Generative AI in the time series domain

In recent months, we have witnessed how generative artificial intelligence has become remarkably integrated into our daily lives.

This phenomenon has been driven by the introduction of innovative tools such as ChatGPT for text generation, Dall-E for image creation or Copilot for code generation.

A further step has recently been taken by applying the knowledge of generative artificial intelligence to the field of time series. Given the deep-rooted human need to anticipate possible future events, from understanding economic cycles in stock prices to predicting demand patterns in retail stores or optimizing electricity consumption in a building, the opportunity to apply generative artificial intelligence to the time series domain presents itself.

The Revolution of Generative Artificial Intelligence in Time Series can bring great benefits in several domains.

Evolution of Time Series analysis and prediction techniques

Historically, traditional statistical methods have been the dominant approach to time series analysis and forecasting. In the last decade, Machine Learning methods have gained popularity, showing promising results. And in recent years, the introduction of deep learning methods is generating controversy, however, recent studies are yielding encouraging advances.

Work is currently underway in the field of generative artificial intelligence based on time series, investigating the benefits of foundational models. These developments could potentially open a new chapter in this field, fostering a deeper understanding of temporal data, improving forecasting efficiency and extrapolation to various domains.

Recently, the first TimeGPT-1 foundational model capable of predicting time series in a wide variety of domains and applications without additional training has been implemented.

History of analytical methods used in Time Series

So, what can Generative AI offer us that traditional methods fail to provide in the world of time series?

Generative AI is based on the implementation of pre-trained models with a large set of time series from various domains, such as finance, web traffic, iot, weather, demand, among others. These models have the ability to generate accurate predictions for unobserved data sets during training.

These are some of the benefits of Generative Time Series AI:

• Modeling nonlinear complexities: Generative AI can capture nonlinear and complex patterns in temporal data, overcoming the limitations of traditional approaches.
• Realistic data generation: It can generate high-quality, realistic temporal data, facilitating the creation of more diverse and representative training sets.
• Extrapolation to new domains: The extrapolation capability of Generative AI enables accurate predictions even in domains not seen during training, offering flexibility in application to different contexts.
• Knowledge transfer: By using pre-trained models on massive temporal datasets, Generative AI can transfer knowledge across domains, improving generalization capability to new datasets.
• Large-scale data handling: By leveraging fundamental models, Generative AI can handle large volumes of temporal data efficiently, causing improved performance in scenarios with large datasets.
• Adaptability to temporal changes: Generative models can dynamically adapt to changes in temporal trends, making them more resilient and flexible to variations in data over time.
• Improved forecasting accuracy and efficiency: Research by Azul Garza and Max Mergenthaler-Canseco compares the most popular metrics in time series forecasting (RMAE and RMSE) with various methods, including traditional statistics, machine learning, deep learning and the foundational TimeGPT model. Results indicate optimal performance for high granularity (monthly and weekly) in the case of the foundational model.

✅ It is recommended to read the original paper: TimeGPT-1 de Azul Garza and Max Mergenthaler-Canseco.

What are the disadvantages of Generative AI over traditional Time Series methods?

Despite the significant advantages, there are some drawbacks:

• Significant investment in computational resources is required.
• It decreases transparency since this type of models are less interpretable and difficult to explain at a high level.
• There is a risk of overfitting.
• Longer computational time for development and implementation.
• It is necessary to have massive data to achieve optimal performance.

There is no doubt that this is a field to be explored and that we will soon have new news in this area.

Time series generation is here to stay, but there is still a way to go. We will be patient.

AI of Things

Data visualization: Choosing the right chart

December 13, 2023