I think that the concept of a "foundation model" for time series is actually a bit flawed as presented in this blog post. A foundation model is interesting because it is capable of many tasks _beyond the target tasks_ that it was trained to do, whereas what the author is looking for is a time-series model that can make out-of-distribution predictions without re-training - which is, in my opinion, a problem that is pretty well solved by existing ARIMA and (especially) Prophet models (Yes, you have to re-fit the model on your distribution, but this is not at all akin to the task of training or fine-tuning an LLM, it's something you can do in seconds on a modern CPU, and yes, there are certain hyperparameters that may need to be selected, but they are actually fairly minimal).
But for a model to make out-of-distribution predictions does not make it a foundation model for time series, really that's just the basic task that all time series forecasting models do. A more interesting question is, does an LLM architecture seem to improve the task of univariate or multivariate time-series prediction? I don't think the answer is yes, although, depending on your domain, being able to use language inputs to your model may have a positive impact, and the best way to incorporate language inputs is certainly to use a transformer architecture, but that isn't what is addressed in this post.
A lot of people try to hedge this kind of sober insight along with their personal economic goals to say all manner of unfalsifiable statements of adequate application in some context, but it is refreshing to try to deal with the issues separately and I think a lot of people miss the insufficiency compared to traditional methods in all cases that I've heard of so far.
Makridakis' conclusion remained true for many years: "statistically sophisticated and complex methods do not necessarily provide more accurate forecasts than simpler ones."
This looks like a great benchmark! We've been thinking of doing a better and more detailed follow-up and this seems like the perfect dataset to do that with. Thanks!
Look i'm optimistic about time-series foundation models too, but this post is hard to take seriously when the test is so flawed:
- Forward filling missing short periods of missing values. Why keep this in when you explictly mention this is not normal? Either remove it all or don't impute anything
- Claiming superiority over classic models and then not mentioning any in the results table
- Or let's not forget, the cardinal sin of using MAPE as an evaluation metric
Good to see positive reception to feedback! Sorry if my message came out as condescending, was not the intent. I recommend reading this piece on metrics https://openforecast.org/wp-content/uploads/2024/07/Svetunko.... It's easy to grasp, yet it contains great tips.
we're grateful for the honest feedback (and the awesome resource!), makes it easier to identify areas for improvement. Also, your point about using multiple metrics (based on use-cases, audience, etc) makes a lot of sense. Will incorporate this in our next experiment.
Short answer: i use multiple metrics, never rely on just 1 metric.
Long answer: Is the metric for people with subject-matter knowledge? Then (Weighted)RMSSE, or the MASE alternative for a median forecast. WRMSSE is is very nice, it can deal with zeroes, is scale-invariant and symmetrical in penalizing under/over-forecasting.
The above metrics are completely uninterpretable to people outside of the forecasting sphere though. For those cases i tend to just stick with raw errors; if a percentage metric is really necessary then a Weighted MAPE/RMSE, the weighing is still graspable for most, and it doesn't explode with zeroes.
I've also been exploring FVA (Forecast Value Added), compared against a second decent forecast. FVA is very intuitive, if your base-measures are reliable at least. Aside from that i always look at forecast plots. It's tedious but they often tell you a lot that gets lost in the numbers.
RMSLE i havent used much. From what i read it looks interesting, though more for very specific scenarios (many outliers, high variance, nonlinear data?)
MAPE can be a problem also if you have a problem where rare excursions are what you want to predict and the cost of missing an event is much higher than predicting a non-event. A model that just predicts no change would have very low MAPE because most of the time nothing happens. When the event happens, however, the error of predicting status quo ante is much worse than small baseline errors.
Thanks for the reply! I am outside the forecasting sphere.
RMSLE gives proportional error (so, scale-invariant) without MAPE's systematic under-prediction bias. It does require all-positive values, for the logarithm step.
> Our dataset consisted of Kubernetes pod metrics collected from a production retail checkout application.
That sums it up and it’s no surprise why Datadog’s toto model performed exceptionally well.
The results would have been much more useful had they opted for a heterogenous mix of data sets. I am thinking of census data and statistics, or financial forecasting (GDP, interest rates), or clinical trial drop-out rates etc. So many interesting problems out there.
At the moment our focus is on observability, hence the narrow scope of our dataset. A pretty good benchmark for observability seems to be Datadog's BOOM- https://huggingface.co/datasets/Datadog/BOOM
But for general purpose time-series forecasting, benchmarks mentioned in other comments like GIFT or M4 might come in handy. We might include them in the follow-up experiment.
I'd be curious what the results would be with the automated Autogluon fit/evals. I suspect given the results here, a weighted average model would likely win out.
Author here, we’re just getting started with these experiments and plan to apply them to more features on our roadmap. Future posts will be more detailed, based on the feedback we received here. Once we finish implementing these features, we’ll be happy to share the code and dataset.
I'm a bit confused by the results table. Were these models tested against the same dataset? Also, a visualization of the test data and forecasts would be helpful as well.
Based on the feedback, we could have done a much better job with these results (lessons for our next experiment). But yes, the models were tested against the same dataset which was aggregated over different granularities (1 minute, 1 hour, 1 day)
Author here, apologies for not making it clear on the post regarding the definition of Zero-shot time-series forecasting, but it's quite widely used and here's the definition of it "Zero-shot time-series forecasting is a framework for time-series prediction that does not require fine-tuning with specific time-series data to be predicted."
I think that the concept of a "foundation model" for time series is actually a bit flawed as presented in this blog post. A foundation model is interesting because it is capable of many tasks _beyond the target tasks_ that it was trained to do, whereas what the author is looking for is a time-series model that can make out-of-distribution predictions without re-training - which is, in my opinion, a problem that is pretty well solved by existing ARIMA and (especially) Prophet models (Yes, you have to re-fit the model on your distribution, but this is not at all akin to the task of training or fine-tuning an LLM, it's something you can do in seconds on a modern CPU, and yes, there are certain hyperparameters that may need to be selected, but they are actually fairly minimal).
But for a model to make out-of-distribution predictions does not make it a foundation model for time series, really that's just the basic task that all time series forecasting models do. A more interesting question is, does an LLM architecture seem to improve the task of univariate or multivariate time-series prediction? I don't think the answer is yes, although, depending on your domain, being able to use language inputs to your model may have a positive impact, and the best way to incorporate language inputs is certainly to use a transformer architecture, but that isn't what is addressed in this post.
A lot of people try to hedge this kind of sober insight along with their personal economic goals to say all manner of unfalsifiable statements of adequate application in some context, but it is refreshing to try to deal with the issues separately and I think a lot of people miss the insufficiency compared to traditional methods in all cases that I've heard of so far.
Ai slop
I wonder how this would perform on the M4 Makridakis competitions (time series competitions)
https://github.com/Mcompetitions/M4-methods
https://en.wikipedia.org/wiki/Makridakis_Competitions
Makridakis' conclusion remained true for many years: "statistically sophisticated and complex methods do not necessarily provide more accurate forecasts than simpler ones."
Maybe things have changed?
(side: Nixtla showed a simple ensemble outperforming Chronos, and the Chronos team responded, but there's some back and forth in the comments: https://www.linkedin.com/pulse/extended-comparison-chronos-a...)
This looks like a great benchmark! We've been thinking of doing a better and more detailed follow-up and this seems like the perfect dataset to do that with. Thanks!
When I worked in Demand prediction (multivariate), it was lgbm that was outperformong across the board.
Look i'm optimistic about time-series foundation models too, but this post is hard to take seriously when the test is so flawed:
- Forward filling missing short periods of missing values. Why keep this in when you explictly mention this is not normal? Either remove it all or don't impute anything
- Claiming superiority over classic models and then not mentioning any in the results table
- Or let's not forget, the cardinal sin of using MAPE as an evaluation metric
Author here, we're trying these out for the first time for our use-cases so these are great points for us to improve upon!
Good to see positive reception to feedback! Sorry if my message came out as condescending, was not the intent. I recommend reading this piece on metrics https://openforecast.org/wp-content/uploads/2024/07/Svetunko.... It's easy to grasp, yet it contains great tips.
we're grateful for the honest feedback (and the awesome resource!), makes it easier to identify areas for improvement. Also, your point about using multiple metrics (based on use-cases, audience, etc) makes a lot of sense. Will incorporate this in our next experiment.
To clarify, you'd prefer rmsle?
Short answer: i use multiple metrics, never rely on just 1 metric.
Long answer: Is the metric for people with subject-matter knowledge? Then (Weighted)RMSSE, or the MASE alternative for a median forecast. WRMSSE is is very nice, it can deal with zeroes, is scale-invariant and symmetrical in penalizing under/over-forecasting.
The above metrics are completely uninterpretable to people outside of the forecasting sphere though. For those cases i tend to just stick with raw errors; if a percentage metric is really necessary then a Weighted MAPE/RMSE, the weighing is still graspable for most, and it doesn't explode with zeroes.
I've also been exploring FVA (Forecast Value Added), compared against a second decent forecast. FVA is very intuitive, if your base-measures are reliable at least. Aside from that i always look at forecast plots. It's tedious but they often tell you a lot that gets lost in the numbers.
RMSLE i havent used much. From what i read it looks interesting, though more for very specific scenarios (many outliers, high variance, nonlinear data?)
MAPE can be a problem also if you have a problem where rare excursions are what you want to predict and the cost of missing an event is much higher than predicting a non-event. A model that just predicts no change would have very low MAPE because most of the time nothing happens. When the event happens, however, the error of predicting status quo ante is much worse than small baseline errors.
My reading of this situation is that MAPE would do the opposite. Means are skewed towards outliers.
Thanks for the reply! I am outside the forecasting sphere.
RMSLE gives proportional error (so, scale-invariant) without MAPE's systematic under-prediction bias. It does require all-positive values, for the logarithm step.
> Our dataset consisted of Kubernetes pod metrics collected from a production retail checkout application.
That sums it up and it’s no surprise why Datadog’s toto model performed exceptionally well.
The results would have been much more useful had they opted for a heterogenous mix of data sets. I am thinking of census data and statistics, or financial forecasting (GDP, interest rates), or clinical trial drop-out rates etc. So many interesting problems out there.
At the moment our focus is on observability, hence the narrow scope of our dataset. A pretty good benchmark for observability seems to be Datadog's BOOM- https://huggingface.co/datasets/Datadog/BOOM
But for general purpose time-series forecasting, benchmarks mentioned in other comments like GIFT or M4 might come in handy. We might include them in the follow-up experiment.
The GIFT Eval benchmark would be a good place to start: https://huggingface.co/spaces/Salesforce/GIFT-Eval
I'd be curious what the results would be with the automated Autogluon fit/evals. I suspect given the results here, a weighted average model would likely win out.
We'll definitely include it in our next experiment (shaping up to be quite big!)
Interesting, what are the usecases youre using the models for? Would like to know more on that, like anomaly detection
That's actually one of the use-cases that we set out to explore with these models. We'll release a head-to-head comparison soon!
That's the thing I'm most interested in out of these. Super interested to see what you find out.
Did you or do you plan to publish any of your code or data sets from this?
Author here, we’re just getting started with these experiments and plan to apply them to more features on our roadmap. Future posts will be more detailed, based on the feedback we received here. Once we finish implementing these features, we’ll be happy to share the code and dataset.
I'm a bit confused by the results table. Were these models tested against the same dataset? Also, a visualization of the test data and forecasts would be helpful as well.
Based on the feedback, we could have done a much better job with these results (lessons for our next experiment). But yes, the models were tested against the same dataset which was aggregated over different granularities (1 minute, 1 hour, 1 day)
[flagged]
[flagged]
Author here, apologies for not making it clear on the post regarding the definition of Zero-shot time-series forecasting, but it's quite widely used and here's the definition of it "Zero-shot time-series forecasting is a framework for time-series prediction that does not require fine-tuning with specific time-series data to be predicted."
Thanks for the reply!
Great read! Really interesting to see how these foundation models like Chronos and Toto are starting to perform well on real-world observability data.
[dead]