Knowledge library

This notebook introduces a published panel for machine-learning research on US equities. It contains monthly observations with 46 firm characteristics and next-month excess returns, spanning several decades. The characteristics are grouped by economic…

This chapter explains how to build and use financial knowledge graphs when questions depend on relationships such as ownership, supply chains, or contagion. It covers entity identity, typed relationship schemas, provenance, and LLM-assisted extraction from…

This notebook compares two Transformer designs for forecasting 21-day ETF returns from eight trailing-momentum features. PatchTST groups consecutive days into patches, reducing the number of attention tokens and embedding local time patterns. iTransformer…

This document explains fractional differencing as a way to make a price series more stationary while retaining some information about its level. It derives the lag weights from the binomial expansion of a fractional difference and shows how truncating small…

This notebook compares squared error, absolute error, and Huber loss for gradient-boosted models predicting short at-the-money straddle returns. The target has a capped gain and potentially severe losses, so the fitting loss may affect how well predictions…

This notebook explains a stochastic discount factor (SDF) model for pricing the cross-section of US firm returns. It learns weights that make discounted asset returns satisfy no-arbitrage moment conditions, rather than optimizing a return-forecasting loss.…

This implementation describes a neural stochastic discount factor model and an adversarial moment network for testing asset pricing relations. The SDF network maps asset characteristics, optionally combined with a macroeconomic regime representation from an…

This model adapter applies PatchTST, a Transformer architecture for long-horizon time-series forecasting, to scalar regression. It accepts sequences arranged by time and feature, rearranges them for the backbone, and reduces the per-feature outputs to one…

The document describes an ablation study of whether learned graph embeddings improve predictions of stock returns beyond ordinary tabular features. It builds a network from historical stock-return correlations, constructs momentum, volatility, and trend…

This notebook benchmarks Random Forest bagging against XGBoost, LightGBM, and CatBoost for predicting stock returns from anonymized firm characteristics. It uses predefined temporal training, validation, and test splits, measures Spearman rank information…

This notebook frames execution of a parent order as a finite-horizon control problem. A tabular Q-learning agent chooses a multiplier on a VWAP-based trade rate using remaining time, remaining inventory, and the most recently completed interval’s spread and…

The notebook treats each window of returns as an empirical distribution and clusters windows according to one-dimensional Wasserstein distance. For equal-sized samples, sorting gives the optimal quantile matching; the distance therefore reflects differences…

This chapter overview maps the research steps between validated market data and model evaluation. It covers split-aware preprocessing, encoding and missing-data decisions, then execution-consistent target construction, including fixed-horizon and event-based…

This document describes a shared workflow for training temporal convolutional networks on daily FX pair data. A model consumes a fixed-length sequence, and a missing expected day makes any window crossing that gap ineligible. The runner derives the valid…

This notebook explains a pedagogical selective state space model based on Mamba and evaluates it for ETF return prediction. Unlike a fixed-parameter state space model, it derives the input-to-state term, state-to-output term, and step size from the current…

This utility constructs a Boolean asset-to-asset adjacency matrix from asset group labels and optional links between groups. Assets assigned to the same macro group are connected to one another, and each supplied cross-group edge connects all members of both…

This notebook builds a retrieval pipeline over excerpts from corporate 10-K filings. It loads a stable set of filings, splits text into sentence windows, creates local embeddings, and indexes the chunks in a vector database. The measured run focuses on…

This configuration note explains why an L1-penalized, three-class logistic regression uses SAGA with an explicitly tightened tolerance. It compares SAGA with liblinear on a Nasdaq 100 microstructure prediction panel, reporting faster fits and better…

This notebook introduces path signatures as fixed-length features that preserve the order and joint geometry of observations in a time-series window. The first level records each coordinate's total change; the second captures signed area between coordinate…

This notebook describes an expanding-window double machine learning analysis of whether 12-to-2-month momentum predicts next-month US firm returns after conditioning on Beta, IdioVol, LME, and Variance. Nuisance models train on earlier decision months, with…

This notebook builds N-BEATS for forecasting SPY closing prices. Fully connected blocks process a complete lookback window and each produces a backcast, which is subtracted before the next block, and a forecast for the future horizon. Constrained blocks use…

This notebook presents a structured state record for research agents that preserves a run's question, cutoff, evidence, unresolved questions, tool history, quality-gate results, and synthesis status. It distinguishes this state from a conversation trace: the…

The notebook trains a vanilla autoencoder on standardized hourly returns for a group of crypto perpetual markets. Its encoder compresses the cross-asset return vector into a two-dimensional latent representation, and its decoder reconstructs the input. The…

The notebook compares Monte Carlo dropout and deep ensembles as ways to estimate uncertainty around daily ETF return forecasts. Dropout remains active across repeated predictions, while ensemble members are separately trained; the spread of predictions from…