Research

Deep neural network surrogates approximate structural finance models at a fraction of the computational cost. Applied to option pricing, they enable frequent model estimation, a measure of tail risk that predicts market crashes, and tests of parameter stability and option return distributions.

We measure how much financial information large language models have memorized and how that memory can bias historical tests. The bias is strongest for low-frequency and aggregate data; identifying affected observations helps preserve reliable backtests.

We separate unexpected news from the component predictable from stock characteristics. These news shocks predict returns more strongly than raw news, with effects lasting up to 18 months, and reveal both investor underreaction and overreaction across news topics.

We examine whether asset returns are better described by a few factors or by many. Large, complex factor models perform better out of sample than simpler models and capture greater sensitivity to long-run macroeconomic activity.

Using earnings calls and financial news, we find that adding context helps large language models only up to a point. Beyond an optimal length, extra information reduces predictive accuracy; larger models can process more context before this decline.

We identify interpretable concepts inside large language models and measure how they shape economic forecasts. This approach can also steer optimism, pessimism, and risk aversion while keeping other model behavior fixed.

We use the probabilities a language model assigns to its next tokens to measure uncertainty in financial predictions. Higher internal confidence predicts more accurate news classifications and stronger portfolio performance than self-reported confidence.

Characteristic correlations help explain how characteristic portfolios move together, predict returns, and respond to the business cycle. Stabilizing factor exposures makes low-variance components useful for estimating the stochastic discount factor and produces out-of-sample performance comparable to leading machine-learning models.

Using investor-level holdings, we show that retail investors tend to trade stocks and cryptocurrencies in the same direction. This shared demand helps explain the rising correlation between the two markets, especially for stocks favored by crypto traders.

Hedge funds run by managers with similar elite educational backgrounds show unusually correlated returns. Evidence surrounding the Galleon insider trading scandal points to shared information networks, which investors appear to value.

Deep Regression Ensembles combine random input weights with ridge regression fitted layer by layer. Across benchmark datasets, the architecture matches or exceeds many standard deep networks while substantially reducing computation.

We characterize an optimal encoder-decoder representation that compresses data into the dimensions most useful for learning. Removing harmful dimensions can improve model performance and make learning more stable under distribution shift.

Using deep replication, we estimate market makers’ risk aversion from individual S&P 500 option contracts. The estimates form a smile across option moneyness and show a higher premium for puts after 2008, consistent with inventory and unhedgeable risk.