Alireza Koochali

• Accurately anticipating future events is essential in decision-making domains such as healthcare, finance, and security, where managing risk is crucial. Probabilistic forecasting is key to understanding potential risks and their impacts. This thesis advances probabilistic forecasting through three primary contributions: (i) innovative forecasting models using Generative Adversarial Networks (GANs), (ii) GAN-based methods for time series data augmentation to address data scarcity, and (iii) a new, unbiased framework for evaluating generative models in time series domain. Key contributions include the introduction of ForGAN, a GAN-based model for probabilistic forecasting, VAEneu, a Variational Auto-Encoder-based probabilistic forecaster, and CRPS Loss, a loss function optimized for these models. CRPS-ForGAN, which integrates CRPS Loss with ForGAN, consistently outperforms baseline models in one-step ahead forecasting tasks across multiple datasets. For multi-step forecasting, auto-regression and attention-based seq2seq models are introduced, with proposed models leading in performance in an extensive experimental setting. These models also generate forecasts efficiently, suitable for real-time applications. To tackle data scarcity, the thesis proposes a GAN-based model for time series data augmentation, particularly for rare extreme events, improving forecasting accuracy by 8.6% in real-world wastewater management applications. Additionally, the Structured Noise Space GAN (SNS-GAN) is presented for class conditional time series generation, achieving up to a 64% improvement in synthetic data quality over baseline models in a time series data generation study. Finally, the thesis introduces the Fréchet Inception Time Distance (FITD) and InceptionTime Score (ITS) as novel metrics for evaluating generative models in time series. Extensive experimentation across diverse datasets confirms their effectiveness, setting a new standard for model evaluation in this field.