Knowledge library

This document explains a CME futures experiment using TabM, a parameter-efficient neural ensemble, on the same engineered feature rows and walk-forward folds used by linear and gradient-boosting models. TabM shares most weights across ensemble members to…

The document presents NLinear, a minimal sequence model for forecasting NASDAQ-100 stock returns from intraday features. It uses a fixed trailing window, subtracts the latest observation so the model works with changes rather than levels, and maps the…

The document describes a pipeline for turning SEC 8-K filings into structured corporate events. A language model extracts event relationships, while explicit schema checks validate entity names, relation types, categories, and dates. Deterministic…

This module describes migrating a registry after a library changed the numbering of walk-forward folds while leaving their date windows and fitted values unchanged. Since fold identifiers appear throughout training specifications, prediction keys, input…

This notebook describes producing predictions for a US equities case study’s holdout window after selecting a configuration on validation folds. It resolves the selected configuration, preserves its checkpoint choice, and constructs a new training…

This notebook explains how hidden Markov models infer unobserved market states from returns and recent volatility. It first establishes simple comparison rules: a volatility index threshold for stress and price relative to a long moving average for trend. A…

The notebook explains how to compare prediction models for S&P 500 options without allowing incomplete or mismatched evaluation samples to distort the results. It identifies each model by family, configuration, and checkpoint, checks that predictions meet…

This notebook builds uncertainty intervals for ETF forward returns using Ridge regression and three conformal methods: split conformal prediction, conformalized quantile regression, and adaptive conformal inference. It evaluates them with walk-forward folds…

This notebook uses double machine learning to estimate whether a perpetual-futures premium z-score is associated with the following eight-hour return after adjustment for six pre-treatment controls. It compares effects across high- and low-volatility regimes…

This notebook explains a conditional autoencoder that models stock returns as characteristic-dependent factor loadings multiplied by latent factors. It first estimates characteristic-managed portfolio returns with a joint cross-sectional least-squares fit,…

This notebook compares linear models on ETF features built from overlapping return windows and related momentum, volatility, and oscillator measures. Such predictors are collinear, making ordinary least squares coefficients unstable. The experiment fits…

Fractional differencing applies a non-integer power of the lag difference operator to log prices. Its binomial weights decay across past observations, so the result combines price changes with some surviving level information. The fractional order controls…

This notebook describes an LSTM model that predicts returns from ordered windows of stock features in a US equities panel. Unlike a flat tabular model, the LSTM processes sessions sequentially and carries a hidden state through each window. Learned gates…

This notebook presents TimeGAN, a recurrent generative adversarial model for producing synthetic financial sequences. Its five modules learn an embedding and reconstruction, predict successive latent states, generate sequences from noise, and distinguish…

This notebook applies two sequential detectors to real validation prediction errors from two US equities linear-model configurations over a calendar year. A two-window mean-shift detector looks for changes in error magnitude, while a bad-day frequency…

This notebook presents a supervised autoencoder for predicting ETF forward returns. The network compresses a fund’s feature row into a low-dimensional representation and predicts returns directly from that representation. Unlike conditional autoencoders in…

This document describes how to derive a single holdout training interval from an existing walk-forward validation specification. It takes the declared holdout dates from the study configuration, uses the earliest training start across validation folds, and…

This analysis compares feature-triage ledgers across nine trading case studies. It distinguishes Benjamini–Hochberg false-discovery-rate significance from a second route to a PROCEED decision, which combines a case-specific effect-size floor with fold sign…

This notebook fine-tunes three transformer checkpoints for three-class financial sentence sentiment and compares their accuracy, macro F1, confusion matrices, parameter counts, and training time. It uses a stratified train, validation, and test split of the…

This notebook builds principal-component factors from a panel of CME futures returns, not from engineered characteristics such as carry, momentum, or volatility. PCA finds directions that explain variation across products; the first may resemble a common…

This document introduces two latent-factor approaches for futures returns. Principal component analysis finds uncorrelated directions that explain the most return variation, while a neural stochastic discount factor seeks combinations related to the…

This notebook screens candidate features for a strategy that sells at-the-money straddles and holds them to expiration. For each trading day, it compares the cross-sectional ranking of available names by a feature with the ranking of their eventual straddle…

This document explains how to build daily forward-return labels for ETFs using adjusted closing prices at the observation date and a later trading session. It creates a primary monthly-horizon label and a shorter weekly variant, with labels computed on each…

This notebook compares split-conformal prediction, conformalized quantile regression, and adaptive conformal inference for forecasting ETF forward returns. In walk-forward folds, it trains on one subset, calibrates interval errors on another, and evaluates…