Knowledge library

This notebook explains how to combine probability forecasts from multiple agents and how to calibrate the resulting probabilities. It presents Neyman extremization, which moves the mean forecast away from a base rate according to panel size and an assumed…

This data utility supports work with the Chen, Pelger, and Zhu asset-pricing dataset, which includes stock returns, firm characteristics, macroeconomic series, and pre-split characteristic tensors. Its conversion method builds Parquet files from the…

This notebook examines whether financial features help explain a 10-session delta-hedged options return label. It compares Ridge models using one implied-volatility feature, groups of implied-volatility-dependent and independent features, and the full…

This module describes a shared way to prepare walk-forward training and validation folds for multiple model families. It separates expensive, model-independent work—slicing time windows, applying fold-specific model-based features, removing rows with missing…

This notebook evaluates LSTM and NLinear sequence models for futures forecasting. Unlike models that use one feature row per decision, sequence models consume ordered windows of past observations, potentially learning patterns that fixed summaries miss.…

This notebook compares validation predictions from multiple model families for perpetual-futures funding research. It explains that information coefficient measures rank association between predictions and subsequent returns, AUC measures ranking for up or…

This implementation describes how to construct batched sequence windows for panel data used in deep-learning studies. Windows follow the panel’s expected calendar rather than compressing each asset’s observations into consecutive rows. Missing sessions…

This notebook presents practical cleaning methods for market and firm data, including duplicate handling, gap flags, domain filters, reversal-based spike detection, outlier clipping, categorical encoding, and alignment across datasets with different…

This notebook examines whether SHAP attributions remain stable when predictions are similar or when models are refitted with different random seeds and architectures. Using ETF forward returns, it searches for sample pairs whose predictions are close but…

This notebook fits gradient-boosted tree models to options-market features that encode forecasts such as implied volatility, skew, term structure, and variance risk premium. It compares tree capacity and regression loss choices across forward-return and…

This case study tests weekly long-short carry-ranked strategies across a diversified set of CME futures, using daily data, walk-forward model evaluation, and costs for commissions, spreads, and roll slippage. It compares signal quality across model families,…

This notebook compares predictive, structural, and causal model evidence for a cross-sectional S&P 500 study combining equity features with options-derived measures such as implied volatility, skew, and term structure. It evaluates weekly forward-return…

This notebook demonstrates a trade-level diagnostic workflow that links realized failures to a model’s SHAP explanations. It trains a LightGBM model on lagged price and volatility features plus macro inputs to forecast next-session SPY returns. A fixed…

This notebook compares behavior cloning with inverse reinforcement learning (IRL) for learning from execution records. Its demonstrations come from a TWAP schedule, whose risk-neutral objective is known: spread trades evenly to limit impact while accepting…

The code computes standardized performance metrics from out-of-sample prediction tables, separately by fold and as headline aggregates. It supports regression and classification, including information coefficient, error measures, discrimination metrics, and…

This notebook uses double machine learning to test whether the association between the implied-minus-realized volatility spread and future equity returns persists after adjustment for observed confounders. It compares the adjusted estimate with naive…

This notebook assembles a research agent that searches for evidence and returns a structured probability for a binary question. It describes action parsing and validation, a turn budget, and a saved forecast artifact that records the rationale and supporting…

This notebook explains PCA as a feature-free latent-factor baseline for forecasting ETF returns. It decomposes each label’s training-window return matrix, using a loading for each fund and estimated factor premia to assemble forecasts. Because the same fund…

This notebook walks through a deployment cycle for a daily FX pairs model using Interactive Brokers as both the live data source and execution venue. It trains a Ridge regression model on a historical panel, computes a compact set of daily return, momentum,…

This notebook presents a neural stochastic discount factor model for a panel of CME futures. The model uses both product returns and observable characteristics, such as carry, momentum, and volatility, to construct latent factors and generate predictions for…

This utility describes a workflow for collecting public SEC 10-Q, 10-K, and 8-K filings and converting them into a common tabular dataset for downstream research. It applies form-specific text extraction: quarterly filings yield the management discussion…

This module describes two components in a multi-agent forecasting pipeline. A debate agent alternates between bullish and bearish arguments, requiring each side to address the other’s claims and provide a probability estimate with supporting evidence. It…

This notebook shows how to use PCMCI to search for delayed relationships among ETF and volatility returns without specifying a causal graph in advance. It contrasts panel-based conditional-independence testing with pairwise Granger tests, which account for a…

This notebook explains how PatchTST uses fixed-length patches of sequential market features to forecast NASDAQ-100 labels. Attention relates patches across a trailing hour while reducing the sequence length compared with attention over every minute. The…