Large Scale Causal Foundation Models [Whitepaper]

Most forecasting models are pattern matchers. They tell you that two things moved together. They can't tell you which one caused the other, or what happens when you actually intervene. That gap is why five models inside one enterprise produce five different answers to the same growth question.

Our new research paper sets out how we closed that gap.

Large-Scale Causal Foundation Models introduces FOUNT, the architecture behind POEM365 and the world's first Large Causal Model. It combines causal discovery with a transformer architecture, so it learns cause and effect across hundreds of interconnected targets at once, not just correlations inside a single one.

Download the whitepaper

What the paper covers

• How FOUNT uses transformer attention to discover causal structure, grounded in Judea Pearl's causal framework, then reasons about interventions and counterfactual "what if we act" questions
• How one unified model forecasts many connected targets simultaneously, capturing the synergies and halo effects that siloed models cannot see
• How POEM365, pretrained on more than 15,000 real brand datasets, applies this to enterprise sales, supply chain, and financial planning

The results

We benchmarked FOUNT against the leading foundation models from Google, Salesforce, and Intel:

• 36% average accuracy improvement across standard benchmarks (ETT, Electricity, Weather), and 75% on ETTh1
• On the M5 forecasting competition, a WMRSSE of 0.516, beating the Kaggle competition winner's 0.520
• The winner used an ensemble of more than 200 models. FOUNT did it with one.

That last point is the one worth sitting with. A single causal model, trained once, outperforming a 200-model ensemble purpose-built to win one competition. That is what a foundation model for causation makes possible.

Who should read it

If you lead data science, analytics, or growth at an enterprise, this is the technical foundation under everything we say about Enterprise Decision AI. It is written for people who want to see the architecture, the benchmarks, and the methodology, not just the claims.

Download the whitepaper