This paper investigates the presence of self-attribution bias among mutual fund managers and evaluates its impacts on trading outcomes. I develop a novel GPT-based Natural Language Processing (NLP) architecture designed to extract attribution information from mutual funds' self-assessments of performance in their shareholder reports. On average, mutual fund managers exhibit a significant self-attribution bias—they are 40.6% more likely to attribute performance contributors versus performance detractors to internal factors. Funds displaying stronger self-attribution bias tend to engage in excessive trading and excessive risk-taking in the subsequent reporting period, which negatively impacts their performance. In addition, funds exhibit a higher self-attribution bias following higher performance, despite the fact that biased attribution only influences fund flows when funds perform poorly. Overall, these findings suggest that biased attribution likely stems from cognitive bias rather than strategic choices.