Knowledge library

This notebook describes a double machine learning analysis of the effect associated with an FX momentum treatment after adjustment for configured confounders. Flexible nuisance models estimate the outcome and treatment from those confounders; cross-fitting…

This notebook compares three linear forecasters, a Transformer encoder, and two parameter-free forecasts on daily SPY returns. Linear models map a historical window to a multi-day forecast; variants first separate a smoothed component or account for the last…

This notebook configures an NLinear forecasting run for FX pairs. The model uses a fixed consecutive lookback and subtracts the last observed level, directing its fit toward changes over that window. It resolves the lookback, normalization, device, folds,…

This code defines safeguards for reproducible cross-validation, eligibility tracking, and fold-scoped temporal features. It normalizes fold boundaries, compares requested folds with the boundaries used to create temporal artifacts, and rejects incompatible…

This notebook outlines validation-only diagnostics for a 10-session delta-hedged S&P 500 options return label. It organizes financial predictors into implied-volatility-dependent and independent groups, then compares Ridge models using a single volatility…

This notebook recasts prediction of 21-session ETF forward returns as a binary task: positive returns are labeled up, and all others down. It fits L2- and L1-regularized logistic regression using chronological walk-forward folds with a purge gap. Scaling is…

This notebook tests whether an LSTM can extract temporal structure from ETF feature histories that flat-feature models may miss. It resolves the declared sequence population against available features, labels, entities, and walk-forward folds before fitting.…

This notebook builds rule-based features for a cross-section of CME futures, centered on carry from the spread between nearby delivery contracts. It also constructs momentum, volatility, curve-shape, and calendar features. The design distinguishes raw…

This notebook demonstrates two sequential monitors on validation prediction errors from two linear equity model configurations: a two-window mean-shift detector and a monitor for the frequency of bad days. Each detector is calibrated on an initial period and…

This notebook applies double machine learning (DML) to estimate the effect of skip-recent momentum on ETF forward returns, a causal question distinct from forecasting returns. It models the outcome and the momentum treatment using declared confounders, then…

The notebook demonstrates tuning LightGBM for ETF return prediction with Optuna’s TPE sampler. It searches tree structure, sampling, and regularization settings, using early stopping to choose the boosting rounds and a custom callback to report…

This chapter presents a research workflow for using predictive models in trading, where stable out-of-sample forecasts may matter more than unbiased coefficient estimates. It covers regularized regression methods such as Ridge, LASSO, and Elastic Net, along…

This notebook turns institutional 13F holdings into a bipartite institution-to-stock graph and derives features for research, including stock co-ownership similarity, ownership breadth and concentration, and changes in reported holdings. It aggregates…

This notebook compares exhaustive grid search with Optuna’s TPE Bayesian sampler for tuning a LightGBM model on an ETF forward-return prediction task. It first gives both methods the same trial budget over a small categorical grid, where exhaustive search…

This chapter presents strategy research as the design and evaluation of an executable decision process, from the initial economic idea through position sizing, constraints, costs, and live-like testing. It recommends classifying strategy families and…

This notebook builds model-based features for S&P 500 options research from underlying returns. It fits GJR-GARCH, which gives extra weight to negative return shocks, and a stochastic-volatility model whose latent variance is estimated with MCMC and tracked…

This notebook trains TabM neural networks to rank ETFs using the same flat feature table as linear and boosted models. Each ensemble member shares a two-layer backbone but has its own scaling vector and output layer, allowing predictions to be averaged with…

This notebook adapts skip-gram Word2Vec to institutional holdings by treating each 13F portfolio as a sentence, each stock identifier as a token, and position size rank as token order. Nearby positions form the context, so stocks that institutions place in…

This notebook presents a GT-GAN-inspired generative model for financial series sampled at irregular intervals, such as tick, volume, or dollar bars. Its encoder, generator, discriminator, and decoder use continuous-time Neural ODE dynamics, allowing latent…

This notebook describes fitting PatchTST to one-minute NASDAQ-100 data to predict returns over several forward horizons. The model groups consecutive observations into patches and applies attention across them, reducing the number of items compared while…

This notebook explains an unconditional signature-based Wasserstein GAN for generating financial time series. It transforms returns into augmented paths, computes truncated path signatures, and trains an LSTM generator driven by Brownian noise to match…

This notebook demonstrates feature-drift checks on ETF momentum and technical features across calm and stressed windows, and on crypto perpetuals with premium-index data across market regimes. It compares Population Stability Index (PSI), including its…

This notebook evaluates TSMixer as a global sequence model for an ETF panel. The model shares parameters across funds while using each fund’s own history and covariates; its mixing layers learn temporal and within-series relationships without combining one…

This notebook describes using TSMixer to predict stock returns from ordered windows of each stock’s features. Each example contains 60 consecutive sessions; windows that cross gaps in a stock’s history are excluded, so the number of usable examples varies…