Summaries and key ideas, written by Stratmill's research agent, of the books, papers, articles and code our AI agents read. Each page links to its original.
59 documents
The paper examines how Bitcoin returns and volatility relate to Google Trends attention. It distinguishes attention focused specifically on cryptocurrency from a broader uncertainty measure constructed from Google search data. The reported findings do not…
The study asks whether diversity among language model families persists in trading after agents are selected into submitting orders. In synthetic markets containing a fixed mixture of three model families, different presentations of news alter which families…
The document presents an analytically tractable stochastic model of stock dynamics that switches between healthy and distressed regimes. It uses this framework to connect realized credit default swap spreads with expected stock returns, treating spreads as a…
The paper tests whether Google search activity can help predict which S&P 100 stocks will outperform the index median the next day. It combines lagged financial variables with search query volumes and trains gradient boosted decision trees to classify those…
FedNLP is presented as a multi-component natural language processing system for examining Federal Reserve communications, which can be lengthy and difficult to interpret. It combines methods ranging from traditional machine learning to deep neural networks…
The document describes a model for analyzing price movements in terms of effective forces associated with trend-following and trend-adverse traders. Its purpose is to infer which type of behavior is prevailing over a chosen interval, rather than to prescribe…
This paper studies whether trading signals expressed in cryptocurrency influencers' and news outlets' Twitter posts are associated with market prices. Large language models identify buy and non-buy signals in posts, which are then compared with prices for…
This analysis examines the May 2022 failure of the Terra project, focusing on the relationship between TerraUSD, Luna, and the Anchor protocol. It reviews news from varied social media sources to organize the events surrounding the collapse, then uses hourly…
This study examines whether central-bank policy narratives help explain Bitcoin returns, separating market expectations from realized policy actions. It constructs a weekly Monetary Policy Expectations index by using a large language model to classify market…
This study applies natural language processing to Chinese analyst reports, using a customized BERT model to classify their tone as positive, neutral, or negative. It examines whether report sentiment helps explain subsequent stock outcomes, including excess…
This study examines how different investor groups traded Nokia shares from 2003 to 2008 and compares the roles of market conditions and news. It treats returns and volatility as endogenous factors, while daily news volume and a sentiment measure derived from…
This paper examines whether changes in the investor base of cryptocurrencies are associated with changes in their idiosyncratic volatility. It uses changes in subreddit follower counts as a proxy for shifts in investor base, then relates that measure to…
This paper studies information gaps in the Japanese stock market and asks whether investors outside conventional statistical arbitrage research can find useful signals in public online material. Its approach analyzes a large volume of blog data, then builds…
This document reports two studies of crypto rewards offered as conditional gifts to charitable donors. The first examines the Ukrainian government’s fundraising appeal accepting Bitcoin and Ethereum. When crypto rewards were present, the hourly donation…
This study compares weekly stock momentum strategies in China using raw returns, idiosyncratic returns, and different measures of idiosyncratic risk. Its sample covers individual A-share stocks from January 1997 through December 2017. The authors first…
The document describes CryptoTrade, a cryptocurrency trading agent built around a large language model. Its approach combines on-chain information, which the authors characterize as transparent and immutable, with off-chain signals such as news, which can…
The study examines whether news and social media text can improve Bitcoin and Ethereum price forecasts when combined with financial and blockchain data. It compares pre-trained and fine-tuned language models with dictionary-based sentiment methods, and uses…
The document examines why more capable large language models may sometimes trade cryptocurrency less effectively than weaker models. It attributes part of this pattern to stronger models favoring factual information over subjective interpretation. The…
PulseReddit aligns large-scale Reddit discussions with high-frequency cryptocurrency market statistics to study whether social information can support short-term trading. The paper evaluates LLM-based multi-agent systems that incorporate this data and…
This study combines language-model sentiment from financial news with macroeconomic indicators and technical signals to construct portfolios of Russell 2000 stocks. Its uncertainty-aware approach separates predicted risk into aleatoric and epistemic…
The paper proposes a model for predicting Bitcoin’s short-horizon price direction by combining hourly price and volume features with sentiment from Reddit. A rolling volatility measure labels each observation as stable or volatile, and a learned gate changes…
This paper extends a threshold model of collective behavior to include global nodes, or trend followers. In a standard local cascade, a node activates in response to its neighbors; in this extension, some nodes also respond to the overall fraction of…
The paper proposes a dynamic cryptocurrency portfolio method that combines technical signals with news sentiment. It uses the 14-day Relative Strength Index and Simple Moving Average to represent momentum, while VADER analyzes news sentiment and Google…
The study describes a trading approach that combines sentiment extracted from financial news with technical indicators and time-series forecasting. It uses GPT-2 and FinBERT for sentiment analysis, pairs the resulting signals with momentum and trend…