JointFM-0.1: A Foundation Model for Multi-Target Joint Distributional Prediction

2026-03-24 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, long

Summary

JointFM is a novel foundation model designed for multi-target joint distributional prediction, inverting the traditional Stochastic Differential Equation (SDE) modeling paradigm. Instead of fitting SDEs to data, JointFM is trained on an infinite stream of synthetic SDEs to directly predict future joint probability distributions of coupled time series. This approach eliminates the need for task-specific calibration or finetuning, operating in a purely zero-shot setting. In experiments, JointFM reduced energy loss by 14.2% relative to the strongest baseline when recovering oracle joint distributions from unseen synthetic SDEs. The model employs a Transformer-based architecture with Variable-Factored Attention and supports output heads like Multivariate Gaussian Mixture (MV-GMM) and Skewed Student-t Mixture to capture complex dependencies, heavy tails, and asymmetry. This enables instant, general-purpose quantitative modeling for domains like portfolio optimization and probabilistic grid balancing.

Key takeaway

For research scientists developing quantitative models for coupled time series, JointFM offers a paradigm shift by providing instant, zero-shot joint distributional predictions. You should consider integrating this synthetic-physics pretraining approach to bypass brittle, computationally expensive SDE calibration and simulation, enabling real-time risk-aware decision-making in complex, dynamic environments.

Key insights

JointFM is a zero-shot foundation model predicting future joint probability distributions by pretraining on synthetic SDEs.

Principles

• Pretrain on diverse synthetic SDEs for generalization.
• Directly predict joint distributions, bypassing SDE fitting.
• Inference cost is independent of underlying dynamics complexity.

Method

JointFM trains on an infinite stream of procedurally generated SDE systems, simulating history realizations for context and tens of thousands of future paths for ground-truth joint distributions, then predicts parameters for flexible joint distributions.

In practice

• Use JointFM for real-time portfolio risk assessment.
• Apply to energy market dispatch and supply-chain hedging.
• Integrate into agentic AI workflows needing probabilistic reasoning.

Topics

• Foundation Models
• Joint Distribution Prediction
• Stochastic Differential Equations
• Time Series Forecasting
• Zero-Shot Learning

Best for: Research Scientist, AI Researcher, AI Scientist, Data Scientist

Related on AIssential

• The digital quant: instant portfolio optimization with JointFM — JointFM is a zero-shot foundation model predicting full multivariate distributions for instant, domain-agnostic time-series forecasting.
• Understanding Key Features of Time Series Foundation Models from Epidemic Forecasting — Heterogeneous pretrained representations improve epidemic forecasting, especially with aligned pretraining domains.
• Decoupled and Divergence-Conditioned Prompt for Multi-domain Dynamic Graph Foundation Models — DyGFM enables multi-domain dynamic graph foundation models by decoupling semantics and dynamics, and managing divergence.
• Calibrating Scientific Foundation Models with Inference-Time Stochastic Attention — Stochastic Attention improves scientific foundation model uncertainty calibration via a lightweight, inference-time modification.
• Learning What Not to Impute: An Uncertainty-Aware Diffusion Framework for Meaningful Missingness — Diff-Joint jointly infers meaningful missingness and imputes observation-based missing values using a diffusion framework.
• LatentWave: JEPA Pretraining for Wireless Foundation Models — JEPA pretraining for wireless foundation models enhances representation transferability by predicting masked regions in latent space.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.