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.
715 documents
This notebook evaluates NLinear, a minimal sequence model that subtracts the last observed level from each lookback window and maps the remaining values to a forecast. Its simplicity makes it a useful benchmark for more complex sequence architectures. The…
The document describes a walk-forward study of TabM neural networks predicting monthly returns from firm characteristics. Because the shared training runner fixes the objective to squared error for regression and cross-entropy for classification, the…
The document presents a workflow connecting standardized market-data loaders, a registry of prebuilt features, and signal diagnostics. Feature computation can be specified by names, parameter dictionaries, or YAML, supporting repeatable pipelines and…
The document turns a cross-asset momentum hypothesis into a monthly exploratory test using a 100-ETF universe. It constructs adjusted month-end prices, measures 12-to-1-month momentum, then ranks eligible ETFs into five equal-weight groups and compares their…
This notebook presents a family-level view of five latent-factor approaches applied to an equity option analytics case study. PCA estimates common movements directly from the return panel; IPCA maps characteristics to exposures linearly; a conditional…
This chapter treats a backtest as an attempt to falsify a strategy through explicit assumptions about signal timing, execution, rebalancing, sizing, costs, constraints, and benchmarks. It compares vectorized and event-driven simulation by their treatment of…
This source describes safeguards for reconstructing a model request from its immutable training specification and validating the resulting run. It checks that the request uses a supported identity version, canonical execution tier, matching model family and…
This notebook presents instrumented principal component analysis as a model of cross-sectional returns. Instead of assigning each firm’s characteristics direct return weights, IPCA estimates a shared linear map from characteristics to factor exposures,…
The notebook explains how to read Kalshi’s binary Federal Reserve rate contracts as probabilities and how to turn a ladder of rate thresholds into an implied distribution. It stresses that the feed contains YES bids, so prices are lower bounds affected by…
This notebook compares four neural network designs on the same task: predicting the next daily return from a window of past returns, pooled across eight exchange-traded funds. It describes how a dense network, a one-dimensional convolution, an LSTM, and a…
This notebook explores an economic data panel built from Federal Reserve Economic Data series. It establishes that the shipped panel uses calendar dates, including weekends and holidays, and shows how accompanying metadata identifies each series’ meaning,…
This notebook explains why ordinary t-tests can overstate the strength of an information coefficient (IC) when daily IC observations are dependent. It constructs a momentum signal on ETF data, examines the IC autocorrelation, and introduces Newey–West…
This notebook fits a long short-term memory network to NASDAQ-100 microstructure features to predict forward returns over several minute-scale horizons. Unlike a flattened linear model, an LSTM processes each observation in sequence and learns how much past…
This chapter explains how to evaluate financial machine-learning models while respecting the information available at each decision time. It distinguishes training, validation, and a final holdout, and explains why randomly shuffled k-fold validation lets…
This notebook queries registries for multiple trading case studies and produces cross-dataset tables for signal, allocation, cost, and risk comparisons. It reports diagnostics on the validation-ranked configuration cluster, including the top-to-tenth Sharpe…
This notebook builds N-BEATS, a fully connected neural forecasting architecture that processes a whole input window and predicts a future path. Its interpretable variant constrains blocks to express components through polynomial trend and Fourier seasonality…
This notebook isolates portfolio allocation by fixing each case study’s selected signal and comparing allocation-stage backtests. It considers methods such as equal weight, inverse volatility, score weighting, mean-variance optimization, risk parity, and…
This notebook compares portfolio allocation methods across case studies while holding each study’s selected signal prediction fixed. It loads allocation-stage backtests tied to the same spine prediction, then examines methods such as equal weight, inverse…
This notebook describes reconstructing a NASDAQ limit order book from DataBento market-by-order messages for a single symbol and trading day. It lays out a modular engine that tracks individual order state, aggregates orders into price levels, and maintains…
This notebook compares LightGBM models with a ridge baseline for predicting ETF returns over two horizons. It explains why trees can discover regime-dependent interactions, such as momentum behaving differently under market stress, while greedy splits can be…
This notebook evaluates a DeFi total value locked series as alternative data for trading ether. It organizes the review around four questions: whether the series relates to forward returns, whether the data is sound and reconstructible, whether its use is…
This notebook compares structured graph-style retrieval with embedding-based vector retrieval on institutional holdings from 13F filings. It constructs a point-in-time holdings snapshot at a fixed cutoff, limits the stock universe, and creates questions…
This notebook demonstrates a workflow for connecting losing trades to the model explanations associated with their decisions. It builds a single-asset SPY feature panel from price and macroeconomic data, trains a LightGBM model on earlier observations, and…
This notebook describes a 15-minute backtest workflow for predictions on NASDAQ-100 stocks. It first runs a random-signal plumbing check: because random trading should lose after costs, a persistently profitable result can indicate problems such as…