Knowledge library

This notebook organizes cross-case-study comparisons of LightGBM models, using linear models and TabM as reference families where available. It evaluates configurations across tree sizes, regression losses, checkpoints, and classification objectives. The…

This document describes a Gymnasium environment for training an agent to liquidate a position over a fixed horizon. The observation combines remaining inventory and time with spread, market depth, volatility, and a normal or stressed regime. A continuous…

This notebook compares LightGBM with an MLP, TabM, and, when available, TabPFN for predicting ETF forward returns from tabular features. Models are evaluated on shared chronological walk-forward folds, with rank information coefficient and training time as…

This notebook compares ordinary least squares with Ridge, LASSO, and Elastic Net for predicting 21-day forward returns across ETFs. The models use the same declared walk-forward folds, with a purge gap matched to the label horizon. Each fold clips feature…

The document examines the ADIA Lab causal discovery challenge as supervised inference within a known synthetic data-generating family. The task assigns variables to eight roles around a treatment and outcome, using many labeled dataset and graph pairs. It…

This module builds a mean forecast ensemble from registered prediction sets, with particular attention to reproducibility and avoiding selection bias. It resolves qualifying model configurations from the registry rather than maintaining a hand-written member…

This notebook expresses a four-phase forecasting workflow using plain Python orchestration, CrewAI, and LangGraph. It compares where state is held, how failures are inspected or recovered, what dependencies each approach adds, and how tracing works. The…

This notebook examines linear-model regularization for a universe of currency pairs whose returns are linked because each pair is a quote between shared currencies. It distinguishes feature collinearity, which ridge, lasso, and elastic net can address, from…

This chapter surveys ways to turn fitted statistical procedures into features for trading models. It covers diagnostics and stationarity, structural breaks, fractional differencing, Kalman filtering, spectral and path-signature methods, ARIMA and GARCH…

This notebook frames a parent order as a finite-horizon control problem. At each interval, tabular Q-learning chooses a multiplier on a VWAP-based trade slice, conditioning on time and remaining inventory plus the most recently observed spread and volatility…

This notebook describes a one-feature-at-a-time screen for candidate signals across 20 currency pairs. It computes each feature’s daily cross-sectional rank agreement with the next-session return using only walk-forward validation windows. The evaluation…

This notebook presents principal component analysis as an unconditioned latent-factor baseline for forecasting stock returns in an option analytics study. PCA uses the historical return panel to estimate common movements and each stock’s exposure to them,…

This notebook compares exhaustive grid search with Optuna’s TPE sampler for tuning a LightGBM return-prediction model. It first gives both methods the same trial budget over the same discrete parameter grid, where enumeration covers every combination and…

This module defines a reader-facing workflow for CME futures research, covering model requests, prediction and backtest execution, candidate pools, and final selection. It reads the configured label sweep from a shared setup declaration, preserves its order,…

This notebook demonstrates an operational cycle for an ETF equity strategy: refresh market data, compute financial features, retrain a Ridge model, generate current cross-sectional predictions, replay them through a backtest engine, and stage a top-ranked…

The document presents GReaT, a framework that serializes rows of mixed numerical and categorical financial data as text, fine-tunes a language model on those records, and parses generated text back into tables. Its example builds ETF features from historical…

The document explains how uncertainty from a model can provide information beyond its point estimate. It contrasts GARCH, which defines volatility from past returns, with a stochastic volatility model that treats volatility as a latent autoregressive process…

This notebook uses principal component analysis to represent the common movements in a broad US equities panel. PCA extracts directions that explain shared variation in returns; each stock’s loading measures its exposure to each direction. A prediction…

This notebook evaluates security rules for a document-grounded financial assistant using six hand-built cases. The cases cover prompt injection, fabricated figures in untrusted material, attempted actions, and citations to content that was not retrieved. Two…

This notebook refits the FX configuration already selected by validation and registers its predictions for a later holdout backtest. It resolves the selection from solvent validation backtests, constrained to a frozen candidate set, then rebuilds the…

This notebook presents a temporal convolutional network (TCN) for predicting ETF forward returns from sequences of trailing return features. Causal convolutions ensure an output uses only observations up to its date, while exponentially increasing dilation…

This notebook explains how to turn fitted GJR-GARCH and hidden Markov models into trading features while controlling look-ahead bias. GJR-GARCH estimates conditional volatility and captures the greater impact of negative return shocks; a two-state Gaussian…

This notebook surveys methods for finding changes in the process that generates a financial series, using daily SPY data as its example. It contrasts the Zivot–Andrews test, which estimates one break while testing for a unit root, with multiple-break…

This utility prepares prediction artifacts for live-demo notebooks by normalizing differing column names and date formats into a common timestamp, asset, and prediction structure. It groups duplicate timestamp-asset observations and averages their scores.…