Information metrics benchmarked for predicting time series with text annotations
When Does Text Inform? Benchmarking Information-Theoretic Metrics for Multimodal Time-Series Forecasting
Artificial IntelligenceInformation Theory
Summary
Texts added alongside time-based data can help predict what happens next, but it’s hard to tell if the text really helps. The authors created fake data with text that is correct, wrong, or irrelevant to know exactly how much helpful information the text provides. They tested six different ways to measure this helpfulness and found all methods can identify the useful text and pick the best annotations for predictions without needing to train models. They also tested these methods on real data and provided practical advice on how to use them.
What this means in practice
- •For data scientists: Select the most informative text annotations to improve time series forecasting accuracy without retraining models.
- •For business intelligence teams: Audit mixed text data sources to identify which annotations enhance forecasting reliability for decision-support systems.
Authors
Emma Andrews, Gianmarco Mengaldo
Abstract
Multimodal forecasting models that combine time series with text annotations promise richer prediction through textual context, but how do we know whether a text annotation meaningfully contributes to the forecasters prediction? This is an information-theoretic question, but to evaluate whether information-theoretic metrics can reliably measure the predictive value an annotation provides, a ground truth benchmark is needed, and none currently exist. We create a synthetic time series signal with annotations in three categories: semantically correct, incorrect, and irrelevant. Because the data generation process is fully controlled, ground-truth information content is known exactly, enabling principled evaluation of six complementary mutual information estimators (KSG, MINE, InfoNCE, CCA, PID and V-information). We show that all six estimators identify correct annotations as most informative, and are able to audit the quality of mixed text corpora, choosing the annotations that result in the best downstream forecasting results without the need for model training. Our benchmark identifies limitations of each estimator, and these are validated on seven real-world datasets, which show how estimator performance differs on weak signals. Finally, we establish practical rules for implementing these metrics for annotation auditing and fusion selection.