Third Annual Risk Conference 2025
What it is about
Miquel Noguer i Alonso, of the Artificial Intelligence Finance Institute, presented at the Third Annual Risk Conference (September 2025) a neural network architecture for financial prediction that combines data of different natures (market, sentiment and fundamental events). It is a technical presentation and the slides are in English.
The 19 slides set out the challenge of integrating asynchronous and heterogeneous data, the principles of the MS-MRFN architecture, its three encoders (returns at several frequencies, sparse sentiment and fundamental events), the fusion module and prediction head, the model equations and the training protocol.
Objectives
- Explain why financial prediction requires integrating data with different frequencies and formats.
- Describe the components of a multimodal, multi-resolution network.
- Present the training protocol that avoids lookahead bias.
Key takeaways
- Financial data combine dense, high-frequency market series, sparse sentiment signals and infrequent fundamental reports.
- The architecture respects the native sampling frequencies, explicitly models the staleness of the data and strictly enforces time ordering.
- It uses three parallel encoders: temporal convolutional networks with LSTM for returns, a similar stack for sentiment and a transformer with cross-attention for fundamental events.
- A modulation layer (FiLM) adjusts how much relevance is given to fundamental information according to its staleness, and an adaptive fusion produces the final prediction.
- Training uses walk-forward validation, with chronological training, validation and test splits, to avoid lookahead bias.
Connect with our networks