Can decomposing forecasting into stages unlock numerical and contextual reasoning?
This explores whether breaking time-series forecasting into separate stages for contextualization, dual-resolution outlook, and synthesis allows systems to combine the strengths of numerical models and language models more effectively than either alone.
Time-series forecasting has had two roughly disjoint research traditions. Time-Series Foundation Models (TSFMs) excel at numerical extrapolation from historical patterns but are unaware of real-world textual signals — news, events, contextual catalysts. LLMs are emerging as zero-shot forecasters with the opposite profile: strong at contextual reasoning, uneven at numerical extrapolation. Each has what the other lacks.
Nexus bridges them through a multi-agent decomposition. The framework breaks forecasting into three stages. First, Contextualization: structure the raw multimodal context — historical data, news, events, domain signals — into a form the downstream agents can use. Second, Dual-Resolution Forecast Outlook Generation: produce projections at two temporal scales, isolating macro-level trajectories from micro-level event-driven catalysts. Third, Forecast Synthesis and Calibration: a Synthesizer Agent merges the dual-resolution outlooks with domain reasoning into a final calibrated forecast with explicit interpretable reasoning.
The architectural move is the dual resolution. Macro-level reasoning captures seasonal patterns, long-term trends, baseline trajectories — the regime where TSFM-style extrapolation works. Micro-level reasoning captures the event-driven catalysts and shocks — the regime where contextual LLM reasoning matters. By separating them and then synthesizing, the framework gets the strength of each without forcing a single model to handle both simultaneously.
The empirical result: on real-world Zillow real-estate and stock-market datasets, Nexus consistently outperforms both TimesFM-2.5 (a strong TSFM) and CoT LLM baselines. Decomposition into specialized stages adapts from seasonal signals to volatile event-driven information without relying on external statistical anchors or monolithic prompting.
For forecasting practitioners, this argues against pure-TSFM or pure-LLM approaches in domains where context matters. Real-estate prices respond to news (interest-rate decisions, regional events). Stock prices respond to announcements. Energy demand responds to weather and policy. These cases all benefit from architectures that explicitly separate the numerical and contextual contributions rather than asking one model to do both.
Inquiring lines that read this note 21
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
What enables genuine semantic understanding in language models? Can diffusion models match autoregressive performance on language generation tasks? What capability trade-offs arise from domain specialization through fine-tuning? How do recommenders balance exploiting fresh signals against maintaining preference stability? Why don't LLMs reliably translate capability into accurate outputs? How do prompting refinements mask underlying biases and model frequency patterns? What causes reasoning models to fail or wander off track? Can models improve accuracy without degrading reasoning quality? Can intelligent routing over smaller models outperform scaling a single large model? Can prompt-based context override biases that were embedded during pretraining? How do capability benchmark scores systematically misrepresent true model abilities? How should systems decide whether to retrieve or reason alone? How should designers communicate what AI systems truly are and can do? What reasoning architectures enable models to solve complex problems efficiently? How well do AI systems understand human social norms? Does model confidence reliably signal actual accuracy in practice? How effectively can language models perform reasoning, especially combined with symbolic methods? How does evaluation scope and dimensionality affect what we measure? How does decomposing tasks improve reasoning and prevent failure propagation? How does AI adoption across firms reshape employment and inequality?Related concepts in this collection 3
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Can LLMs actually forecast time series better than we think?
Explores whether language models possess stronger forecasting ability than current benchmarks suggest, and what role workflow design plays in revealing or hiding that capability.
same paper, the broader observation
-
Does separating planning from execution improve reasoning accuracy?
Can modular LM architectures that split problem decomposition from solution execution outperform monolithic models? This explores whether decoupling these cognitive operations reduces interference and boosts performance.
adjacent: same decomposition principle in a different domain
-
Can reasoning systems maintain memory across retrieval cycles?
Existing retrieval systems treat each lookup independently. But what if reasoning required a persistent memory workspace that evolves as contradictions emerge and understanding deepens?
adjacent: another multi-stage decomposition for evidence-integrating reasoning
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Nexus: An Agentic Framework for Time Series Forecasting
- Approaching Human-Level Forecasting with Language Models
- A Decomposition Perspective to Long-context Reasoning for LLMs
- SoftCoT: Soft Chain-of-Thought for Efficient Reasoning with LLMs
- Can Large Language Models do Analytical Reasoning?
- Divide-or-Conquer? Which Part Should You Distill Your LLM?
- GenAI as a Power Persuader: How Professionals Get Persuasion Bombed When They Attempt to Validate LLMs
- Large Causal Models From Large Language Models
Original note title
multi-agent decomposition into contextualization dual-resolution outlook and synthesis stages enables time-series forecasting that integrates numerical patterns with unstructured contextual data