What global models are and how they are transforming business forecasting
Global models are transforming the field of time series forecasting, giving companies the ability to generate accurate forecasts with limited data. Inspired by advances in artificial intelligence, these foundation models are trained on large volumes of data and can extrapolate their predictions to new scenarios, making them a game-changer in key sectors.
By Alberto Llera (PhD)
What foundation models are
Time series forecasting is crucial for decision-making in sectors such as energy, transportation, and healthcare. Traditionally, statistical models such as ARIMA or ETS have dominated this field. Although neural networks, such as recurrent neural networks (RNNs) and LSTM networks, were designed for this type of task, their practical use in traditional train/test environments has been limited.
However, the arrival of foundation models has transformed this landscape. Inspired by advances in language models, such as those underlying technologies like ChatGPT or DALL·E, these models are trained on vast amounts of unsupervised data and are then able to extrapolate their learning to new tasks. Similarly, global models for time series forecasting work by training on enormous amounts of data from different time series, enabling them to make predictions on series they have never seen before.
How global models are used at LISDATA
At LISDATA, we have adopted these global models as part of our strategy to optimize forecasting in environments with limited data. The speed at which this field is evolving is impressive. Just a decade ago, the idea of using global models for time series forecasting seemed unfeasible, but today it is already an emerging reality.
In our R&D team, we conduct rigorous testing with different models and have observed that, in some cases, foundation models already outperform traditional train/test approaches. This is especially relevant for industrial sectors that lack sufficient historical data. Thanks to transfer learning, these models can be trained with a minimal dataset, and as more data is collected, the models progressively improve.
Practical applications of global models
From an industrial perspective, working with global models offers a significant competitive advantage: we can implement predictive systems from scratch, optimizing their performance over time without the need for large volumes of initial data. This not only accelerates companies’ ability to predict future behavior, but also allows organizations in early stages of digitalization to harness the power of artificial intelligence immediately.
In projects such as InnoTwin, we have explored the potential of global models in different contexts, such as predicting product quality in manufacturing processes. This project allowed us to apply techniques such as transfer learning to create a single model capable of predicting the quality of previously unseen products. This not only reduces development time but also optimizes resource utilization, since a single model can adapt to multiple products and tasks.
How we bring R&D to businesses
At LISDATA, we are immersed in developing our own foundation models for scenarios where data is scarce. Imagine a poorly digitalized industry where we have just started installing sensors. Traditionally, we would have had to wait months to collect enough data to generate reliable forecasts. With foundation models, however, we can start making predictions from the very first moment, and as data accumulates, we can fine-tune the model to improve its accuracy.
This approach is transforming the way companies can approach forecasting and decision-making. Every day we move closer to having models that, in addition to learning from large amounts of data, efficiently transfer that knowledge to new tasks. Global models are set to become a key part of business digitalization, helping companies harness the power of artificial intelligence even when initial data is limited.
What the future holds for global models
In the coming years, we will clearly see a significant increase in the use of global models, both in academic research and in industry applications. The ability to make predictions without extensive or specific training, and to continuously improve as more data is collected, makes these models a powerful tool for any company looking to stay competitive in the digital era.