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.
4,510 documents
This notebook trains Skip-gram Word2Vec on labeled financial news sentences and explains how its vectors represent words that occur in similar contexts. It describes the roles of vector size, context window, minimum frequency, prediction mode, negative…
This notebook demonstrates fine-tuning three transformer checkpoints for three-class financial sentence sentiment and evaluating them with accuracy, macro F1, and confusion matrices. It uses a stratified train, validation, and test split so class imbalance…
This notebook explains how to design a search tool for a forecasting agent so evidence has a consistent structure, a traceable origin, and an auditable path into the model. A shared client protocol returns typed results across providers and includes a cutoff…
This notebook evaluates four news-derived signals—weighted surprise, average sentiment, sentiment change, and article coverage—against forward stock returns. It computes a daily cross-sectional Spearman information coefficient, summarizes its mean,…
This notebook evaluates four signals derived from news text: weighted surprise, average sentiment, sentiment momentum, and article coverage. It uses forward returns prepared by an earlier feature-building step, then calculates a daily cross-sectional…
This notebook compares Polymarket’s crypto-settled event contracts with the regulated Kalshi venue as sources of alternative data. It explains how access rules, settlement assets, position limits, and listing policies shape which participants influence…
This notebook presents a multi-round forecasting debate between bull and bear roles. Each side argues for a higher or lower probability of an event, sees the other side’s prior argument in later rounds, and reports a probability and supporting evidence. The…
This notebook implements an ESG headline workflow that selects news with keywords, assigns each selected headline to an environmental, social, or governance category, and scores sampled headlines with FinBERT sentiment. It measures the selected pool’s…
This notebook applies token-level SHAP to a pinned FinBERT model that classifies financial text as negative, neutral, or positive. It wraps model inference to preserve the checkpoint’s label order, then explains class probabilities by measuring how masked…
This record documents a multi-agent pipeline that estimates whether the United States will enter a recession by the end of 2026. Agents search for economic indicators and forecasts, form individual probabilities, compare rationales through debate, and pass…
This notebook runs three otherwise identical research agents on a recession question, either by replaying a saved capture or by running live models and search in parallel. It records each agent’s probability, confidence, evidence-gathering activity, and…
This notebook turns financial headlines into stock-level signals and evaluates them against forward returns. It embeds headlines, measures news surprise as semantic distance from a rolling embedding baseline, and combines surprise with sentiment direction to…
This notebook runs or replays a panel of identical research agents answering a shared forecasting question, then inspects their forecasts, search activity, and conversation traces. Separate per-agent tracing preserves attribution during parallel calls. The…
This notebook evaluates a pretrained financial sentiment model on FinMarBa headlines and explains why the result is not a clean test of sentiment transfer. FinBERT was trained against human judgments of text sentiment, while FinMarBa’s negative, neutral, and…
This notebook presents a pipeline for turning financial headlines into stock-level signals. It cleans and standardizes a news corpus, removes exact and prefix-matched duplicates within ticker-date groups, and uses sentence embeddings to represent headline…
This notebook builds a forecasting agent that alternates between web searches and probability forecasts. A small two-method language-model interface lets the same loop use a mock, local, or commercial provider. The agent receives a prediction-market question…
This notebook explains how to extract insider transactions from raw SEC Form 4 XML for quantitative equity research. It uses an XML parser to keep each trade’s code, date, share count, price, and direction attached to its transaction block, while separately…
This notebook applies pretrained FinBERT to financial headlines from FinMarBa and examines why its accuracy is far below figures reported for its Financial PhraseBank evaluation. The central diagnostic is label provenance: PhraseBank labels reflect…
This document outlines a simple visual workflow that retrieves a Fear and Greed classification, routes the result through conditional branches, and sends trade instructions to an execution node. A scheduled trigger starts the process; the workflow fetches…
This workflow proposes using holders of a successful token as a starting group for discovering other assets they own. It filters holders by token ownership share and portfolio value, removes addresses labeled as exchanges, project treasuries, or protocol…
This workflow describes a bidirectional grid for cryptocurrency perpetual futures. From an initial reference price, it places long levels below and short levels above; positions are opened as price reaches grid steps and are intended to close when price…
The strategy uses an exchange’s Bitcoin futures short-position series as a sentiment input and applies RSI to that series to identify turning points. It proposes short exposure when the RSI of short positions crosses above a high threshold and long exposure…
This short-term futures approach monitors Binance liquidation data, filters small events, and compares recent liquidation volume with its historical average using a Z-score. A surge with a strongly one-sided liquidation mix creates a candidate signal. The…
This workflow describes an automated crypto perpetual-futures system that combines short- and long-term EMA crossovers with news sentiment. It evaluates completed candles to identify golden or death crosses, then gathers recent RSS coverage and has an AI…