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.
5,018 documents
This notebook traces ETF ranking signals through portfolio formation and a simplified backtest. Ridge regression and logistic classification are trained with walk-forward splits, then compared with momentum and equal-weight portfolios. At monthly rebalances,…
This notebook fits PatchTST, a transformer-style sequence model, as one member of a predeclared population of models for S&P 500 options research. It divides a lookback series into fixed-length patches, embeds them as tokens, and uses attention to relate…
This document defines two diagnostics for labels whose forward return windows overlap: pooled panel autocorrelation and effective sample size based on label uniqueness. Autocorrelation pairs observations only within the same entity and at the requested…
This notebook builds an LSTM allocator that maps sequences of ETF returns and rolling volatility features directly to long-only portfolio weights. A softmax output keeps weights nonnegative and fully invested. Instead of forecasting returns and feeding those…
This notebook compares three ways to classify financial news sentences as positive, neutral, or negative: TF-IDF word and word-pair features with logistic regression, averaged pretrained GloVe vectors with logistic regression, and a pretrained FinBERT…
This notebook constructs features from minute-level NASDAQ-100 data using three procedures: a rolling heterogeneous autoregressive regression for near-term variance forecasts, a Fourier transform to describe periodic structure in recent activity, and…
This notebook shows how to analyze a financial event graph without confusing when an event happened, when it became public, and when the graph pipeline extracted it. It loads timestamped 8-K relationships, measures disclosure and extraction delays, and…
The document introduces latent factors as shared return drivers inferred from co-movement across futures rather than supplied as named predictors. It compares two approaches fitted within each training fold. Principal component analysis (PCA) finds linear…
The document describes how to compare validation predictions from linear, gradient-boosting, tabular deep-learning, and sequence models for S&P 500 options. Before reporting diagnostics, it checks that each registered prediction population is complete…
The document explains how double machine learning (DML) estimates whether buyer-initiated volume share affects future returns after accounting for spread, short-term volatility, and trailing returns. It residualizes both treatment and outcome against those…
This notebook explains how feature stores keep training data and live model inputs aligned. It describes feature views through their entity key, event timestamp, time to live, and feature columns, then contrasts point-in-time joins for historical training…
This document explains a five-gate process for evaluating nine trading case studies, from signal quality through validation, transaction costs, holdout performance, and evidence quality. The gates are cumulative, so a strategy reaches the final stage only if…
This notebook walks through a deployment rehearsal for a crypto funding-rate direction model. It trains a three-class LightGBM model on historical Binance-derived perpetual data, then fetches live hourly bars and funding rates from OKX, aggregates data to…
This notebook describes fitting a long short-term memory network to one-minute NASDAQ-100 microstructure features to predict forward returns at several horizons. The model processes a trailing sequence step by step, learning which earlier observations to…
The notebook introduces a conditional autoencoder that extends instrumented principal component analysis by replacing its linear mapping from stock characteristics to factor exposures with a neural network. It retains the exposure-times-factor-return…
This utility module builds tables and figures for evaluating predictive models across validation and holdout data. Its diagnostics include confidence intervals around information coefficients, validation-to-holdout decay, selection-adjusted backtest…
This document describes the architecture of a neural policy network that maps asset-level time series and context into bounded portfolio risk weights. Its per-asset backbone combines context embeddings, feature-wise modulation, variable selection, an LSTM,…
This document records a multi-stage forecast of whether the Federal Reserve will raise rates during 2026. Three agents search for evidence and produce probabilities, followed by debate and a supervisor update. Their views differ: two assign a low probability…
This notebook runs the NLinear member of a declared three-model sequence-learning population for S&P 500 options. It resolves the requests and checkpoints for the full population before fitting this member, so later notebooks can execute the LSTM and…
This notebook describes the holdout prediction stage of a US firm-characteristics study. It takes the model configuration selected using validation results, reconstructs its training specification and checkpoint, then fits it again using data that ends…
This configuration note explains why an L1-penalized, three-class logistic regression uses SAGA instead of scikit-learn’s liblinear solver. It cites multiclass compatibility in newer scikit-learn versions and reports that SAGA completed much faster in…
The notebook describes an expanding-window double machine learning analysis of whether 12-to-2-month momentum predicts next-month US firm returns after adjustment for beta, idiosyncratic volatility, market capitalization, and variance. It fits nuisance…
This notebook presents a pipeline for extracting structured corporate events from SEC 8-K filings and loading them into a Neo4j knowledge graph. A language model processes filings in batches to produce event records with subjects, relations, objects, and…
This notebook compares Optuna hyperparameter searches for XGBoost, LightGBM, and CatBoost on a firm-characteristics dataset. Each library receives its own search study, using a tree-structured Parzen estimator, median pruning, early stopping, and persistent…