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.
715 documents
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 explains why a strategy’s observed Sharpe ratio needs adjustment when it was selected from many trials. It presents the Deflated Sharpe Ratio (DSR), which compares the selected result with a benchmark that rises with the number and dispersion…
This notebook explains how to fine tune a transformer for financial named entity recognition, which identifies spans such as organizations, people, amounts, dates, and percentages in text. It focuses on BIO boundary labels and the alignment problem between…
This notebook adapts a supervised autoencoder to daily US equities. The model reconstructs ranked price and volume characteristics while auxiliary and main classification heads train a shared latent representation to predict whether returns are positive…
This notebook demonstrates a data-quality workflow for daily US equity OHLCV data. Structural checks cover nulls, duplicates, chronological order, price consistency, negative values, stale prices, and extreme returns. It then distinguishes structural…
This notebook presents a workflow for reducing a large set of ETF features to candidates for later modeling. It ranks features using cross-sectional Spearman information coefficients against forward returns, computed by date and then averaged. Newey–West…
This notebook explains how a stochastic discount factor (SDF) prices a cross-section of equity option signals. Unlike factor models that estimate exposures and factor returns, the SDF approach estimates one random variable that makes asset returns price…
This document introduces latent factors as common return patterns across stocks, with each stock's loading describing its exposure to a pattern. It contrasts principal component analysis, which extracts factors and stock loadings from the return panel alone,…
This document explains how to construct forward-return and direction labels for NASDAQ-100 intraday research. It defines the return using an explicit execution convention: observe a bar close, enter at the next bar, and measure the outcome over a configured…
This notebook studies TabM, a neural network design for tabular data, on monthly firm characteristics. It compares three presets that increase hidden width and ensemble membership while keeping other training settings fixed. The shared runner fixes the…
This chapter outlines post-deployment governance for machine-learning trading systems. It separates technical failures, where identical inputs produce different outputs, from statistical decay, where outputs no longer predict returns. Its framework combines…
This index introduces latent factors as shared patterns extracted from the way stocks’ returns move together, with each stock’s loading indicating its exposure to a pattern. It contrasts principal component analysis, which estimates factors and stock…
This analysis explains how to read information coefficients for a catalog of CME futures prediction models. It defines each date’s IC as the rank correlation between predicted and realized returns across products, then averages those daily values across…
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 brings an ETF case study’s registered backtests together into an uncertainty-aware strategy assessment. It reads performance measures with block-bootstrap confidence intervals and uses paired bootstraps to compare successive pipeline stages,…
This notebook explains how a macro panel stamped with the period it measures can leak future information into a trading backtest. It estimates publication dates by adding the period length to the stamped date and then applying a conservative release lag…
This notebook uses TreeSHAP to explain LightGBM return predictions, with global feature importance, individual prediction explanations, dependence plots, and pairwise interaction values. It compares SHAP rankings with permutation importance and tree-based…
The notebook demonstrates a pipeline for extracting supplier, customer, and competitor relationships from company annual filings and storing them as a knowledge graph. A local language model generates candidate subject–relationship–entity triples from filing…
The document explains how to create currency-pair features from three fitted time-series models: a state-space filter that estimates a slowly changing price level, a two-state hidden Markov model for dollar volatility regimes, and an ARIMA model whose…
The document presents a shared state machine for automated trading halts, with closed, open, and half-open states. It applies that lifecycle to four conditions: portfolio drawdown, daily loss, consecutive losing sessions, and system latency. After a recovery…
This notebook asks whether sequence models can improve on hand-built summaries of perpetual funding premium history. It compares NLinear, a simple linear model that reads an ordered window, with a two-layer LSTM. Both use the same feature order, walk-forward…
This document explains why stop-losses, trailing stops, and time exits cannot be evaluated in a case study whose backtest holds weights across a month and observes only the realized monthly forward return. Such rules depend on the price path between entry…
This document describes a strategy assessment process for registered NASDAQ-100 microstructure backtests. It reads existing runs rather than training models or rerunning backtests, traces selected strategy lineages, compares candidates with an equal-weight…
This document presents a staged method for matching company names from alternative data, filings, news, and price sources to securities. It first joins on available identifiers in a deliberate trust order, preserving unmatched records and checking that…