This paper investigates whether Large Language Models (LLMs) adapted for time series forecasting can predict financial market returns, which are often considered nearly random. The authors use the Chronos model, a pretrained time-series foundation model with 11 million parameters, and apply it to daily residual returns of the largest U.S. stocks from 1978 to 2016. Residual returns are extracted using three methods: IPCA, PCA, and Fama-French factors, following the dataset of Guijarro-Ordonnez et al. (2021). The study employs both zero-shot (no fine-tuning) and fine-tuned versions of Chronos, where fine-tuning involves daily retraining on the available data. Portfolio weights are determined by ranking predicted returns and taking long/short positions proportional to rank distance from the median, with volatility scaling to manage risk. The main finding is that Chronos, even without any financial training data, can generate significant alpha, with gross Sharpe ratios exceeding 1.3 across all residual datasets. Fine-tuning further improves performance, achieving Sharpe ratios above 2.0. These results surpass traditional benchmarks such as Short-Term Reversal (STR), autoARIMA, and a smaller CNN-Transformer model. The paper concludes that LLMs can identify inefficiencies in financial time series that are nearly indistinguishable from noise, offering a promising tool for quantitative trading strategies. The authors also note that fine-tuning requires careful control to avoid overfitting, and that the pretrained model's zero-shot performance is particularly convincing as it minimizes overfitting concerns.