FACTS-MAS is a domain-agnostic multi-agent forecasting framework that decomposes heterogeneous evidence into five specialized channels: temporal dynamics, exogenous context, seasonality, structural information, and discrete events. Each agent produces an independently testable forecast through a common interface, and a leakage-safe historical validation measures its skill relative to a naive reference. The skill estimates weight the agents’ forecasts, allowing uninformative evidence to be removed or down-weighted rather than forced to contribute. A calibration step selects domain-specific constant means from validation folds. We evaluate the system on weekly for-sale housing inventory from Zillow and weekly unemployment insurance claims from FRED, covering 15 U.S. metropolitan areas at 4-, 8-, and 13-week horizons. FACTS-MAS outperforms a naive last-value baseline and the TimesFM foundation model, while matching the performance of a factor-adjusted exponential smoothing model. These results support FACTS-MAS as a systematic protocol for testing, calibrating, and recombining evidence channels before assembling the strongest single model for a given domain. Code, data, and results are available at https://github.com/Madumita24/FACTS-MAS.
Review