Introducing GeonatiQ

Carbon LLM cont…

How it works: Decompose Carbon Prices & Identify News Attention Our LLM breaks down carbon prices into high frequency, low frequency, and long-term residual components. It then learns which type of news each component attends to the most, extracting signals that traditional models miss.

Insights from R&D:

High-frequency signals Correspond to macro/market sentiment headlines (e.g., bear market fears).

Low-frequency signals Relate to structural/partnership news (e.g., Amazon rainforest monitoring projects).

Residual/long-term signals Map to monetary policy & central bank updates (e.g., ECB rate decisions).

So?

A 5-month news attention history was used to predict 30-day forward prices. Achieved MAE: 2.2 EUR, with strong directional accuracy.

Key takeaway:

The LLM “self-learns” how different news types map to market horizons, creating a powerful, explainable forecasting engine for EU ETS carbon prices.