Time series forecasting is central to many analyses, whether predicting system behavior or implementing predictive maintenance to detect and anticipate anomalies. However, predicting future values remains a complex task, practically impossible in certain fields. Today, statistical approaches and machine learning methods are widely used, offering relatively reliable forecasts. Yet, given the complexity of many scenarios—especially in industrial settings—statistical methods quickly reach their limits, while machine learning techniques often do not fully exploit their potential, primarily due to high costs associated with acquiring and processing necessary training data.

In recent years, foundation models have become popular, particularly in natural language processing (NLP) and computer vision, exemplified by models such as GPT and DALL-E. These large-scale models, trained on vast datasets, are designed to adapt to a wide variety of tasks. Foundation models are now also emerging in the field of time series.

The FoMoTSF project, which stands for “Foundation Models for Time Series Forecasting,” initially aims to evaluate the performance of foundation models applied to time series forecasting, particularly within specific industrial use cases. To this end, a comprehensive open-source library named onTime has already been developed, enabling quick, reliable, and universal benchmarking of forecasting models on both public and private datasets. To date, this library has facilitated the evaluation and comparison of two leading foundation models alongside other benchmark methods, whether statistical or machine learning-based. Moreover, additional foundation models are currently being integrated and evaluated.

The second goal of the FoMoTSF project is to enhance the best-performing and most suitable existing foundation model to apply it effectively to real industrial scenarios. Achieving this objective would demonstrate the significant advantages of employing such models in industrial applications, where the demand for reliable, cost-effective forecasts is particularly high.