A Novel Method for Time Series Prediction with Small Data: Integrating Data Augmentation, Normalization Techniques, and Machine Learning

Abstract In recent years, machine learning techniques for time series prediction have gained significant traction due to their effectiveness and versatility. However, these algorithms often require a minimum amount of data to ac hieve optimal performance and avoid overfitting. In many real-world applications, such data is scarce or difficult to obtain, creating challenges for training accurate models. This limitation highlights the importance of developing methods that can work effectively with small datasets. To address this issue, we propose AIDAN, a novel approach that integrates artificial intelligence, data augmentation, and normalization techniques to enhance predictions for small, low-frequency time series. AIDAN employs transformations to diversify data samples while preserving their essential characteristics and explores the impact of normalization to stabilize data for training machine learning algorithms. We evaluated the performance of our approach by varying the prediction horizon across multiple domains, including environmental and socioeconomic data, for the ten largest global economies using heterogeneous datasets. Our findings revealed that AIDAN outperformed both a set of traditional time series prediction methods (ARIMA, ARIMA-GARCH, and Exponential Smoothing) and a simple machine learning ensemble (Jitter + diff + MLP), achieving between 6% and 9% higher accuracy for short-term predictions and approximately 15% for long-term predictions. This research provides an effective framework for advanced machine learning-driven time series prediction in low-frequency and data-scarce scenarios.