In this dissertation, I investigate three applications of machine learning in financial forecasting. Thefirst study investigates the best techniques for forecasting corporate earnings, and sheds light on what the most accurate earnings expectation is, and what the market expectation appears to be. We find consistent evidence that the best machine learning forecast outperforms analysts’ forecasts. However, the best machine expectation does not beat the analyst forecast by a meaningful amount in most cases, except for two distinct instances: (1) the earnings forecast is for small firms, and (2) the earnings forecast is for a longer horizon. Second, in cases where there are meaningful differences between analyst and machine expectations, earnings response coefficient (ERC) tests imply that investors’ expectations appear to be mostly aligned with the best machine forecast. In my second study, I investigate the best interest rate forecasting techniques, and show that existing techniques perform poorly compared to a simple forecast of zero change. In light of this, I propose a new interest rate forecast which focuses on removing the maturity risk premium from forward rates and demonstrate that this new approach outperforms for long horizon forecasts of interest rates. Given these findings, I decompose excess bond returns to show that the primary driver of excess bond returns for short holding periods is a bonds carry, while for long holding periods its the bonds maturity risk premium. This risk premium is plausibly invariant across both time and across the maturities of forward rates. In my third study, we propose and test the "Sticky Information Cost" (SIC) hypothesis to understand how investors acquire information in uncertain financial markets. SIC asserts that information processing costs for investors are influenced by a firm’s slow-changing information environment, closely linked to its fundamental uncertainty. Using direct measures for information processing costs and the return predictability of analysts’ biases as a proxy for information acquisition, we find opposite relationships between uncertainty and information acquisition when comparing across firms and over time. These results hold across various uncertainty measures and other earnings-related anomalies, supporting the SIC hypothesis while challenging existing theories.