TimeGPT: Generative AI for time series

Hello, everyone. Thank you. Thank you very much for, for being here. Um my name is Azul Garza, I'm CTO and co-founder of, of Nix Lab and I will be presenting Time GP T uh the first foundational models for time series forecasting.

Uh at Nixle, we are building a future for everybody. Uh we are a start up focused on time series research and deployment and basically, we are doing exactly the same thing that Open A I did for language, text and speech and Stability A I did for images and video, but for time and in particular for temporal data and this is really important because at the end of the day, businesses don't speak in images or languages they speak in time series. You can think for example, in sales iot data, um electricity demand finance among other use case and to solve the time series analysis problem, we are building foundational artificial intelligence to understand uh classify and predict time series.

And we do that using time series research, which is the systematic approach to unveil the the future and then research enables anticipating, for example, ocean tides or predicting closing values of the dow jones and time series are the dna of finance, commerce, energy, and, and technology. And at nixle, we we begin creating the most comprehensive open source time series ecosystem in the world. If you go to our github repo, you can see that we have different open source libraries specialized in particular ways to model time series. For example, we have statistical classical models, uh auto arima auto ets, the data model. But also we have open source libraries specialized in machine learning models, for example, like g bmx g boost. And of course, we have our neural forecast library specialized in uh deep learning models for time series forecasting.

And a lot of companies around the world uses our open source solution uh to produce accurate uh forecast for their use cases. For example, mozilla microsoft uh wayfair data bricks lyft ford, all of them uses uh the open source nic a software uh in in production. And uh we learned a lot uh in the open source space in the forecasting in the forecasting field.

Uh but one particular insight stood out and it's that forecasting in some cases and a lot of cases can be extremely hard, confusing and expensive. And this is you because uh if you want to deploy uh forecasting model into production, you need to collect the data, you need to clean the data, you need to perform hyper parameter optimization to select the right model for the right time series. And of course, you need to uh deploy that model into production and then you need to analyze the, the results of the, of the forecasting of the forecasting solution. So it's extremely hard to put uh forecasting pipelines into production.

But at nl we like, we really like big problems and also we love big solutions. That's why we are introducing time g 51 which is the first foundational models for a time series, time series data. We build time gp t one collecting uh the biggest amount of time series data in the world from a lot of different sources. For example, finance, electricity, web traffic, retail, tourism, iot healthcare, transport, economics among other sources. And then we trained a transformer based architecture with that data to produce time gp t in 10 gp t works exactly the same as cha gp t. But this is really important. We don't use uh a pre trained uh large language model. We are building an specific large temporal model for time series data and it works exactly the same uh in the sense that you just need to upload your data. You don't have to train the model on your own data. And immediately with a really low latency, you will have uh extremely accurate forecast in time gp t uh is more accurate and also is a lot faster orders of magnitude faster.

And of course, as chat gp t and other large language models, you can fine tune our large temporal model time gp t on your own enterprise data. And of course, with time gp t one, you don't need machine learning expertise. Time gp t one is simple. We exposed time gp t as an api. So basically you can get your forecast just using uh api calls. But also we released a python sdk. You can access time gp t uh using python. It's just as simple as just import the main class time gp t set up your token and then just solve the forecast method with your own data. And you will have forecast provided by, by, by time gp t.

And we are really excited about time gp t because with this solution, we think that the future is really for, for everybody. We launched the be eight weeks ago and the results are being really, really amazing, more than 5000 companies signed up and uh from a diverse set of sectors. For example, tech consultancies, retail, manufacturing, finance, energy, healthcare. Here are some logos of the, of the companies that want to try uh and have tried uh time gp t google for to seek my american express, nike coca cola and sleep.

And which is really interesting about time gp t. And um the people that want to use to use it is that the use cases are really broad. For example, people wants to, people want to use time you to demand forecast uh to inventory management, waste reduction anomaly detection, health care, iot predicted maintenance, finance. But also we got some unexpected things to, for example, if people um wants to predict circadian rhythms or for example, wrestling, wrestling audience, we released a paper and we compared time gp t against a lot of other other solutions, uh statistical models, machine learning models, deep learning models. But also we compared time gp t against the solutions of the big tech companies. In this case, we compared 10 gp t one against the solutions provided by google in particular vertex a i and bit one.

And we did an experiment um comparing tupd against vertex a i and bigquery us using new york city bike trips uh more than uh 300 times series. And these are the results. 10 gp d is in the first in the first row. Uh the next one is vertex a i, the deep learning model. Vertex a i behind the scenes uses a uh a transformer model. The difference with uh with vertex a i is that you need to train uh that model, it is not a pre trained model. And also we compared against big query, which is the statistical model. And as you can see uh in terms of ms e and may, which are uh the two metrics. Uh the, the two metrics most used in in in the field time upt uh beats uh vertex a i and, and bigquery.

The conclusions of this experiment is that time gp t is 5% more accurate than the, than vertex a i, 20% more accurate than uh big query. Uh time gp d only took two minutes to, to forecast the 300 times series against four hours of, of vertex a i. And this is really important because those four hours are being used only uh to train the model to train the model behind vertex a i. And the deployment, the, the most exciting part about time upt took less than hour, the than one hour against eight hours of, of vertex a i. And finally 10 upt only takes two lines of code uh to be deployed in, into production.

And that's why we are really, really excited about time gp t. And if you want to try time gp t on your own data, just go to, to that web page and submit your interest.

Um thank you very much.