Fin-GAN is a Generative Adversarial Network (GAN) designed for probabilistic forecasting and classification of financial time series. The paper introduces a novel economics-driven loss function for the generator, which places GANs into a supervised learning setting and enables the production of full conditional probability distributions of price returns given historical data. This approach moves beyond traditional point estimates and allows for uncertainty estimates. Numerical experiments on equity data show that Fin-GAN outperforms classical supervised learning models like LSTMs and ARIMA in terms of Sharpe Ratios. The method also helps alleviate mode collapse, a common issue in GANs, and allows for the modeling of joint co-movements of different stocks. The paper compares Fin-GAN with other models, including LSTM, ARIMA, and ForGAN with binary cross-entropy loss, and demonstrates its effectiveness in forecasting daily stock ETF-excess returns and ETF raw returns. The Fin-GAN loss function is designed to improve Sharpe Ratio performance and incorporates terms that enhance classification and uncertainty estimation. The paper also discusses the implementation of Fin-GAN, including data preprocessing, training setup, and evaluation metrics. The results show that Fin-GAN achieves superior performance in terms of Sharpe Ratios and provides more accurate forecasts with uncertainty estimates.Fin-GAN is a Generative Adversarial Network (GAN) designed for probabilistic forecasting and classification of financial time series. The paper introduces a novel economics-driven loss function for the generator, which places GANs into a supervised learning setting and enables the production of full conditional probability distributions of price returns given historical data. This approach moves beyond traditional point estimates and allows for uncertainty estimates. Numerical experiments on equity data show that Fin-GAN outperforms classical supervised learning models like LSTMs and ARIMA in terms of Sharpe Ratios. The method also helps alleviate mode collapse, a common issue in GANs, and allows for the modeling of joint co-movements of different stocks. The paper compares Fin-GAN with other models, including LSTM, ARIMA, and ForGAN with binary cross-entropy loss, and demonstrates its effectiveness in forecasting daily stock ETF-excess returns and ETF raw returns. The Fin-GAN loss function is designed to improve Sharpe Ratio performance and incorporates terms that enhance classification and uncertainty estimation. The paper also discusses the implementation of Fin-GAN, including data preprocessing, training setup, and evaluation metrics. The results show that Fin-GAN achieves superior performance in terms of Sharpe Ratios and provides more accurate forecasts with uncertainty estimates.