Phase 2

Overview

Phase 2 is INIT’s internal R&D initiative focused on building an advanced demand-forecasting platform for highly seasonal products. The project set out to push the boundaries of time series prediction by combining classical statistical methods with modern AI/ML techniques — culminating in the development of GLPatch, a novel forecasting model that achieved state-of-the-art benchmark results.

The Challenge

Forecasting demand for highly seasonal products is one of the hardest problems in applied time series analysis. Traditional statistical models struggle with the non-linear patterns, multiple overlapping seasonalities, and abrupt trend shifts that characterize real-world seasonal demand data. Off-the-shelf ML models, while powerful, often fail to capture the underlying signal structure without careful decomposition.

Key difficulties included:

• Multiple overlapping seasonal cycles (weekly, monthly, annual) with varying amplitudes
• Abrupt regime changes and trend shifts in demand patterns
• Noisy real-world data requiring robust signal extraction
• The need for a model architecture that balances statistical interpretability with deep learning flexibility

Our Solution

1. Signal Decomposition Study

We conducted an extensive study of signal decomposition techniques, evaluating IIR (Infinite Impulse Response) and FIR (Finite Impulse Response) filtering approaches for separating seasonal components from trend and residual signals. This allowed us to isolate the core seasonal patterns before feeding cleaner signals into the forecasting model — a critical preprocessing step that significantly improved prediction accuracy.

2. GLPatch — A Novel Forecasting Architecture

Building on insights from the decomposition study, we developed GLPatch, a custom forecasting model that combines patch-based temporal embeddings with learned seasonal representations. The architecture processes time series data in structured patches, enabling the model to capture both local dynamics and long-range seasonal dependencies in a computationally efficient manner.

3. Comprehensive Benchmarking

We rigorously benchmarked GLPatch against established forecasting methods across multiple datasets, evaluating performance using both MSE (Mean Squared Error) and MAE (Mean Absolute Error) metrics. The model was tested across a wide range of forecasting horizons and seasonal patterns to validate its generalization capabilities.

4. Intellectual Property & Technology Maturation

The results were strong enough to pursue formal IP protection, and the technology was matured to TRL 8 (Technology Readiness Level 8) — system complete and qualified, ready for deployment in production environments.

Results

Phase 2 delivered exceptional benchmarking performance and produced deployable forecasting technology:

• 18 first-place MSE results across benchmark datasets, demonstrating superior prediction accuracy
• 22 first-place MAE results, confirming robustness across different error metrics
• TRL 8 achieved — technology qualified and ready for production deployment
• IP filings initiated to protect the novel GLPatch architecture and methodology
• Reusable forecasting framework that can be adapted to new seasonal forecasting domains