Why does real-world driving data matter in predicting crashes?
Predicting something that is rare isn’t easy. In driving, most trips end safely, so an AI trained to predict crashes needs to learn from a very large number of real examples.
That’s why we’ve been training our AI for more than 20 years using real-world driving data from 106 countries – and it’s one of the things that was highlighted in our AI validation. Independent validator Anders Arpteg said: “Because crash events are statistically rare, effectively training an AI to predict them requires a massive volume of historical and global data. Given the model performance, it is clear that sufficient data from many years and countries has been used.”
Here, I will explain how training an AI with real driving data, over many years and in many countries, translates into the ability to predict crashes.
Why does an AI model need so much data to predict crashes?
Statistically, crashes are rare, so the AI model needs enough real examples to learn from, including trips that didn’t end in a crash. GPS data doesn’t show risk on its own. But AI can detect subtle patterns across entire trips, including behaviors that wouldn’t typically be identified as “risky”. Being able to recognize which behaviors contribute to crash risk and which don’t requires a large number of real examples.
How does the AI use real-world data to predict crashes?
I always like to use visual examples to explain our AI model. Think of building with toy blocks. Imagine if you knew what billions of people had built using the same pieces. As someone new started to build, you’d soon be able to predict what they were building, based on previous examples. The more examples you have, the more reliable the prediction. Read more about this.
Our AI works in a similar way. It uses a form of pattern recognition to spot patterns in driver behavior and compare these with real-world profiles with known outcomes. Read more about patterns in driver behavior.
Why does it matter that the data is from the real world?
It means the AI learns from things that have really happened, including trips with and without crashes. It also knows whether claims followed – including at-fault and not-at-fault. This means that when the AI identifies a pattern that is linked to higher crash probability, it’s based on evidence from real trips, not assumptions about which driving behaviors might be risky. This is also why our AI doesn’t base risk predictions on proxy indicators such as harsh braking.
Why is it important for the data to come from different countries?
If the AI was only trained using data from one country, it wouldn’t be as effective at measuring risk across different geographies. That’s because driving doesn’t look the same on different roads, in different traffic conditions and different cultures. Training the AI on data from 106 countries means that it learns how behaviors correspond with crashes, wherever the driving takes place. This is reflected in the validation, which found consistent results across continents, vehicle types and driver categories.
Why does it matter that the AI training covers more than 20 years?
Over 20+ years, our AI has learned from more than 7 billion real-world driving profiles, each linked to a known outcome. And because outcomes such as crashes and claims only become known afterwards, that evidence can only be built up over time. That’s why a database like ours isn’t something that can be built overnight. Watch a video of how our AI works.
What did the independent validation of Greater Than’s AI show?
The validation assessed the methodology, data handling, and performance reporting behind the AI model. The main conclusions were that:
- The AI successfully isolates a 25x differential in crash probability between the drivers identified as highest risk and lowest risk.
- The drivers identified as highest risk by our AI are 91x more likely to be involved in a crash associated with a severe or major claim vs those identified as lowest risk.
Read more about the AI validation
FAQs
How is the AI trained to predict future crashes?
It compares a driver’s behavioral patterns across trips with a database of billions of real-world driving profiles with known outcomes.
Why does real-world data matter for predicting crashes?
It means the AI learns from what actually happened, so its predictions are based on real crash and claim outcomes, not assumptions about which driving behaviors might be risky.
How many countries does the training data come from?
Greater Than’s AI has been trained with real-world driving data from 106 countries.
