Knowledge library

This document explains how a forecasting agent’s search interface can constrain its evidence and make its activity auditable. A provider-neutral protocol returns typed results with source, publication date, relevance score, and content, while a cutoff filter…

This notebook examines linear-model regularization for predicting returns across currency pairs. It distinguishes collinearity among input features from dependence among the assets: currency pairs share underlying currencies, so their returns and rankings…

This configuration note explains the choice of the saga solver and a tighter convergence tolerance for L1 logistic regression on three-class labels. It contrasts saga with liblinear, citing runtime measurements on a Nasdaq microstructure panel and reported…

This notebook studies a supervised autoencoder for predicting forward returns from firm characteristics. It retains a narrow neural-network bottleneck but removes the conditional-factor interpretation and pricing restriction used by related models. The…

This notebook distinguishes predicting returns from claiming that momentum causes them. It treats a stock’s prior-year return, excluding the latest month, as the treatment and forward return as the outcome. Recent volatility, illiquidity rank, and volume…

This notebook fits PatchTST as one member of a declared sequence-model population for S&P 500 options. PatchTST divides a lookback window into fixed-size patches, embeds them as tokens, and uses attention to learn relationships among portions of the history.…

This notebook aggregates LightGBM results across multiple trading case studies and compares them with linear and TabM baselines. It examines the highest-IC configuration for each case study, reports heteroskedasticity-and-autocorrelation-consistent…

This notebook tests whether learned graph embeddings add predictive information to a tabular model for US stocks. It builds a network from absolute return correlations measured before the target period, constructs momentum, volatility, and trend features…

This document describes an LSTM sequence model for an S&P 500 options study. Unlike cross-sectional models that rely on engineered features to summarize a symbol’s history, the LSTM receives chronological windows and learns which parts of the sequence…

The notebook builds prediction intervals around gradient boosting forecasts with split conformal prediction. It fits a model on an earlier training segment, measures absolute residuals on a later calibration segment, and uses a finite-sample order statistic…

This notebook presents a pipeline for turning financial headlines into stock-level signals. It cleans and standardizes a news corpus, removes exact and prefix-matched duplicates within ticker-date groups, and uses sentence embeddings to represent headline…

This notebook describes how to create holdout predictions for an FX strategy after its configuration has already been selected using validation results. It resolves the highest-Sharpe validation backtest among admitted candidates that remained solvent, then…

This notebook compares DQN, PPO, and A2C in a simulated cryptocurrency trading environment. It calibrates a GARCH(1,1) volatility process on hourly Bitcoin perpetual-futures returns, then simulates paths that preserve volatility clustering while leaving…

This notebook builds a forecasting agent that alternates between web searches and probability forecasts. A small two-method language-model interface lets the same loop use a mock, local, or commercial provider. The agent receives a prediction-market question…

This notebook demonstrates a read-only question-answering workflow over an institutional holdings graph built from 13F filings. Instead of generating database queries from natural language, it routes exact supported questions to hand-written Cypher templates…

This notebook describes a patched transformer for sequence prediction in an S&P 500 equity and options study. It divides a time window into contiguous blocks and represents each block as a token, allowing attention to compare distant parts of the window…

This notebook explains instrumented principal component analysis (IPCA) for a changing US equities panel. Traditional PCA assigns each stock a fitted loading, which cannot naturally cover names absent from the training sample and stays fixed as a company's…

This chapter frames reinforcement learning as a tool for sequential control, distinguishing it from supervised forecasting. It focuses on tasks where actions affect later outcomes and rewards are comparatively concrete: trade execution, market making, and…

This ETF feature notebook explains how fitted statistical models can leak future information through their estimated parameters, even when their formulas use past observations. Its remedy is a walk-forward refit schedule: allow an initial burn-in, fit using…

This notebook studies gradient boosting for a US equities panel, varying tree capacity, objective function, and training checkpoint across three forward-return horizons. It explains how shallow trees can learn feature interactions, while increasing capacity…

This notebook demonstrates SHAP explanations for a LightGBM model predicting ETF forward returns. TreeSHAP supplies feature contributions for individual predictions and aggregate importance summaries. Beeswarm and dependence plots show how feature values…

This notebook describes how to generate holdout predictions for a crypto perpetual funding strategy after its configuration has been selected using validation results. It resolves the highest-Sharpe eligible validation backtest, carries over the model…

This shared utility module supports causal analysis of trading data with time-series double machine learning (DML). Its documented tools include walk-forward estimation with an embargo, nuisance models, and refutation checks. Block permutation preserves…

This notebook explains TSMixer, a time-series neural network that alternates two operations: a shared linear map across lookback days for each feature, and a shared feature MLP applied independently at each day. Residual connections and pre-normalization…