What people are not seeing about TimeGPT
This is the Three Data Point Thursday, making your business smarter with data & AI.
Want to share anything with me? Hit me up on Twitter @sbalnojan or Linkedin.
Let’s dive in!
"The old wheel turns, and the same spoke comes up. It's all been done before and will be again." (Sherlock Holmes, The Valley of Fear)
“...we believe in AI's transformative potential for temporal tabular data.” that sounds like a rather specific and strange mission, doesn’t it? And yet, those specific missions are sometimes the best.
The company behind this statement just announced a cool new generic machine learning model called timeGPT - the ChatGPT of time series data - kind of. More like v1, but still better than 0.1.
And yet, the responses to the announcement are mixed, negative, and dismissive.
Honestly, I have no idea why. TimeGPT is amazing news! But most people dismiss two important things:
It’s not just about time GPS; time series models are becoming available at large - especially generic ones will provide a huge advancement across machine learning.
Timed data is becoming more and more important; just because it wasn’t in the past doesn’t mean it won’t in the future.
Why time is becoming more important
There’s a secret to all data: all data was created in time. All data is time series data. Time is always a feature. It’s a natural attribute, unlike all others that need to be modeled.
There is even an international standard for it, unlike most other types of features like “Did the customer churn?”
That alone makes time series data so powerful.
But that’s not all there is. Data growth is accelerating, which means recent data is simply larger and more information-rich. If you discard time as a feature, you’re basically ignoring a very important piece of information - the information density of the data itself.
Even if you don’t believe in the information richness of recent data, adding time as a feature only means you’re letting your algorithms figure out whether it is true.
There’s no real downside to adding time as a feature into your models.
Except for the obvious facts, most skeptics will point out: time series machine learning is super hard and requires experts!
Why people are skeptical about time series ML models
Because time series data is so standardized, it paradoxically takes an expert to model it. The field has advanced so much in comparison to other specific fields. This means that time series models are different from all the other models! Because they do NOT need modeling.
This, in turn, however, means that if you have a dataset where time is only one feature, then you’ll end up with a dual approach. It means you’re going an extra round for this one because you’re modeling out all other features except time. Or you’re going to ignore the special nature of time and treat it as a simple feature, turn it into “1_day” and “2_day_ago” labels ready for consumption, just like “red” and “blue.”
On high dimensional data, if you’re modeling out 100 features, you’ll then be much more inclined to add time as a modeled feature instead of using an advanced time series model.
On mid-dimensional data, if you have 20-50 features, you will focus on reducing the computation and fine-tuning time - leaning towards fast-boosted models, still not leaning towards time series models.
On low dimensional data, if you have just 1-10 features, you’re going to end up doing a lot of hand modeling even for your time series model, thus leaning towards one of the specialized models, but probably not for a generic model.
So from a high perspective, there isn’t a single class in the problem space that requires a generic time series model.
And yet, if you look closely at those arguments, all of them will go away once time series models get just a little bit better and data growth accelerates a little bit more.
Since such developments don’t happen at the same time across the board, you’re likely already missing out on a few potent use cases.
Note: Check out the discussion on Hacker News to dive deeper into this perspective.
What is timeGPT, really?
TimeGPT is a new generic model (as a Service) provided by Nixtla, founded in 2021. With a $4,5m USD seed round this April, they focus on the time series problem in general.
They have already published a ton of open-source projects totaling 7,5k GitHub stars. They truly believe in the importance of temporal tabular data, as do I.
Time series analysis traditionally has been a field of super special models called RIMA, TBATS, and many other names. The key is to find a good model that fits your data, adjust it, and then use it to forecast.
The whole idea of a generic model is that you don’t need to do that. You don’t need to fit it into your data in a specific way. You might still want to fine-tune and do some RAG, as you do with LLMs, but you do NOT want to do specific modeling.
That’s what timeGPT is trying to do.
Why should this be interesting for you?
Forecasting is extremely difficult; as Max Mergenthaler explains at the launch of timeGPT, only a handful of companies in the world have the resources to do it right.
And yes, forecasting is always, and I mean always, about time. So basically, until we have an excellent generic timeGPT, almost everyone is excluded from using one of the most essential features in his data.
That sucks!
Where will timeGPT fall short for now?
TimeGPT will basically be like ChatGPT-1; it will be “interesting” and somewhat good, especially in unknown domains, but with a little bit of practice, most experts will be able to cook up something that’s better.
However, we can expect 2-3 generations from now on to be blown away by a time series foundation model, one that we want to use all the time for all time series forecasting. One that integrates well into other models and is usable out of the box.
As we’ve seen with other transformer-based models, 2-3 generations simply mean getting a ton of additional data, something that can be done in a matter of months.
I’m excited and am looking forward to the next versions next year.
Here are some special goodies for my readers
👉 The Data-Heavy Product Idea Checklist - Got a new product idea for a data-heavy product? Then, use this 26-point checklist to see whether it’s good!
The Practical Alternative Data Brief - The exec. Summary on (our perspective on) alternative data - filled with exercises to play around with.
The Alt Data Inspiration List - Dozens of ideas on alternative data sources to get you inspired!