Multiple input variables
Pricing, promotions, weather, macro indicators
Platform overview
FOUNT represents a fundamental shift in how forecasting systems are built and how decisions are made from data.
Most forecasting models today - from classical statistical approaches to modern transformer-based foundation models - are built to recognize patterns. They learn from historical data, detect correlations, and extend those patterns into the future. This works well when the world behaves exactly as it has before.
But real-world systems don't behave that way.
Markets shift. Consumer behavior evolves. External shocks disrupt trends. And when that happens, pattern-based systems fail - not because they lack data, but because they lack understanding.
FOUNT was designed to solve this exact problem. Instead of learning only what tends to happen, FOUNT learns why it happens. It integrates causal discovery directly into a transformer architecture, allowing it to understand how variables influence each other over time. This shift - from correlation to causation - is what makes its forecasts more stable, more interpretable, and far more useful for decision-making.
The limitation of traditional forecasting is not computational - it is conceptual.
Even the most advanced models like TimesFM, Chronos, or Moirai operate as highly sophisticated pattern recognizers. They can identify that two signals move together, but they cannot distinguish whether one actually drives the other.
This leads to fragile systems. A model might learn that sales increase during a certain period, but it cannot tell whether that increase was driven by pricing, promotions, seasonality, or external demand. As long as those hidden drivers remain stable, the forecast holds. The moment they change, accuracy collapses.
FOUNT addresses this by explicitly modeling cause-and-effect relationships. It separates genuine drivers from coincidental signals, ensuring that predictions remain reliable even when underlying conditions shift.
FOUNT does not treat forecasting as a single-output prediction problem. It treats it as a dynamic, interconnected system.
It learns how multiple variables interact, how multiple outcomes influence each other, and how changes propagate through the system over time. This allows it to do something most models cannot: move from prediction to reasoning.
At its core, FOUNT operates as a unified model that simultaneously handles:
Multiple input variables
Pricing, promotions, weather, macro indicators
Multiple output KPIs
Revenue, demand, churn, margins, LTV
Multiple domains
Cross-industry, cross-geography datasets
This unified design is critical. In real businesses, nothing operates in isolation. Sales impacts inventory. Promotions affect margins. Marketing influences both short-term demand and long-term customer behavior. These relationships are not secondary - they define the system itself.
FOUNT learns these relationships directly, rather than approximating them through separate models stitched together after the fact.
The impact of this approach is not just theoretical - it shows up clearly in performance.
#6 competition
0.516 WMRSSE
Outperforming the Kaggle-winning solution that relied on an ensemble of 200+ models
Benchmark datasets
Significant gains
Across ETTh, Electricity, and Weather - standard time series benchmarks
This is not just a marginal improvement. It is evidence that a single causal model can outperform highly engineered ensembles when it understands the underlying system.
FOUNT is not positioned as an isolated research breakthrough. It serves as the foundation for POEM365 - a large-scale enterprise forecasting system trained on over 15,000 real-world brand datasets.
Demand and supply planning
Revenue and margin forecasting
Marketing and investment optimization
Long-term strategic planning
FOUNT changes the role of forecasting inside an organization.
| From | To |
|---|---|
| Reporting | Reasoning |
| Prediction | Decision support |
| Isolated models | Unified intelligence |
| Guesswork | Understanding |
Free tier available. No credit card required. Multi-KPI and single-KPI forecasting - production-ready in minutes.