A preprint study asks when auxiliary context can genuinely help a time-series forecaster. Its answer is conditional: two things need to be true at once. The target must not be dominated by a last-value shortcut, and the context must contain information beyond recent history. In ordinary language, a model needs more than a strong tendency to repeat the latest reading, while the added material needs to tell it something the recent numbers do not already reveal. Without both conditions, the study’s hypothesis says the benefit attributable to context should be negligible.
This framing shifts attention from whether a model can accept extra input to whether the input is doing useful work. The study examines text-conditioned fusion, where information from text is routed into a time-series forecaster. It evaluates eight main datasets, including six Time-MMD sub-datasets and 50 S&P500 stock series from FinMultiTime. The sample is therefore a mixture of forecasting datasets and stock series, rather than one single population. That matters when reading the percentages: they describe results across named settings, not one universal average.
A test for the extra signal
To screen for that extra signal before training, the paper uses a mutual-information test. Mutual information is a measure of how much knowing one quantity tells us about another. The procedure first residualizes the target against the latest observation, removing the part associated with that last-value signal. It then projects the context onto its top 20 principal components, a compact representation, and estimates the remaining dependence with a Kraskov k-nearest-neighbor method using k = 3. Finally, it shuffles the context 200 times. The procedure is designed to ask whether the target information left over after the latest observation is related to the compressed context.
The result is not a simple yes-or-no switch. The diagnostic returns TRY_FUSION when p < 0.05. It returns SKIP_FUSION when p ≥ 0.05 and the test has power of at least 0.8, or 80%, meaning the researchers judge the negative result sufficiently well tested under the rule. When p ≥ 0.05 but power is below 0.8, it returns INCONCLUSIVE. That last category is important: a failure to find a relationship is not treated as evidence against fusion when the test may not have been strong enough to detect one.
Results split sharply by dataset
MoME’s controlled text-modulation results varied widely. Across six datasets, the routing contribution was 50.9% on HealthUS, 41.5% on Environment, 32.0% on SocialGood and 28.7% on HealthAFR. The corresponding figures were much smaller on FinMultiTime, at 6.7%, and Energy, at 4.3%. The Energy estimate was not statistically distinguishable from zero. The pattern is consistent with the study’s central condition: context can look valuable in some settings and marginal in others, depending on what recent history already captures and what the extra input adds.
The percentages come from a particular architecture, not from a generic fusion model. MoME is a 14.3-billion-parameter mixture-of-experts model. Its Expert-level Language Modulation pathway, or EiLM, has approximately 74,000 parameters, less than 0.001% of the full model. When modulation is toggled, the expert structure and routing weights remain unchanged. In practical terms, the reported routing contribution is tied to the modulation pathway inside this model setup. It should therefore be read as a result about MoME’s text modulation across the tested datasets, not as a blanket figure for every way of combining context with a time series.
A narrower promise for context
The practical reading is a pre-training checkpoint. First ask whether the target is dominated by its most recent value. Then ask whether the auxiliary context still carries information after that latest observation is accounted for. The paper’s diagnostic turns those questions into three labels, with TRY_FUSION reserved for a significant result, SKIP_FUSION for a non-significant result with adequate power, and INCONCLUSIVE when power is low. That approach puts the burden on evidence that the context adds something, while leaving room for more testing when the data cannot give a clear answer.
The study’s strongest positive numbers and its decision rule answer different questions. The MoME results show how routing contributions varied in the tested settings, while the mutual-information screen asks whether extra signal remains after the latest value is accounted for. Together, they make the study’s message narrower than a promise that context always helps. Context earns a place in the forecast only when it adds signal beyond the latest value and the diagnostic can detect that signal.
Paper data and sources
Original title: When Does Context Routing Help? A Systematic Study of Multi-Modal Fusion in Time Series Forecasting
Authors: Ruizhe Zhou, Gaoyuan Du, Xiaoyang Liu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text