Four weeks ago Google shipped TimesFM 3 and the headline feature was covariates: the ability to tell a forecaster about things other than the series it is predicting. We wrote about it at the time and called covariates the genuinely new part.
QiYao-M, posted to arXiv on 28 September by Hanyin Cheng, Linfeng Wang, Zhengbo Qu, Yang Shu and colleagues at East China Normal University and Huawei, makes a sharper claim about the same idea. Covariates are not one category.
The work is theirs, not ours. What follows is a reading of their paper, with the results attributed and linked, plus the part we think matters commercially and which the benchmark table rather buries.
1. The distinction
Their definitions are precise and worth quoting, because the whole design follows from them.
Endogenous modalities are "derived from the time series itself and provide alternative views of the same underlying temporal dynamics." Their examples are statistical feature descriptions and line plots of multi-order differences, extracted from the numbers you already have.
Exogenous modalities "originate from external sources and provide additional evidence that may influence future evolution." Text, satellite imagery, tabular data from somewhere else.
In the language of somebody running a demand forecast: endogenous is a restatement of your own sales history. Exogenous is the weather, the promotion calendar, the competitor's price change, the bank holiday.
The authors' complaint about existing multimodal forecasters is that they push both through "largely shared mechanisms, overlooking the distinct forecasting roles" of the two. Their argument is that the roles are genuinely different:
- An endogenous signal evolves with the thing you are predicting. It has a future, and that future is coupled to the series. So the model should learn how it moves from history into the future.
- An exogenous signal is evidence about the future rather than part of it. You cannot learn its dynamics from the series, and you often have almost no pretraining data for it, because everybody's external signals are different.
That asymmetry is the paper, and once stated it is hard to unsee. It is also a fair description of why covariate support in forecasting tools so often disappoints: the tool treats a promotions flag and a rolling average as the same kind of input.
2. What they built
Four components, and the naming is at least descriptive.
For the endogenous side, an Endo-Multimodal Predictor and Endo-Multimodal Supervision, which together make the model explicitly learn how those signals evolve from history to future rather than just consuming them as context.
For the exogenous side, an Exo-Multimodal Retrieval Enhancer. This is the interesting one. Rather than training the foundation model to use your external signals, it retrieves historical cases with similar multimodal conditions and uses those to condition the forecast. The authors are explicit that this "enables rapid downstream adaptation without updating the TSFM parameters."
And because there is almost no exogenous multimodal pretraining data in the world, a fourth piece: Endo-Modality Proxy Training, which trains the retrieval module by sampling endogenous modalities as stand-ins, so the retriever can be trained at all.
3. The numbers, read plainly
The evaluation is broad, which is to their credit. Unimodal benchmarks are GIFT-Eval across 23 datasets and TIME across 50. Multimodal benchmarks are Time-MMD spanning agriculture, climate, energy, health and social good, MoTime covering Taobao-Fashion and Tianchi, and FinMultiTime covering HS300 and SP500. Baselines include Chronos-2, Toto-2.0, TiRex-2, PatchTST-r2, Zeus, ChatTime and Aurora among others.
| Result | Figure |
|---|---|
| MASE vs strongest baseline (TiRex-2), unimodal | 1.0% better |
| CRPS vs strongest baseline, unimodal | 1.3% better |
| Average MSE reduction from exogenous information | 5.1% |
| Multimodal metrics ranked first | 14 of 18 |
Now the honest reading, because the framing matters more than the arithmetic.
One percent is not the story. A 1.0 percent MASE improvement over TiRex-2, across a 23-dataset aggregate, is a real result in a competitive field and it is close to nothing for a business deciding what to deploy. It will not survive contact with your data as a reliable edge, and treating it as the reason to adopt would be a mistake.
The 5.1 percent is the story, and it is a different kind of claim. That figure is not "our model beats their model." It is "giving the model external evidence is worth 5.1 percent MSE, and here is a way to do it that generalises." Those are different propositions and only the second one is about your business, because you are the one who has the promotion calendar.
This is the same shape of misreading we flagged in AGI talk measures the wrong axis: the number that gets quoted is the model-versus-model aggregate, and the number that decides anything operationally is somewhere else in the table.
4. The part with the commercial payoff
Strip the naming and the exogenous design says something a forecasting team will recognise immediately.
In most production setups, the real cost of a new signal is not the signal. It is the retraining cycle it triggers. Somebody wires up the weather feed, and then the model has to be retrained, revalidated against holdout, compared against the incumbent, signed off and redeployed. That is weeks. It is why so many forecasting systems run on a feature set somebody froze eighteen months ago, and why the obvious idea of "we should tell it about the bank holidays" keeps not happening.
A retrieval module that conditions on similar past cases, with the foundation model's weights untouched, changes what that costs. Adding a signal becomes a wiring job and an evaluation, not a training project. If it does not help, you unwire it.
We would want to test three things before believing it end to end, and the paper does not answer them:
- What happens on a signal with no historical analogue? Retrieval works by finding similar past conditions. A genuinely novel shock, which is exactly when you most want the external signal, is the case with no close match.
- How does it degrade as the exogenous set grows? Adding one feed is the easy demo. Twenty correlated feeds is the real system.
- What is the retrieval latency in a production forecast run? A nightly batch does not care. An in-request forecast does.
The first of those is the one we would test first, because it is where the failure would be both most likely and most expensive.
5. The sceptical question about the other half
The endogenous half deserves a harder look, and we would raise it as a question rather than a verdict.
Rendering a series as statistical descriptions or as plots of multi-order differences does not add information. It is the same numbers in another format. So whatever gain arrives from the endogenous machinery is a gain from making structure that was already present easier for the model to use.
That can be a genuine and useful effect. Representation matters and there is a long history of it mattering. But it sits closer to feature engineering than to new evidence, and it raises the obvious question of whether a larger or better-trained unimodal model would learn the same structure directly from the numbers. The 1.0 percent unimodal margin is consistent with either answer.
We would want the endogenous and exogenous contributions ablated separately before attributing much to the first. The paper reports strong results in scenarios both with and without exogenous modalities, which is the right experiment to have run, and the public abstract does not break the two apart in a way that settles it.
6. The baseline nobody wants to run, again
We made this argument when Google shipped TimesFM 3 and it applies unchanged, so we will keep it short.
A foundation model that beats another foundation model by one percent on a 23-dataset aggregate has told you nothing about your series. Retail demand with a heavy weekly cycle, warehouse throughput, transaction volume and energy load are different problems, and the aggregate is an average over a distribution your data is probably not drawn from.
Run seasonal naive. Same backtest, same holdout, same error metric, no special pleading. It takes an afternoon. Then run the foundation model in the same harness. If the gap is not large and consistent, ship the baseline, because it has no inference cost, no dependency, no licence question and no retraining story.
If the gap is large, then price the pipeline against it. The lift has to pay for the serving cost, the monitoring, and the person who understands it. That calculation is the same one we set out in the small language models note and in what tokens actually cost: cost per useful output, not score per benchmark.
7. Where this lands for operations work
The endogenous and exogenous split maps onto real systems more cleanly than most architectural distinctions do.
In the intelligent warehouse the endogenous signals are throughput, pick rates and historical volumes. The exogenous ones are the weather, the carrier's service notice, the promotion the commercial team ran without telling anybody. The second group is where the forecast breaks, and it is exactly the group that is hardest to get into a model.
The same split runs through agentic commerce and quantitative signal discovery, where the distinction between a derived indicator and genuine external evidence is the difference between a feature and an edge.
One more thing worth noting from the release: the Hugging Face weights carry an MIT licence. As we argued about TimesFM 3, for anything you intend to run in a product the licence frequently decides more than the benchmark table, and MIT is about as unproblematic as it gets.
8. What we take from it
Three things, stated at the level of confidence we actually hold them.
High confidence: the distinction is correct and useful. Treating derived views and external evidence as one category is a real modelling error, and naming it is worth more than the percentages attached to it. We expect to use this distinction when designing forecasting systems regardless of whether we ever run this model.
Medium confidence: the retrieval approach is the commercially important idea. Adapting to a new external signal without touching model weights removes the main practical barrier to using external data. The three questions in section 4 have to be answered first.
Low confidence: the specific numbers. One paper, days old, no independent replication, reported by the authors. The 5.1 percent is more interesting than the 1.0 percent, and neither is a reason to change anything until it has been reproduced on your own series.
We have not run QiYao-M. This is a reading of a paper and a model card published by researchers at ECNU and Huawei, whose work this is.
If you are forecasting demand, throughput or volume and you suspect your model is ignoring the external signals that actually move it, talk to us, or see how we build systems for businesses.
9. Sources
The work below belongs to the researchers and institutions named. Links go to the originals.
- Hanyin Cheng, Linfeng Wang, Zhengbo Qu, Yang Shu, Zhongwen Rao, Meng Wang, Yijie Li, Xin Jiang, Bin Yang and Chenjuan Guo (East China Normal University, Huawei), QiYao-M: Multimodal Time Series Foundation Model with Role-Aware Modeling of Endogenous and Exogenous Modalities, arXiv 2609.34842, 28 September 2026. Definitions, architecture and all figures quoted above. Also on alphaXiv.
- QiYao-M model card on Hugging Face, for the MIT licence and the component descriptions.
- Google Research, TimesFM 3, covered in our earlier note on forecasting as next token prediction, for the covariate support this work sharpens.
- Benchmarks referenced by the authors: GIFT-Eval, TIME, Time-MMD, MoTime (Taobao-Fashion, Tianchi) and FinMultiTime (HS300, SP500).
This note sits in our Research track, alongside the operations, foundation models and multimodal threads. The full library is at Research.