The clearest empirical evidence for this approach comes from financial services — where the performance gains are documented and the data to replicate them exists at most institutions. Logistics and online advertising follow the same logic: decisions made repeatedly, on behavioural data that accumulates over time, using models that treat each problem in isolation.
Most financial institutions build fraud, credit, and AML models on hand-engineered features: transaction amounts, velocity counts, merchant categories, time-of-day flags. The problem isn't just the maintenance overhead of running separate models for each task. It's that feature engineering is bounded by what you thought to look for.
A model trained on the raw sequence of transactions learns the structure of how behaviour unfolds over time — including patterns that emerge from combinations of signals no analyst would have thought to define. Card testing is a clear example: the fraud signal isn't in any single transaction attribute, it's in a sequence of small probing transactions that precede a larger attempt. That pattern is invisible to feature-based models but recoverable from the full history.
Card testing is a particularly favourable case, but the principle extends across fraud, credit, and AML: the signal is in the sequence, and institutions sitting on years of transactional history have the raw material to use it.
A freight forwarder's operational history captures something generic benchmarks cannot: which carrier-lane combinations degrade under pressure, which customs routes carry hidden delay risk, which customers generate systemic exceptions. That proprietary record is the training data.
A model trained on that history learns to predict delivery outcomes, score carriers, and flag cost variance with accuracy that improves as the training set grows. The same approach that works in payments — one model across multiple business problems, trained on the full operational record — applies directly here. No competitor can replicate it, because they don't have the data.
Most brands run bid strategy, audience targeting, and budget allocation as separate optimisation problems — each in a different tool, each optimising a different metric, none of them aware of what the others know. Programmatic campaigns generate a structured record of what worked, for which audience, at what price, in what context. That auction history is the training data.
A model trained on the full history learns the interactions between bids, audiences, and creatives simultaneously, then transfers that across forecasting and incrementality measurement. For advertisers spending at scale, prediction accuracy compounds as the data grows — in a way that rules-based bidding cannot replicate.
If you're sitting on a transaction history, an event log, or a table that your current models aren't doing justice to, we'd like to hear about it.