Summaries and key ideas, written by Stratmill's research agent, of the books, papers, articles and code our AI agents read. Each page links to its original.
1,124 documents
This notebook explains how to evaluate probability forecasts and govern a forecasting pipeline. It compares Brier and log scores, expected calibration error, and sharpness; shows how reliability diagrams reveal where forecasts are over- or under-confident;…
This deployment demonstration retrains a Ridge model on historical daily data for a cross-section of major FX pairs, fetches current bars through an Interactive Brokers paper session, ranks the pairs, and builds a long basket. It walks through connecting to…
This tutorial turns ETF cross-sectional signals into long-only, equal-weight top-ten portfolios and compares Ridge regression and logistic classification with momentum and equal-weight baselines. Models are fitted on walk-forward training windows, with a…
This study tests whether the poor fit of a linear model to monthly US stock returns comes from its squared-error objective. It holds the unwinsorized forward-return target, features, and walk-forward folds fixed while comparing boosted tree models trained…
This notebook trains a Proximal Policy Optimization agent to liquidate a fixed order over a set horizon. It compares the learned pacing policy with TWAP and an Almgren–Chriss schedule, measuring implementation shortfall and examining when each strategy…
This notebook presents the Rademacher Anti-Serum (RAS) protocol as a way to adjust performance estimates for selecting from a finite class of strategies or signals. Empirical Rademacher complexity estimates how well the candidate performance matrix can align…
This notebook introduces NLinear as a baseline for predicting stock returns from ordered feature windows. Each example contains 60 consecutive sessions for one stock. For every feature, the model subtracts its latest value, maps the resulting history to a…
This reference describes two public SEC filing sources for studying equity positions and insider activity. Quarterly 13F reports provide institutional holdings, either for a curated set of managers or through bulk data covering a full filing window. Form 4…
This notebook reconstructs the prevailing national best bid and offer (NBBO) for each AAPL trade during the regular session on March 16, 2020. It chronologically combines quote and trade events, carries the latest bid and ask forward, and removes trades…
This notebook fits TabM neural models to an S&P 500 options short-straddle return target and compares three model capacities on the same feature panel used by linear and boosting approaches. TabM shares a neural backbone across ensemble members, while…
This notebook demonstrates deep hedging for a short European call. It simulates geometric Brownian motion paths, uses Black–Scholes delta hedging as a benchmark, and trains a semi-recurrent neural network to choose underlying positions that minimize a…
This notebook evaluates position-level controls for an S&P 500 equity-and-options allocation, comparing fixed stop losses, trailing stops, and time exits with each strategy's own no-overlay baseline. It selects an eligible strategy lineage using validation…
This document describes a foreign exchange OHLCV dataset covering G10 majors and crosses, with daily and four-hour bars from OANDA. It outlines how the data can be downloaded, loaded, filtered by pair or date range, and explored through coverage summaries…
This document demonstrates Instrumented PCA, in which asset factor loadings are modeled as linear functions of characteristics observed before the return. It uses alternating least squares to estimate the characteristic-to-loading map and realized factors,…
This document describes a reproducible US equities benchmark drawn from the Chen, Pelger, and Zhu asset-pricing replication archive. It contains roughly 1.2 million monthly stock observations from 1967 through 2016, with 46 firm characteristics,…
This document explains how TSMixer processes a stock’s recent feature history as a sessions-by-features matrix. Its blocks mix information down feature columns across time, then across features within each session. Stacking blocks lets those operations…
This notebook uses registered walk-forward GBM predictions for ETFs and CME futures to compare equal weighting, inverse conformal-width weighting, and weighting proportional to positive prediction scores. It calibrates entity-specific Mondrian…
This notebook compares zero-shot Chronos and TinyTimeMixer forecasts with an LSTM and a penalized linear model on an ETF panel. Every model receives the same univariate historical context, but the pretrained models forecast that context series rather than…
This notebook estimates whether a crypto perpetuals premium z-score is associated with subsequent eight-hour returns under an intervention, adjusting for price volatility, funding rate, and the premium’s deviation from its recent mean. Double machine…
This notebook evaluates whether option-market measures can rank future returns across stocks. It fits declared linear models on option-derived features, using walk-forward validation, and compares penalty strengths for ridge, lasso, and elastic net. The…
This notebook explains how GARCH turns volatility clustering into a per-session feature. It first uses return plots and an ARCH-LM test to check whether squared returns depend on their own lags. In a GARCH(1,1) model, the response to a new shock and the…
This study asks whether option-market quantities can rank future stock returns. Its features include implied volatility across maturities, put-call skew, term-structure slope, and the variance risk premium. Because many measures are represented in several…
This notebook adapts an adversarial stochastic discount factor design to daily U.S. equities. A portfolio-weight network forms a factor from next-day excess returns, while a competing moment network helps train the pricing kernel; a separate beta network…
This document develops a daily feature panel for ranking twenty currency pairs at the New York 5 PM close. It distinguishes rankable signals, such as standardized multi-horizon returns and channel position, from market-state measures such as volatility,…