Resumos e ideias principais, escritos pelo agente de investigação da Stratmill, dos livros, artigos científicos, artigos e código consultados pelos nossos agentes de IA. Cada página inclui uma ligação para o original.
1,124 documentos
This document describes a one year out of sample backtest of S&P 500 option straddles. It applies predictions from a model refit on pre holdout history, together with the previously selected strategy, allocation, concentration, weekly entry schedule, hedge…
This notebook specifies and executes a double machine learning analysis of the effect of the variance risk premium on short-option returns through expiry. Before execution, it resolves the treatment, outcome, confounders, timing, nuisance model, temporal…
This analysis compares predictions from several model families trained on monthly US stock characteristics to forecast next-month returns. It focuses on cross-sectional information coefficient, which measures how well a model ranks stocks within each month.…
This case study fits regularized linear models to returns from short at-the-money straddles held to expiry. The trade collects call and put premiums, giving it a capped maximum gain but potentially very large losses when the underlying moves sharply.…
The document explains how an experiment registry can track a model from its training configuration through predictions to backtest results. Each stage receives an identifier derived from a canonicalized specification, allowing repeated identical runs to…
This notebook compares methods for discovering relationships among a panel of ETF returns: NOTEARS for contemporaneous linear directed acyclic graphs, VAR-LiNGAM for lagged and instantaneous structure, PCMCI for conditional-independence links, and Granger…
This notebook demonstrates tuning LightGBM hyperparameters with Optuna's TPE sampler, using cross-sectional information coefficient as the objective. It combines early stopping to choose the number of boosting rounds with a custom pruning callback that…
This chapter presents portfolio construction as the process of converting return forecasts, risk estimates, and constraints into weights, leverage, and rebalancing decisions. It lays out a research workflow for documenting allocator choices, avoiding…
This notebook explains how to evaluate position-level exits and combine them with portfolio-wide controls. Fixed stop-loss, take-profit, and time exits are contrasted with trailing and tightening stops; a scaled exit reduces a position at successive profit…
This notebook studies how ridge, lasso, and elastic net behave when a crypto perpetuals feature matrix measures one economic quantity—the premium—many different ways. Premium levels, changes, volatility, standardized positions, ranks, and related funding…
This document describes data access and alignment conventions for crypto perpetual futures and related on-chain series. It explains that premium-index bars are timestamped at their opening time: an eight-hour bar records the premium leading into the funding…
This study converts registered model predictions into comparable S&P 500 option strategies. On weekly decision dates it ranks predicted returns, filters to a liquid universe, and sells equally weighted at-the-money straddles on the highest-ranked symbols.…
This notebook brings together five latent-factor approaches for modeling the S&P 500 options case study’s equity return cross-section. PCA estimates common movements from returns alone; IPCA maps characteristics to exposures linearly; a conditional…
This notebook demonstrates the Rademacher Anti-Serum protocol as a way to account for selecting a winner from a class of tested strategies. It estimates empirical complexity from candidate performance paths, illustrating how dependence among candidates…
This notebook develops a two-model exit policy for hourly crypto perpetuals. An entry classifier identifies unusually strong forward returns, while an exit classifier predicts whether the next forward return will be negative. The exit model receives…
This notebook synthesizes results from nine market case studies into a cumulative strategy-screening funnel. It tests, in order, whether a model has positive information coefficient, whether its selected configuration has positive validation Sharpe, whether…
This notebook explains when a strategy is clearest as precomputed arrays and when it benefits from a sequential simulation that carries positions, fills, cash, realized profit and loss, or equity forward through time. Array-based backtests are attractive…
This notebook assesses whether total value locked can serve as an alternative-data signal for ether returns. TVL aggregates the dollar value of crypto assets deposited in decentralized finance protocols. Because it is a price-valued stock rather than a…
This notebook evaluates one previously selected S&P 500 options configuration on a holdout period. It reuses the registered model predictions and strategy settings, including the signal schedule, allocation, hedge rule, and trading costs, without tuning them…
This notebook teaches how to decode NASDAQ TotalView-ITCH binary messages and store them as structured data for later market microstructure analysis. It explains message framing, fixed-width field layouts, big-endian values, timestamps measured from…
This notebook explains a supervised autoencoder for predicting the direction of future US equity returns across multiple horizons. Its encoder feeds a reconstruction decoder, an auxiliary classifier, and a main classifier. Joint training combines…
This notebook builds a portfolio allocator that places a Temporal Fusion Transformer-style variable-selection network before an LSTM encoder. The selection network embeds each input feature separately and assigns softmax weights, allowing the model to vary…
This notebook compares pandas and Polars on operations used in financial data pipelines, including rolling calculations, grouped OHLCV summaries, window statistics, filtering, joins, lazy file scans, memory use and string processing. It generates synthetic…
This notebook describes a TabM workflow for foreign-exchange pair models. TabM applies a small neural network to each decision row rather than treating the observations as a sequence. The notebook takes architecture and checkpoint schedules from…