My question: What kind of dynamical system observability does Timestral provide? Is it a blackbox y = F(x, params) model for offline use? Can it provide the actual dynamical system equations? Or are you stuck forever querying the Timestral server (which is bad, think about if you're grid system operator doing day ahead, month ahead forecasts, and Timestral goes down)
Typically, in linear, non-linear, and NN models, we fit some noisy data to A) An underlying model with unknown parameters or B) A model free search which attempts to construct combination of "basis" functions or C) A purely black box model which we only care about Input/Output error.
A) is useful because often we know an underlying model we want to match, but don't know the parameter values. If we can determine the parameters, we get insights about the real system. E.g, how are these two variables coupled and how strongly -> We can make business decisions about what happens if we adjust some parameters.
So what exactly does this timeseries foundation model do and provide?
Maybe, the observability is that you have a forecast, and you adjust the covariates and see what happens. But one would want much more insights than a "executive summary" of the forecast.
The important parts for me as a signals processing R&D engineer.
> Classical models such as ARIMA and linear regression remain well-established tools for forecasting time series. These models are usually fitted to a single series and learn only from that series' history.
> Deep learning models such as PatchTST add the capacity to learn nonlinear patterns across many related series, but they are still typically trained from scratch on each dataset, so what they learn doesn't carry over to new domains.
> Time series foundation models bring the pre-training approach to time series. Trained on a large and diverse corpus, our model learns patterns that transfer across series, forecasting zero-shot on data it has never seen, even in domains where data is scarce.
Edit: Looks like it's a pure black box which you can run on premise.
> Output: Probabilistic forecasts, 21 quantiles
> Inputs: Multivariate targets, past-only and known-future covariates
My field is primarily audio/vision ML so I don't have any particular domain expertise, but I don't get how you can lump all "time-series" together in one big bucket. My naive understanding is that the way to improve the accuracy is to gain as much joint information as possible. Like:
> Known-future covariates. Some drivers of sensor behaviour are known in advance, such as production schedules, planned shutdowns, weather forecasts. Timestral can condition on them to improve forecast quality, so a scheduled load increase is expected rather than flagged as a surprise.
To me this should be the entire basis of the system, not a condition to improve forecast quality? Like if you're modelling electricity consumption, if all you have is a window of N previous samples, you're going to get completely beaten by a comparatively simple system that knows the time of day, that there's holidays coming up, the weather forecast, etc.
I would also think that these time-series forecasting models would gain from a sophisticated understanding at least the labels of the covariates? A 5M model has no idea what a "bank holiday" is, it just sees an arbitrary class go from 0 to 1, and won't have any useful inductive biases as to what it means or how to link them to other pieces of information.
E.g., wouldn't you predict €/MWh of electricity far more accurately with access to price of natural gas (itself influenced by many geopolitical factors), breakdown in composition of production of each energy source, predictions about roll-out (requiring a sophisticated LLM capable of understanding news and their implications), etc. What's the use in trying to predict it from 2 variables and a narrow window looking back at just 8192 or 32768 samples?
Or perhaps I misunderstand the use-case and this is more for extremely short term movements?
For some kinds of forecasting it doesn't need to be exactly accurate - eg say you are forecasting traffic for how to scale a load balancer, it can be a bit off it will just cost you a little and you can adjust after - a lot of feedback systems dont need full accuracy. You wouldnt use this to forecast whether to buy shares. It basically replaces various traditional curve fitting tools. If you need a really good forecast yes you are going to build a domain model. At least this is my understanding...
You're right but, if the time series looks like music then you can apply your music domain knowledge. You don't have to only apply knowledge that's general across every domain, part of the magic of large neural nets is that you can do recognition and prediction at the same time thereby using (statistically) things known from training about how to identify a sequence and then also how to continue it.
Sure adding outside info helps even more, but with enough data it should become unnecessary (or so goes the theory).
Do you actually find that for some use cases foundation models work better than AutoARIMA or etc? I do a lot of ecom projections and it’s never been the case for me, most of my cases are best with seasonal naive.
Would love to know actual time series foundation model use cases!
But why would you throw away that data at all? These are tiny models compared to domain-specific audio/music/image models, and any capacity spent on recognition is capacity that could have been used for prediction.
Is there a reason you don't throw in something like cross-attention on an embedding of the sequence description/data labels? I've trained numerous generative models for audio and images and it's always the case that the unconditional model performs worse, trains slower, etc. than the conditional one (with the best performance from a hybrid using classifier-free guidance or similar scheme).
Again I don't have experience with the domain of time-series forecasting directly, but I would think there are generalities between autoregressive generative models and time-series forecasting, given that both involve predicting the next sample(s) given N regularly sampled past inputs.
I'm doing time series forecasting for a biological system. I haven't used these foundation models but my experience is that knowledge of how things are related is baked into what variables you put into the model (which can be more than raw data, you can make any derived product you think would be useful)
but it's up to the model to figure out how they are related (if I knew exactly how they were related I wouldn't be training a forecasting model).
I'm sure someone who actually knows this stuff has a better easy to view it but my impression was that these time series foundation models aren't trained for a specific dataset but more they are trained on a procedure for how to look at datasets and make predictions. It doesn't seem to lump things into a bucket much more than any other model type (which also are just consistent procedures to look at datasets). It's not like you have radically different model types across different fields.
I can see another use for time-series generative models; to create the right shape of noisy data.
To improve the robustness of machine learning models, artificial noise is often introduced into actual training data. Something like Timestral could simplify the process of generating it.
I interviewed for a company which used time series modelling to predict engine failure on ships. Though they didn't call it that, they called it just "AI" and weren't familiar with the time series literature at all. (Neither was I, but I did at least know that the time series literature was a thing.)
They just used a bunch of sensors attached to the engine.
That seems like a scenario where there wouldn't be external sources of information capable of improving your predictions. It also seemed to me like a scenario where good old fashioned statistical methods would outperform LSTMs, which was what they used.
(I didn't get the job - I was offered it, but they rescinded their offer at the last moment, supposedly because their investors had insisted they should hire a web developer instead.)
I appreciate this analysis, but I wonder, setting aside unpredictable current events for the moment, aren’t seasonal and scheduled events also covariates? eg forecasting baseball scores year round?
A lot of misconception around the utility of foundational time series models on HN.
Sure, you could train a XGboost model on a time series with holiday or event covariates, but what if you have 10k different series to forecast for?
You could train 10k different models, one for each series, but then they have very little history to learn from. Alternatively, you could train a single generic model to cover all 10k, but then you have to deal with scale mismatch between series in addition to the feature dev overhead between series. A TSFM saves a ton of work in modeling each of these series and could potentially generalize better to produce more accurate forecasts.
What's the input like? Does it incorporate language understanding, so that each time series has a natural language explanation of the series and all the data fields? If it takes in really descriptive text like that I can see the opportunity for generalization. If it's just a time series of non-descript numbers I see much much less opportunity for cross-series learning.
All the businesses with product catalogs for example, every product can have several time series associated with it.
If a business has 10k employees, you have, at minimum, 10k time series to perform employee analyst operations over. There are several hundred companies in the US with 10k+ employees.
There are many more than 10k postal zipcodes in the US to perform forecasts over.
To meaningfully test a forecasting model, use a data-generating process with a high signal-to-noise ratio where a classical method (empirical conditional distributions, i.e., counting) can capture most of the signal. Then check whether the model captures that signal too. If you only test on high-noise processes, any model looks comparable to SOTA because there's little signal to differentiate them. Also testing on a range of signals with different signal-to-noise ratio would allow us to know how much of the signal the model captures in different scenarios.
Any 'vibe check' on these; I don't have benchmarks but I tried out Le Chonk for general coding work that they claim it works really well at and I was very disappointed, it seemed to make all kinds of obvious errors of the kind that I haven't seen from opus, sol, deepseek flash etc in over a year
I've been using Mistral Large 4 and some of my other friends have as well.
Something is going on with the performance (as in speed of the service, not quality). Some days it's quite fast, some days it's so slow I interrupt it because I thought it was hanging.
I think they're shifting hosting capacity around, or otherwise messing with the model. Note that it's still a "preview" version, meaning they're still training it further for the final release and they are likely still figuring out the optimal hosting configuration as well.
It wouldn't surprise me to see further changes in both quality of the model and speed before the final release.
Odd, I am in France too. Progressive rollout maybe? I find it weird to announce a model with fanfare but not make it immediately available to customers, let alone paid ones.
It was for me too, and it's only become slow very recently. I can't say with certainty, but I think it was around the time Mistral 4 launched that it seemed to slow down.
Very nice. It’s mind-blowing to me that it used to take me months to build and deploy a forecasting model five years ago. Now in minutes I can use something like this, implemented with help from an LLM, and get apparently quite good results (and fine-tune on my data, it seems).
Speed looks nice too. I want an open weight model with this performance now.
Ah thanks!
The Mistral model has a restrictive license, where you need a specific commercial agreement to use it if you make more than 20M$ revenue/month. Ours is Apache 2.
If you compare the 2, it's not a strict "timestral is better" story - the winrates are tight - your probably should try both on your use-case. We have a 42% winrate vs Timestral on fevbench for instance.
Ofc, our main value is not just these models, but also 1) extra capabilities (like our model that can natively account for stockouts/censored data, and others coming soon), 2) our systems and products around the model, and 3) we're all laser focused on time-series/forecasting
This is the pseudo-intellectual version of your boomer relative ranting about the weatherman on TV, “how does this guy still have a job when he’s always wrong!”
Or the poll-bashing that happened after the 2016 election when an outcome with a 29% chance per one model was observed.
Any interesting forecasting problem is underconstrained. Monte Carlo methods are the classic way to mitigate that one-universe problem.
Nobody can predict the future with certainty. All models are wrong, but some are useful.
Meh....time series data is fundamentally different from textual data. I see why a time series "foundation model" seems appealing, especially to the marketing folks at Mistral. Real-world usability is very limited, which is probably the reason why the blog post features an imaginary customer rather than an actual reference.
Time based data is extremely important, and there are tons of techniques out there to train super fast super small models, but they are a pain in the ass to tune. I could see a foundation model that is kind of good at a lot of things, that could be fine tuned or used when you want something more straight forward to use to analyze a bunch of data.
Honestly this might be very good for feature selection to refine a much more specialized model.
You are telling me quant funds prefer a heavy-weight and slow ANN to a fast, specialized model? Maybe in totally non-critical environments, otherwise it's simply too slow, regardless of performance.
Typically, in linear, non-linear, and NN models, we fit some noisy data to A) An underlying model with unknown parameters or B) A model free search which attempts to construct combination of "basis" functions or C) A purely black box model which we only care about Input/Output error.
A) is useful because often we know an underlying model we want to match, but don't know the parameter values. If we can determine the parameters, we get insights about the real system. E.g, how are these two variables coupled and how strongly -> We can make business decisions about what happens if we adjust some parameters.
So what exactly does this timeseries foundation model do and provide?
Maybe, the observability is that you have a forecast, and you adjust the covariates and see what happens. But one would want much more insights than a "executive summary" of the forecast.
The important parts for me as a signals processing R&D engineer.
> Classical models such as ARIMA and linear regression remain well-established tools for forecasting time series. These models are usually fitted to a single series and learn only from that series' history.
> Deep learning models such as PatchTST add the capacity to learn nonlinear patterns across many related series, but they are still typically trained from scratch on each dataset, so what they learn doesn't carry over to new domains.
> Time series foundation models bring the pre-training approach to time series. Trained on a large and diverse corpus, our model learns patterns that transfer across series, forecasting zero-shot on data it has never seen, even in domains where data is scarce.
Edit: Looks like it's a pure black box which you can run on premise.
> Output: Probabilistic forecasts, 21 quantiles
> Inputs: Multivariate targets, past-only and known-future covariates
"We have published a GitHub repository so you can run Timestral on your own data."
[1] https://github.com/mistralai/timestral
> Known-future covariates. Some drivers of sensor behaviour are known in advance, such as production schedules, planned shutdowns, weather forecasts. Timestral can condition on them to improve forecast quality, so a scheduled load increase is expected rather than flagged as a surprise.
To me this should be the entire basis of the system, not a condition to improve forecast quality? Like if you're modelling electricity consumption, if all you have is a window of N previous samples, you're going to get completely beaten by a comparatively simple system that knows the time of day, that there's holidays coming up, the weather forecast, etc.
I would also think that these time-series forecasting models would gain from a sophisticated understanding at least the labels of the covariates? A 5M model has no idea what a "bank holiday" is, it just sees an arbitrary class go from 0 to 1, and won't have any useful inductive biases as to what it means or how to link them to other pieces of information.
E.g., wouldn't you predict €/MWh of electricity far more accurately with access to price of natural gas (itself influenced by many geopolitical factors), breakdown in composition of production of each energy source, predictions about roll-out (requiring a sophisticated LLM capable of understanding news and their implications), etc. What's the use in trying to predict it from 2 variables and a narrow window looking back at just 8192 or 32768 samples?
Or perhaps I misunderstand the use-case and this is more for extremely short term movements?
You're right but, if the time series looks like music then you can apply your music domain knowledge. You don't have to only apply knowledge that's general across every domain, part of the magic of large neural nets is that you can do recognition and prediction at the same time thereby using (statistically) things known from training about how to identify a sequence and then also how to continue it.
Sure adding outside info helps even more, but with enough data it should become unnecessary (or so goes the theory).
Would love to know actual time series foundation model use cases!
Is there a reason you don't throw in something like cross-attention on an embedding of the sequence description/data labels? I've trained numerous generative models for audio and images and it's always the case that the unconditional model performs worse, trains slower, etc. than the conditional one (with the best performance from a hybrid using classifier-free guidance or similar scheme).
Again I don't have experience with the domain of time-series forecasting directly, but I would think there are generalities between autoregressive generative models and time-series forecasting, given that both involve predicting the next sample(s) given N regularly sampled past inputs.
Won't you need to attribute a weight to the data coming from the model and the weight for your data?
I'm sure someone who actually knows this stuff has a better easy to view it but my impression was that these time series foundation models aren't trained for a specific dataset but more they are trained on a procedure for how to look at datasets and make predictions. It doesn't seem to lump things into a bucket much more than any other model type (which also are just consistent procedures to look at datasets). It's not like you have radically different model types across different fields.
To improve the robustness of machine learning models, artificial noise is often introduced into actual training data. Something like Timestral could simplify the process of generating it.
That seems like a scenario where there wouldn't be external sources of information capable of improving your predictions. It also seemed to me like a scenario where good old fashioned statistical methods would outperform LSTMs, which was what they used.
(I didn't get the job - I was offered it, but they rescinded their offer at the last moment, supposedly because their investors had insisted they should hire a web developer instead.)
Sure, you could train a XGboost model on a time series with holiday or event covariates, but what if you have 10k different series to forecast for? You could train 10k different models, one for each series, but then they have very little history to learn from. Alternatively, you could train a single generic model to cover all 10k, but then you have to deal with scale mismatch between series in addition to the feature dev overhead between series. A TSFM saves a ton of work in modeling each of these series and could potentially generalize better to produce more accurate forecasts.
who has these though? i cannot imagine somoone making sense of 10k timeseries.
lol. so what?
Edit : working with GLM5.3 now, it's also slow. Maybe Mistral's under heavy load.
Disclaimer : Not Cofounder of The Forecasting Company here
Something is going on with the performance (as in speed of the service, not quality). Some days it's quite fast, some days it's so slow I interrupt it because I thought it was hanging.
I think they're shifting hosting capacity around, or otherwise messing with the model. Note that it's still a "preview" version, meaning they're still training it further for the final release and they are likely still figuring out the optimal hosting configuration as well.
It wouldn't surprise me to see further changes in both quality of the model and speed before the final release.
I agree with you though. It's slow as hell, and so is their GLM offering.
Speed looks nice too. I want an open weight model with this performance now.
If you compare the 2, it's not a strict "timestral is better" story - the winrates are tight - your probably should try both on your use-case. We have a 42% winrate vs Timestral on fevbench for instance.
Ofc, our main value is not just these models, but also 1) extra capabilities (like our model that can natively account for stockouts/censored data, and others coming soon), 2) our systems and products around the model, and 3) we're all laser focused on time-series/forecasting
Seems unprofessional to me, you already started dunking on this release but added little to no arguments as to why or what is wrong here.
https://www.theforecastingcompany.com/blog/t0-gamma/
https://github.com/Nixtla/nixtla
No need to pay for proprietary TimeGPT.
Calling it "Modified" MIT is just misleading.
Which is why at the end of the day why nobody trusts or blames the forecasting models, specially in all or nothing types of decisions.
Or the poll-bashing that happened after the 2016 election when an outcome with a 29% chance per one model was observed.
Any interesting forecasting problem is underconstrained. Monte Carlo methods are the classic way to mitigate that one-universe problem.
Nobody can predict the future with certainty. All models are wrong, but some are useful.
I fully agree that all models are wrong.
Their main usage is just replaying the past or encoding the modelers assumptions for the future.
But there is no way to know if a model is good or bad.
There would be if we had infinite universes.
Spacetimeformer was imo a huge leap forward that got overshadowed by all the language model progress
Excited to see more effort put into time series forecasting
Honestly this might be very good for feature selection to refine a much more specialized model.
With such strong human capital, both in quantity and quality, Europe should in theory be competitive with US and China in AI.
"Mistral is so bad, it's the only AI my school doesn't bother to block"