🚨 We trained a large language model on 2B+ card transactions and it caught 35% more fraud. Like the labs train models on words, we can train models on transactions. To do that, the model has to read a transaction like it reads a word. So each transaction gets broken into its parts: - Who the merchant is - What it sells - How much moved - Which channel it came through - How long since this customer last transacted These models are pattern matching machines. Claude “predicts” the next word in a sentence whereas our trained model predicts the next transaction a customer might make. Repeat this across billions of transactions, and the model builds a detailed sense of what normal looks like for each individual customer. We never told the model which transactions were fraud. We don’t want it chasing fraud directly. What we want is for it to learn what ordinary spending looks like, so anything off-pattern becomes instantly recognizable. For any given transaction, the model produces 2 things: 1. A compact numerical summary of everything the customer has done up to that point, which in machine learning is called an embedding. 2. A "surprise" score. Basically, a measure of how unexpected this particular transaction looks given the customer's history. We trained a foundation model that discovers the language of transactions. Every person has their own dialect that they speak, and the embedding we built discovers the family of dialects that describe payments. The holy grail here is to take the foundation model and apply it to a brand new client (whose data was not in training set). That's what we did and had amazing results: Across every threshold we tested, adding the embeddings caught 24% to 35% more fraud. At a common operating point, where the riskiest 1% of transactions get flagged for action, the improvement was 28%. Our AI team wrote an entire blog post about it that digs into specifics, drawbacks, and our methodology: sardine.ai/blog/foundation-m… We are also publishing a whitepaper titled ‘Building a foundation risk model’ that I’ll talk about more here in the coming weeks. Follow for more on this process. One important note to end on - we didn't replace our scoring model with a neural network. Our issuing model already scores every transaction using gradient-boosted decision trees, built with XGBoost, and running on 77 features we've engineered over the years. The new embeddings were simply added to it as extra features, and then we retrained the same XGBoost model on the larger set. Again, more specifics in the blog than I can fit here. I’m incredibly encouraged by this progress, both for @sardine and the entire fraud & fincrime fighting community.