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.
55 documents
This notebook applies TabM, a tabular neural-network approach, to foreign-exchange pair prediction rows without treating the data as a sequence. It uses shared runner infrastructure to fit preprocessing within each training fold, save declared weight…
This notebook builds a daily feature panel for a long-short ranking strategy across twenty FX pairs. It aggregates four-hour spot bars into sessions ending at the New York 5 PM rollover, then constructs trailing return, channel, volatility, drawdown, range,…
This notebook explains how double machine learning (DML) estimates whether FX momentum affects future returns after adjusting for configured confounders. It distinguishes this intervention question from prediction: predictive models are compared by…
This notebook studies how position sizing affects FX backtests after the model has already selected which currency pairs to trade. It preserves the winning baseline’s predictions, signal mapping, costs, and execution settings, then varies allocation rules.…
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…
This notebook tests whether gradient-boosted trees improve cross-sectional ranking across currency pairs beyond a penalized linear model. The FX universe contains pairs sharing currencies, so observations are dependent: a move in one currency affects…
This notebook checks whether four-hour spot FX data can support a daily cross-sectional strategy that ranks currency pairs using momentum and carry. It tests whether the declared instruments have prices at each decision point, whether the universe represents…
This notebook describes a double machine learning analysis of the effect associated with an FX momentum treatment after adjustment for configured confounders. Flexible nuisance models estimate the outcome and treatment from those confounders; cross-fitting…
This notebook configures an NLinear forecasting run for FX pairs. The model uses a fixed consecutive lookback and subtracts the last observed level, directing its fit toward changes over that window. It resolves the lookback, normalization, device, folds,…
This notebook audits the complete set of canonical validation predictions for an FX pairs study. It checks model identity, artifact availability, completeness, configured-label coverage, and whether the assembled configurations match the declared menu.…
This notebook audits the registered validation predictions for an FX pairs study. It checks that the catalog includes the configured model population, complete prediction sets, available artifacts, and current model identities. It then summarizes predictive…
This notebook presents correctness and runtime comparisons for VectorBT Pro and VectorBT OSS against ML4T on supported case-study strategies. It covers ETF allocation, USD-quoted foreign exchange, and a US equity panel for both editions; VectorBT Pro also…
This notebook describes a shared workflow for fitting temporal convolutional networks to daily FX pair data. Instead of treating each date as an isolated feature row, the model uses a fixed lookback of consecutive observations. The runner determines eligible…
This case study evaluates daily momentum and carry signals across a small, correlated universe of G10 spot currency pairs. Its workflow moves from feasibility checks and forward-return labels through engineered and model-based features, cross-validation,…
The document explains how to create currency-pair features from three fitted time-series models: a state-space filter that estimates a slowly changing price level, a two-state hidden Markov model for dollar volatility regimes, and an ARIMA model whose…
This notebook reviews a selected foreign-exchange strategy and its holdout lineage using registered research artifacts. It reproduces the choice from a frozen validation candidate set, ranking eligible solvent backtests by validation Sharpe with a…
This notebook explains how to construct one-, five-, and 21-session forward spot returns for a fixed FX-pair universe. It first maps four-hour bars into trading sessions using a New York rollover calendar, then builds returns without removing rows before…
This notebook runs the already selected FX strategy specification against registered holdout predictions. It preserves the strategy’s signal, allocation, risk controls, rebalance rule, costs, and account settings; the prediction set and price-frame identity…
This notebook presents a current audit comparing VectorBT Pro and VectorBT OSS with ML4T on supported real-data strategy workloads. Both VectorBT editions participate in ETF allocation, USD-quoted foreign-exchange allocation, and a US equity panel. Pro also…
This audit compares Backtrader and Zipline with ML4T on real strategy cases, using only asset and contract combinations that each engine and the frozen data bundle can represent natively. It covers ETF allocation and a US equity panel for both engines, plus…
This notebook assesses whether four-hour spot FX data can support a daily cross-sectional strategy that ranks currency pairs using momentum and carry, buys the strongest, and sells the weakest. It does not fit a model or make forecasts. Instead, it checks…
This configuration describes a daily long-short strategy across 20 FX pairs. It sets a New York close decision time, next-bar execution, equal-weight sizing, and ranking signals based on momentum or carry. It also specifies spread and swap-point costs,…
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 case study compares declared fixed and trailing stop rules for FX positions. It first selects one parent strategy per label from a sealed validation cohort, then varies only the position risk rule. Because stops respond to price movements within a…