Idea
A graph neural network platform that improves multivariate time series forecasting accuracy for enterprises and researchers.
Research Paper
Core Innovation
This paper introduces MillGNN, which uniquely integrates cross-correlation and dynamic decaying features to capture multi-scale lead-lag dependencies in time series data. It employs hierarchical message passing to model both intra- and inter-scale lead-lag effects, enhancing forecasting precision beyond existing methods.
Market Size (TAM)
$2–10B TAM, $1–2B SAM; assumption: growing demand for advanced forecasting in finance, energy, and supply chain sectors.
Potential Customers & Pain Points
- Financial institutions needing accurate market predictions
- Energy companies optimizing demand forecasting
- Supply chain managers reducing inventory risks
- AI researchers seeking advanced time series models
Business Model
Subscription-based API access for forecasting services with tiered pricing based on data volume and features.
Competitive Landscape
- Temporal Fusion Transformer
- N-BEATS
- Graph WaveNet
Implementation Challenges
- Complexity of model integration into existing pipelines
- Data quality and preprocessing challenges
- Scalability for very large datasets
Validation Strategy
- Pilot with financial and energy sector partners to benchmark accuracy
- Develop open-source toolkit for academic validation
- Conduct scalability tests on large multivariate datasets
Research Paper Overview
MillGNN: Learning Multi-Scale Lead-Lag Dependencies for Multi-Variate Time Series Forecasting
Summary
MillGNN is a graph neural network method that captures multi-scale lead-lag dependencies in multivariate time series forecasting by integrating cross-correlation and dynamic decaying features; it uses hierarchical message passing to model intra- and inter-scale lead-lag effects, improving forecasting accuracy on 11 datasets compared to 16 state-of-the-art methods.