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.
5,018 documents
This article explains the Intrinsic Curiosity Module as a way to encourage reinforcement learning agents to explore when external rewards are sparse. The module encodes consecutive states, learns to infer the action that connected them, and predicts the next…
The article introduces Harmony Search (HS), a metaheuristic that maps candidate solutions to musical harmonies and parameters to notes. It maintains a memory of candidate vectors and creates new candidates by choosing parameter values from memory, adjusting…
The article explains the Artificial Algae Algorithm (AAA), a population-based metaheuristic in which each algae colony represents a candidate solution. It describes three mechanisms: spiral movement explores the search space, an evolutionary step transfers a…
This tutorial builds intuition for fitting a straight line to a set of points by visualizing the distance between each point and a candidate line. It first corrects an indexing issue in example code that swaps coordinate values, then varies the line’s angle…
The article presents Mamba4Cast as a time-series encoder within an Actor-Director-Critic trading agent. Its described encoder processes market features, applies normalization with noise, adds hourly and daily time embeddings, uses multi-window convolutions…
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 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 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 builds forward-return labels for a strategy that reads listed-option signals and trades the underlying shares. It adjusts daily share prices for splits and dividends, uses a persistent security identifier, and defines returns between executable…
This notebook demonstrates covariate-drift monitoring for ETF momentum features and crypto perpetual futures features across contrasting market windows. It compares Population Stability Index, which measures changes in individual feature distributions, with…
This notebook explains instrumented principal component analysis as a way to make latent-factor loadings depend on observable stock characteristics rather than assigning each stock a fixed loading. A shared mapping from characteristics to factor exposures…
This document introduces a family of firm-return models that assumes a small set of unobserved common factors explains much of the cross-section. Rather than predicting returns directly from characteristics, these models infer each firm's factor exposures…
This document explains how wavelet decompositions and rolling spectral estimates describe financial returns at different time scales. Wavelets divide returns into components spanning progressively slower periods, making it possible to see which scales…
This notebook uses double machine learning to estimate whether futures carry has an effect on subsequent returns after adjusting for volatility, momentum, and cross-sectional carry rank. It distinguishes causal explanation from predictive performance: a…
This document explains how to compare an incumbent and a candidate model in shadow evaluation before allocating capital. It fixes their roles and promotion thresholds in advance, then assesses the candidate on five dimensions: improvement in annualized…
This notebook demonstrates how to configure Feast with Parquet sources, entities, feature views, timestamps, and a time-to-live policy, then retrieve features for training examples and a live-style as-of query. It compares Feast’s output feature by feature…
The notebook introduces a supervised autoencoder for equity factor modeling. Unlike unsupervised autoencoders that learn a bottleneck by reconstructing the cross-section, this model also includes forward returns in its training objective. The target…
This notebook applies a neural stochastic discount-factor model to CME futures. In asset-pricing terms, a stochastic discount factor is a random variable whose product with each asset's return has the same expected value under no arbitrage. The model…
This chapter surveys financial text representations, from dictionaries and word counts through TF-IDF, static embeddings, recurrent networks, and Transformers. It explains the trade-offs: simpler methods are fast, interpretable, and adaptable to finance,…
This notebook builds a financial retrieval-augmented generation evaluation harness with separate measures for retrieval coverage, answer faithfulness, appropriate abstention, and security behavior. It uses six hand-written fixtures, with context drawn in…
This case study evaluates monthly long-short decile strategies built from US firm characteristics, comparing linear models, gradient boosting, and latent-factor approaches under point-in-time accounting lags, walk-forward validation, and era-dependent…
This notebook uses Optuna to tune LightGBM return models while considering both cross-sectional information coefficient (IC) and prediction turnover. A single-objective search maximizes IC for comparison; an NSGA-II multi-objective search identifies…
This analysis compares three ways to convert signed model return forecasts into binary positions: a fixed zero cutoff, a trailing percentile for each symbol, and a cross-sectional percentile across symbols. It applies the rules to registered validation…