DeepSeek-TS+: A Unified Framework for Multi-Product Time Series Forecasting

Author(s): Shenggang Li

Originally published on Towards AI.

Leveraging State-Space Enhanced Multi-Head Latent Attention and Group Relative Policy Optimization (GRPO) for Adaptive Forecasting

This member-only story is on us. Upgrade to access all of Medium.

I was impressed by DeepSeek’s technology — its efficient Multi-Head Latent Attention (MLA) and Group Relative Policy Optimization (GRPO) techniques inspired me to apply them to multi-product time series forecasting.

In our approach, we extend MLA into what we call MLA-Mamba, allowing the latent features to evolve dynamically over time using a state-space model with non-linear activations. This gives our model an adaptive memory that adjusts to trends much like a sales team adapting its strategy during market surges.

At the same time, GRPO introduces a smart decision-making process that continuously refines forecasts by comparing predictions against a baseline, similar to a manager tweaking forecasts on the fly. This dynamic adjustment helps our model respond effectively to sudden changes in sales patterns.

We compare our approach with classical ARMA models and standard GRU-based networks. While ARMA handles linear trends and GRUs capture temporal dependencies, our DeepSeek-TS framework is designed to model complex inter-product relationships and adapt to non-linear dynamics, resulting in more accurate and robust forecasts.

In the following sections, we break down the technical details of our extended MLA (MLA-Mamba) and GRPO frameworks, and demonstrate how their… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI