In briefAdapting a pretrained order-flow model to forecast very short-term prices
SummaryThe paper pretrains a 300M-parameter generative model on Level-3 order-flow data from the crypto exchanges Bitstamp and Bitfinex, then fine-tunes it to predict microprice returns and their realized volatility. The fine-tuned model outperforms classical baselines such as gradient boosting on handcrafted features over sub-second horizons. Its predictive ability also transfers zero-shot to assets it was not fine-tuned on.
In briefUsing tabular foundation models to forecast firms' earnings changes
SummaryThis note introduces tabular foundation models (TFMs) to accounting research and tests them on predicting firms' next-year earnings changes. The TFMs learn in context without post-release training or tuning, and are pretrained on synthetic data to avoid look-ahead bias. Google Research's TabFM is more accurate than all comparison models, including a fine-tuned large language model and several tree-based models, and is faster than the language model. Some tree-based models predict more quickly, so accuracy and speed involve a trade-off.
In briefMaking sense of Rule 605: analysis of monthly broker and market-maker execution reports
SummaryThis paper analyzes monthly reports from market maker Virtu and broker-dealer SoHo covering January 2021 to March 2024. Standard methods such as linear regression and LSTM show poor predictive power for the probability laws of stock orders, even though those laws are relatively stable. However, month-to-month correlations are highly stable and predictable. The authors study the distributions of sixteen datasets derived from 26 stock-trading parameters that brokerages report under regulatory requirements.
In briefChoosing Complexity: How Ability and Overconfidence Shape Task Selection
SummaryThis paper studies, through a theoretical model and experiments, how cognitive ability and overconfidence affect the choice between complex tasks that differ in their success-payment tradeoff. Individuals with higher cognitive ability tend to choose more complex tasks, and overconfident individuals do so as well, conditional on ability. About 75% of decisions pick the payoff-maximizing task, but errors mostly involve choosing tasks that are too complex. The authors suggest pairing task menus with information and feedback that help workers assess their ability and success probability.
In briefDynamic investment in team skills: theory and lab experiments
SummaryThis paper provides the first systematic analysis of dynamic investment in team-specific skill that builds up within a team. The authors develop a dynamic game framework in which such investment yields persistent benefits but carries risk, since those benefits depend on successful team coordination. Laboratory experiments show underinvestment relative to the efficient benchmark, yet investment responds strongly to incentives in line with theoretical predictions. Pre-registered traits, theory of mind and propensity to coordinate, predict both investment and team efficiency.
