FACTS-MAS weights forecasts by each evidence channel's skill
Original title: FACTS-MAS: A Multi-Agent Architecture for Adaptive Forecasting from Heterogeneous Evidence
On housing inventory and unemployment claims for 15 U.S. metros, it beats TimesFM, while factor-adjusted exponential smoothing matches the fused forecast.
FACTS-MAS splits heterogeneous evidence into agents for temporal dynamics, exogenous context, seasonality, structure, and discrete events. Leakage-safe historical validation scores each agent's skill against a naive reference and uses those scores to weight or drop channels. A calibration step picks domain-specific constant means from validation folds. Evaluation covers weekly Zillow for-sale housing inventory and weekly FRED unemployment-insurance claims at 4-, 8-, and 13-week horizons. The authors say the result supports the framework as a protocol for testing channels before fusing them.
- Uninformative evidence is removed or downweighted rather than always fused.
- Horizons tested are 4, 8, and 13 weeks.
- It beats a naive last-value forecast and the TimesFM foundation model.
- Code, data, and results are linked from the paper's GitHub repository.
- Organizations
- Zillow
- Countries
- United States
- AI models
- TimesFM
From the original article
1 min read at arxiv.orgReal-world forecasting often requires integrating heterogeneous evidence whose predictive value varies across horizons, operating regimes, and domains. We introduce FACTS-MAS, a domain-agnostic multi-agent forecasting framework that decomposes evidence into specialized channels for temporal dynamics, exogenous context, seasonality, structural information, and discrete events.
By Madumita Karthikeyan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.