Idea
COSMIC is a zero-shot time series forecasting model that integrates covariates for improved accuracy, benefiting data scientists and enterprises.
Research Paper
Core Innovation
This paper introduces COSMIC, a zero-shot forecasting model that uniquely incorporates covariates via in-context learning. It uses Informative Covariate Augmentation to train without requiring datasets with covariates. This approach enables state-of-the-art forecasting performance both with and without covariates, overcoming limitations of prior models.
Market Size (TAM)
$2–10B TAM, $1–2B SAM; assumption: growing demand for automated forecasting in enterprises and SaaS platforms.
Potential Customers & Pain Points
- Enterprises needing accurate forecasts without extensive retraining
- Data scientists lacking models that handle covariates effectively
- SaaS providers seeking scalable forecasting APIs
Business Model
Subscription-based API access for forecasting services with tiered pricing based on usage and features.
Competitive Landscape
- DeepAR
- N-BEATS
- Prophet
Implementation Challenges
- Integration complexity with existing forecasting pipelines
- Data privacy concerns with covariate usage
- Adoption resistance due to model interpretability
Validation Strategy
- Develop prototype API and test on public forecasting benchmarks
- Pilot with select enterprise customers for real-world validation
- Iterate model based on feedback and expand covariate support
Research Paper Overview
Zero-Shot Time Series Forecasting with Covariates via In-Context Learning
Summary
Pretrained time series models capable of zero-shot forecasting have shown promise in improving forecasting performance and accessibility. Existing models either lack support for covariates or do not incorporate them effectively. COSMIC is introduced as a zero-shot forecasting model that leverages covariates through in-context learning and uses Informative Covariate Augmentation to train without datasets containing covariates. COSMIC achieves state-of-the-art zero-shot forecasting performance with and without covariates, effectively utilizing covariates in forecasting.