Integrating Natural Language Events into Time Series Forecasting through Agentic LLM Orchestration
(2026) In Master's Theses in Mathematical Sciences FMSM01 20261Mathematical Statistics
- Abstract
- This thesis investigates whether Large Language Models (LLMs) can meaningfully reason over natural language event context to produces calibrated adjustments to time series forecasts, and under what conditions that reasoning produces reliable signal. We design an agentic forecasting system in which the LLM acts as a strategic orchestrator rather than a direct numerical predictor: all numerical computation is delegated to a library of validated statistical implementations, while the LLM reasons over textual event descriptions, historical analogues from the target series' own history, and cross-domain precedent cases. This architectural separation isolates the LLM's contribution to the reasoning layer.
The system is evaluated through a... (More) - This thesis investigates whether Large Language Models (LLMs) can meaningfully reason over natural language event context to produces calibrated adjustments to time series forecasts, and under what conditions that reasoning produces reliable signal. We design an agentic forecasting system in which the LLM acts as a strategic orchestrator rather than a direct numerical predictor: all numerical computation is delegated to a library of validated statistical implementations, while the LLM reasons over textual event descriptions, historical analogues from the target series' own history, and cross-domain precedent cases. This architectural separation isolates the LLM's contribution to the reasoning layer.
The system is evaluated through a controlled ablation study on three simulated datasets spanning primary care, parcel logistics, and music streaming. Six conditions systematically vary access to grounding sources, and the pipeline is benchmarked against a numerical foundation model and assessed by both standard accuracy metrics and an independent LLM judge of reasoning quality.
The full pipeline reduces sMAPE by 59% (health center) and 66% (logistics) relative to a no-augmentation baseline, at a runtime overhead of 1.5 times and an average API cost of $0.07 per forecast, and outperforms the foundation model by a factor of 2.0 to 2.5 in MASE on event-driven test windows. The ablation shows that LLM event reasoning requires at least one grounding source to function reliably, and that the dominant source is determined by the structural match between available evidence and the test-period event. The findings characterise LLM event reasoning as a viable and transparent capability for hybrid forecasting where contextual events drive material deviations from baseline patterns. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9229880
- author
- Perntoft, David LU and Atle, Astrid
- supervisor
- organization
- course
- FMSM01 20261
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMS-3553-2026
- ISSN
- 1404-6342
- other publication id
- 2026:E37
- language
- English
- id
- 9229880
- date added to LUP
- 2026-06-03 09:50:25
- date last changed
- 2026-06-03 09:50:25
@misc{9229880,
abstract = {{This thesis investigates whether Large Language Models (LLMs) can meaningfully reason over natural language event context to produces calibrated adjustments to time series forecasts, and under what conditions that reasoning produces reliable signal. We design an agentic forecasting system in which the LLM acts as a strategic orchestrator rather than a direct numerical predictor: all numerical computation is delegated to a library of validated statistical implementations, while the LLM reasons over textual event descriptions, historical analogues from the target series' own history, and cross-domain precedent cases. This architectural separation isolates the LLM's contribution to the reasoning layer.
The system is evaluated through a controlled ablation study on three simulated datasets spanning primary care, parcel logistics, and music streaming. Six conditions systematically vary access to grounding sources, and the pipeline is benchmarked against a numerical foundation model and assessed by both standard accuracy metrics and an independent LLM judge of reasoning quality.
The full pipeline reduces sMAPE by 59% (health center) and 66% (logistics) relative to a no-augmentation baseline, at a runtime overhead of 1.5 times and an average API cost of $0.07 per forecast, and outperforms the foundation model by a factor of 2.0 to 2.5 in MASE on event-driven test windows. The ablation shows that LLM event reasoning requires at least one grounding source to function reliably, and that the dominant source is determined by the structural match between available evidence and the test-period event. The findings characterise LLM event reasoning as a viable and transparent capability for hybrid forecasting where contextual events drive material deviations from baseline patterns.}},
author = {{Perntoft, David and Atle, Astrid}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Sciences}},
title = {{Integrating Natural Language Events into Time Series Forecasting through Agentic LLM Orchestration}},
year = {{2026}},
}