Idea
Probabilistic time-series forecasting tool delivering reliable, fast, and training-free prediction intervals for critical decision-making.
Research Paper
Core Innovation
This paper presents Conformal Seasonal Pools (CSP), a novel training-free probabilistic forecasting method that mixes same-season empirical draws with residuals around a seasonal naive forecast. CSP achieves superior calibration and accuracy compared to learned models like DeepNPTS, while running orders of magnitude faster and requiring no training or parameter tuning.
Why It Matters
Accurate and well-calibrated prediction intervals are essential in industries like healthcare, finance, and energy where poor forecasts can cause costly or dangerous outcomes. CSP's training-free approach drastically reduces computational costs and improves reliability, enabling scalable deployment in safety-critical environments and operational workflows.
Market Size (TAM)
$2–10B TAM for time-series forecasting software; $500M–$1B SAM from safety-critical industries. Driven by demand for reliable, interpretable forecasts and cost-efficient deployment.
Potential Customers & Pain Points
- Healthcare providers – Need reliable patient outcome forecasts
- Financial institutions – Require accurate risk and capital forecasts
- Energy operators – Demand precise grid and load predictions
- Autonomous systems developers – Need trustworthy multi-step trajectory forecasts.
Business Model
SaaS platform offering API access to CSP forecasting with tiered pricing based on usage volume and enterprise features including compliance and integration support.
Competitive Landscape
- DeepNPTS
- Prophet
- DeepAR
- N-BEATS
Implementation Challenges
- Adoption resistance due to preference for learned models
- Integration challenges with existing forecasting pipelines
- Convincing safety-critical sectors to trust training-free methods
Validation Strategy
- Benchmark CSP against leading learned forecasters on diverse real-world datasets
- Pilot deployments in healthcare and energy sectors to demonstrate safety and cost benefits
- Collect user feedback and iterate on API usability and integration capabilities
Research Paper Overview
Training-Free Probabilistic Time-Series Forecasting with Conformal Seasonal Pools
Summary
This paper introduces Conformal Seasonal Pools (CSP), a training-free probabilistic forecaster that outperforms DeepNPTS on multiple benchmarks with significantly better calibration and over 500x faster CPU runtime. CSP provides reliable prediction intervals critical for safety- and decision-critical applications without requiring any model training or learned parameters.