature it was trained to find. This means a team goes through and labels items on an image, instructing the computer on what a roof is, what a tree is, what a driveway is, and how all of these things can look across millions of different images. Doing this thoroughly and across a wide array of imagery with rigorous oversight is the central component in building a reliable model because without it, sometimes the model will misidentify major things, like confusing asphalt shingle with concrete tile. These errors can make a drastic difference in how a carrier will underwrite a property. Which brings us to the other core item: adjusting the model when it’s wrong and giving it new information to consider when labeled test and validation data disagrees with the model output. At Arturo, we work hand-in-hand with our customers through this iterative model validation process, thus ensuring the experiences and depth of knowledge from an underwriting or claims department is being taken into account. The AI team at Arturo, solely focused on property, is growing rapidly, and they bring to the table the final, critical piece of the puzzle: cognitive diversity. A machine learning model is ultimately only as good and unbiased as the people who thoughtfully design it, and a large and diverse team ensures a wealth of experiences and perspectives to make the Arturo models the best they can be. Our cognitive diversity is a key ingredient in the culture of innovation and technical resilience we foster at Arturo.When an on-site inspector goes to a home, spending time and gas on the drive and the walkaround, they are able to recognize the trees around a home, including how long the branches are and how close to the house they are. With Arturo, our AI can assess a full property in seconds and at a fraction of the cost.

‍Growing confidence

‍So that brings us to the final question: how do we know the model is good? Although anyone working with artificial intelligence strives every day to make a more accurate model, the reality is no model is perfect. How do you know what information to trust and to what extent? At Arturo, we give you that assessment of trust up-front. As part of our classification models’ outputs,  we produce a probabilistic confidence score which indicates how certain a given model is of its particular outcome. For example, if the roof material classification model determines a home has a tile roof, it will also output a percent-score indicating how likely it is that the roof material is actually tile. We employ a similar trust up-front mindset in the case of our characterization models, such as identifying areas associated with various roof conditions (e.g., rust, debris, streaking, etc.) and provide statistical confidence scores of the given model.This helps carriers compare and validate the models against ground truth within their own portfolio, and it allows them to decide for themselves what range of confidence scores they trust in their automated decision workflows and when they think further human-in-the-loop investigation of a property is required. This way, they can allocate resources more effectively, only sending out an inspector when absolutely necessary. For many customers, this creates a far more seamless experience, especially around point-of-quote where abandonment rates tend to be high because of friction and lengthy lead times.

‍Decide with intelligence

‍The biggest risk for the P&C insurance industry today is lacking intelligence to make the right decisions.On-site assessments can only generate very high-level information about a property, one at a time. If insurance carriers don’t have a way to verify the most recently recorded exterior condition of all properties in their portfolios, critical information is being lost. Having access to this information, both current and historical, on-demand is an added convenience that Arturo’s models provide.Aerial images taken from planes, stratospheric balloons, satellites and drones provide helpful vantage points from which a wide range of property attributes and environmental risk factors can be determined. With a strong underpinning of ground truth baked in our model development, we ensure that accuracy is built in our AI models from the start. We label hundreds of thousands of properties for our initial work, and we also leverage millions of properties from our customers to further refine our work. When these images are processed using Arturo’s AI models, they can uncover dozens of insights about a property in seconds. From increasing quote completion and conversion by 13% to improving accuracy and reducing premium leakage by 20%, there are significant quantifiable upsides to taking advantage of the power of machine learning to drive top line growth, improve margins, and mitigate portfolio risk across the policy continuum.