We examine how large language models (LLMs) interpret historical stock returns and price charts when prompted to forecast short-horizon returns. While individual stock returns tend to reverse, LLM forecasts overextrapolate trends. Simulations show that extrapolation is stronger for less persistent series, similar to humans, and difficult to eliminate. LLM return forecasts are overoptimistic yet understate extreme upside returns, resulting in confidence intervals that are too narrow. When information is presented in prices rather than returns, expectations become more pessimistic. The findings suggest LLM forecasts exhibit patterns similar to human-like behavioral biases that are context-dependent and resist correction through prompt engineering.