Knowledge library

This notebook builds model-based conditional volatility features for S&P 500 shares with a GJR-GARCH(1,1) model, then compares an option-implied volatility spread measured against realized volatility with one measured against a model forecast. Since…

This CME futures study varies gradient boosting tree capacity and loss function while recording predictions at multiple training checkpoints. Tree leaf count controls how finely a model partitions the feature space, while squared, absolute, and Huber losses…

This notebook studies how Ridge regularization choices affect information coefficient on an ETF panel and how hyperparameter search can inflate reported performance. It first evaluates a wide alpha grid across walk-forward folds, scaling features within each…

This notebook explains instrumented principal component analysis (IPCA), which models each ETF’s factor exposures as a shared linear function of that fund’s observed features. Alternating least squares estimates the feature-to-exposure map and the factor…

The notebook presents an autonomous research workflow in which a language model uses general-purpose tools to inspect files, run commands, query a backtest registry, and examine data. Rather than hard-coding research actions, the operator discovers…

This notebook fits a PCA latent-factor model to a panel of stock returns, without using option-surface features, characteristics, or target information. PCA extracts leading directions of historical cross-sectional co-movement, estimates each stock's…

This shared feature-engineering module defines construction rules and audits for financial predictors used across multiple case studies. It records each feature family with its hypothesis, inputs, rolling lookback, information lag, role, and potential…

This analysis compares learned signals for monthly cross-sectional US stock returns using accounting and market characteristics, with next-month return as the target. It reviews linear models, boosted trees, deep tabular methods, latent-factor estimators,…

This chapter presents agentic workflows for financial research tasks where gathering, filtering, and synthesizing uneven evidence is the main challenge. It contrasts these workflows with conventional models that work well on stable structured inputs, and…

This notebook defines and validates monthly return labels for a cross-sectional US firm-characteristics study. It checks how the data provider pairs each row’s characteristics with the return outcome, emphasizing that the panel is already aligned and must…

This notebook builds an end-to-end portfolio allocator that places a Temporal Fusion Transformer style variable-selection network ahead of an LSTM. Separate gated residual networks embed individual features, and learned soft weights combine those embeddings…

This notebook explains how to evaluate financial models without allowing information unavailable at a decision date to leak into training or model selection. It distinguishes training, validation, and a final holdout set, and shows why randomly shuffled…

This notebook compares four ways to discover relationships in a panel of ETF returns: NOTEARS for contemporaneous linear DAGs, VAR-LiNGAM for lagged and instantaneous structure, PCMCI for conditional-independence tests, and Granger tests as pairwise…

This notebook studies how Ridge regularization affects information coefficient (IC) performance and how hyperparameter search can bias reported validation results. It first evaluates a wide alpha grid with walk-forward folds, scaling features inside each…

This notebook presents Tail-GAN, a generative adversarial approach designed to reproduce financial scenario tail risk, especially value at risk (VaR) and expected shortfall (ES). It uses differentiable sorting so quantile-based risk measures can contribute…

This notebook examines how to combine multiple probability forecasts. It presents simple averaging as a baseline and explains Neyman extremization, which moves a panel mean farther from a base rate according to the number of forecasters and an assumed…

This notebook turns futures carry—the price gap between nearby expiries—into model-based features. It describes one-step ARIMA forecasts per product, rolling Fourier measures of seasonal cycle strength, and a two-state hidden Markov model that infers broad…

This overview explains a family of models for predicting returns from firm characteristics. Rather than treating each characteristic as a direct return predictor, the models assume returns are driven by a small set of unobserved common factors and estimate…

This document presents a pedagogical Mamba-style selective state space model for predicting ETF returns from sequences of momentum features. Unlike attention and dense temporal mixing, a state space model carries a fixed-size state through the sequence with…

This document describes a conditional autoencoder that models firm exposures to latent return factors as a nonlinear function of firm characteristics. Like IPCA, it estimates a factor structure in which predicted returns combine firm-specific exposures with…

This document explains a TabM neural network for ranking ETFs from tabular features. Its shared two-layer backbone feeds multiple members, each with a learned scaling vector and output layer; averaging their predictions aims to reduce initialization noise…

This chapter overview explains how gradient-boosted trees build on decision trees and random forests for financial tabular prediction. Boosting adds learners sequentially to correct earlier errors, helping capture nonlinear relationships and threshold…

This document describes a monthly US equity dataset designed for machine learning research in asset pricing. It contains 94 rank-transformed firm characteristics, including accounting measures and technical features, alongside returns. Firms are anonymized,…

This notebook evaluates security controls for a financial document question-answering system using six hand-written cases. The cases represent indirect prompt injection, fabricated data in untrusted material, attempted actions, and invalid citations. It…