The aim of this project is to use large language models to extract information from financial texts and make numerical predictions of key financial variables based on them. I am currently working on two separate projects in this area. In the first project, which was already part of my previous NAISS medium allocation, I used firms annual Form 10-K reports to predict stock returns at different horizons. Using interpretability techniques, I then analyze which information matters for which horizon. This allows me to understand which type of information from the corporate reports the market reacts to and how fast it reacts. This addresses one of the key questions of finance, namely how efficient markets are at processing different types of information. I have initial results for this project showing that the LLM can predict returns at the one-day horizon but have so far lacked the resources to test other horizons. In a second project, I use the transcripts of earnings calls, which are sessions where the top management of a firm answers questions of financial analysts regarding the firm's performance in the preceding quarter, to predict future firm earnings with an LLM. Several research papers in finance, including one of my own projects that was supported by my previous NAISS medium allocation, show that errors in the forecasts of financial analysts indicate mispricing and market inefficiency in financial markets. However, this work relies heavily on ex post realized forecast errors, which introduces several methodological issues. Instead, a clean approach would compare the forecasts of financial analysts to an unbiased statistical optimal forecast made at the same time. However, existing approaches that use standard machine learning methods on quantitative data fail to provide a useful benchmark because they offer only marginal improvements in accuracy over the analyst forecasts. I hypothesize that this is because they lack the qualitative information that the analysts have access to and aim to bridge this gap by combining a forecast generated with traditional machine learning methods with a forecast from an LLM that adds the qualitative information. Note that I am submitting this proposal with a start date two months before the end of my previous proposal because I no longer need the Dardell resources from the previous proposal but instead need significantly more Alvis resources. I have received a temporary increase of my Alvis resources for my previous allocation, but this will run out in February, after which I will no longer be able to reasonably use Alvis.