Knowledge library

This chapter argues that durable trading research depends on a disciplined process that can adapt to changing markets, noisy evidence, and implementation costs. It introduces distinctions among structural breaks, regimes, data drift, and concept drift, then…

This evaluation notebook screens financial and model-based features for their relationship to the next eight-hour return of crypto perpetual futures. It builds the evaluation sample from walk-forward validation windows, keeps the final holdout outside the…

This NASDAQ-100 case study investigates whether 15-minute microstructure signals can survive trading costs. The workflow evaluates order flow, quote staleness, and relative spreads, then compares a full-universe strategy with one restricted to…

This notebook studies how regularization affects a CME futures model built from correlated feature families, including carry, rolling Sharpe, momentum, and volatility measures. It compares Ridge, Lasso, and ElasticNet: Ridge shrinks coefficients while…

This notebook describes adapting Diffusion-TS to generate synthetic ETF return sequences. Its denoiser predicts the clean sequence as a sum of polynomial trend, Fourier seasonal, and residual components, while a combined time-domain and Fourier-domain loss…

The chapter argues that repeatable research discipline is a more durable source of trading advantage than any individual strategy. It presents a workflow built around falsifiable hypotheses, out-of-sample evaluation, and corrections for multiple testing to…

The document explains why TabM is used to model perpetual funding panel data: its shared neural backbone combines all input features, while nonlinear layers can represent interactions. Unlike a tree that chooses one feature at each split, this design can use…

This evaluation workflow screens financial and model-based features against forward returns for a NASDAQ-100 intraday panel. It first reserves the final test period, excluding observations whose forward labels would settle within that window. For each…

This analysis estimates whether skip-recent momentum is associated with a causal change in ETFs’ forward returns, rather than asking how well momentum predicts returns. Double machine learning models the outcome and the treatment using declared confounders,…

This notebook studies gradient-boosted models for forecasting returns and direction in crypto perpetuals. Its grid varies tree capacity and loss function, including squared-error, absolute-error, and Huber objectives, while scoring models at multiple points…

This document presents the PatchTST neural-network backbone for forecasting multivariate time series. It divides each input series into overlapping patches, projects each patch into a learned feature space, adds positional encodings, and processes the…

This notebook estimates whether the premium z-score of a crypto perpetual affects subsequent eight-hour returns when three declared drivers—price volatility, funding rate, and average premium deviation—are held fixed. Because treatment is continuous, the…

This notebook presents an unconditional Sig-Wasserstein GAN for generating financial time series. It replaces a learned real-versus-generated discriminator with a distance based on expected path signatures, mathematical summaries of a path’s sequential…

This notebook explains how to reconstruct a model experiment catalog from a case-study registry and compare its contents with MLflow. It follows experiments through training runs, prediction sets, and backtest runs, linking each stage to the preceding one.…

This notebook compares ordinary least squares with Ridge, LASSO, and Elastic Net for predicting forward returns across an ETF universe. All models use the same walk-forward folds, with a purge gap matched to the label horizon, so differences can be…

This chapter outlines how to build time-consistent fundamentals and alternative-data research across equities, macroeconomic series, futures, crypto, and prediction markets. Its central methods are bitemporal storage and as-of queries, source-aware…

This notebook compares linear models for predicting ETF returns from overlapping price-derived features. Since multiple return horizons and indicator families can encode similar information, ordinary least squares may assign unstable opposing coefficients to…

This guide maps causal questions to Python libraries and clarifies the distinction between estimating effects under an assumed causal structure and discovering possible structures from data. It introduces double machine learning for adjusted treatment…

This notebook introduces a stochastic discount factor model for pricing a cross-section of ETFs. Instead of separately estimating fund exposures and factor returns, it trains networks to find a common discount-factor series whose covariation with fund…

This notebook applies TabM, a weight-sharing neural ensemble, to a panel of cryptocurrency perpetual-futures features. Its members share a neural backbone but have separate scaling vectors and output layers, combining efficient ensembling with nonlinear…

This notebook screens financial and model-derived features for their ability to rank ETFs by subsequent returns. It calculates a date-by-date information coefficient by comparing each feature’s cross-sectional ranks with forward-return ranks, then examines…

This notebook frames a focused test: whether a learned sequence representation of perpetual-funding features improves on hand-built summaries of the premium, the gap between perpetual and spot prices. It compares NLinear, which applies a linear mapping to…

This notebook examines whether SHAP feature attributions are stable enough to support model pruning, risk decisions, or reporting. Using an ETF prediction task, it first searches for test samples whose predictions are close while their SHAP profiles differ.…

This notebook evaluates candidate features for 20 currency pairs one at a time against forward returns. It limits the analysis to walk-forward validation periods, stamps each session with its fold, and leaves a buffered gap before the holdout so labels…