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 notebook refits the configuration selected by earlier validation stages using pre-2021 history, then publishes predictions for a 2021 holdout. It retrieves the chosen configuration from a recorded candidate set or applies the same ranking rule when that…
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 screens financial and model-based features for their relationship with the next eight-hour return of crypto perpetual contracts. For each settlement, it computes cross-sectional Spearman correlations across eligible contracts, then summarizes…
This notebook aggregates existing walk-forward model results across financial case studies to compare LSTM, NLinear, TSMixer, TCN, and PatchTST with linear, gradient-boosted, and tabular neural-network baselines. Its main measure is average daily Spearman…
This notebook examines whether published investment factors offer credible return premia and diversify one another. It draws on long-history and cross-asset series from AQR alongside Fama-French equity factors, using serial-correlation-aware statistics and a…
This notebook treats total value locked as alternative data and evaluates the hypothesis that DeFi capital conditions can help predict ether returns. It defines TVL as the dollar value of crypto assets deposited in protocols, then emphasizes that it is a…
This notebook examines why a model’s information coefficient does not determine the Sharpe ratio of a strategy built from its predictions. Across a set of case studies, it compares signal rankings with backtest outcomes and considers how entry rules,…
This notebook implements a daily educational adaptation of an adversarial stochastic discount factor model. A portfolio-weight network constructs a factor from next-day excess returns using characteristics and market state available at the prior close. An…
This analysis assesses whether selected trading signals appear reliable by comparing information coefficient (IC) evidence with the Sharpe ratios produced after predictions are converted into positions. It aggregates model-family results across asset…
This notebook aggregates previously run walk-forward results to compare LSTM, NLinear, TSMixer, TCN, and PatchTST across case studies, alongside linear, gradient-boosted, and TabM baselines. It reports average daily Spearman information coefficients with HAC…
This exploratory notebook profiles hourly CME futures data across products and asset classes, using the E-mini S&P 500 as an example. It explains the hierarchy from product to expiring contract to continuous series, summarizes coverage windows, and checks…
The notebook previews S&P 100 10-K and 8-K filings intended as inputs to a financial knowledge graph. It describes schema checks for required fields, unique company and accession keys, consistent form labels, year alignment, and text-length metadata. Annual…
This document describes a reader-facing interface for examining backtest results stored in a registry. It supports summaries of run counts, inspection of specifications and individual runs, ranking results, comparing allocation methods, viewing concentration…
This notebook analyzes timestamped financial event edges from a knowledge graph, distinguishing when an event happened, when it became public, and when the extraction pipeline created the edge. It calculates disclosure and extraction lags, builds snapshots…
This chapter review describes a hypothesis-driven process for defining trading strategies before backtesting. It connects documented data assumptions and immutable configuration to exploratory analysis, event studies, and a structured strategy term sheet.…
This notebook trains a vanilla autoencoder on hourly returns for a basket of crypto perpetual markets. Returns are standardized using the training period, and a neural network compresses the multi-asset input into a two-dimensional latent representation…
The document describes a shared analysis workflow for comparing model configurations across case studies. It first selects one configuration per case study, then derives fold-level and cross-study summaries from that selection so later analyses use a…
This analysis reconstructs NASDAQ limit-order lifecycles from ITCH messages, following orders from submission to their first cancellation-related event or execution. It distinguishes adds, deletes, partial cancels, replacements, and executions, then measures…
This notebook defines forward price-return labels for a fixed panel of crypto perpetual contracts and explains how label construction affects every later model and backtest. It shifts bar-open timestamps to the time completed data becomes available,…
This source module builds catalog tables from experiment registries for model predictions and backtests. It defines reserved fields for provenance, identity, completion status, artifact availability, and metrics. The catalog carries multiple measures that…
This capstone explains how to evaluate probability forecasts with Brier score, log score, expected calibration error, and sharpness, and how reliability diagrams reveal where forecasts are overconfident or underconfident. It describes fitting calibration…
This notebook evaluates LSTM and NLinear sequence configurations on CME futures products. Sequence models consume ordered windows of observations, testing whether the recent path contains information beyond engineered summary features. The LSTM learns a…
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 notebook adapts TSMixer to predict forward ETF returns from historical momentum features. Its architecture alternates a shared linear transformation across the time axis with a feature MLP applied separately at each date. Pre-normalization and residual…