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 how bar timestamps can create look-ahead bias in financial machine-learning datasets. MetaTrader 5 time bars are labeled with their opening time, although their high, low, close, and other completed-bar features are only known later.…
The article explains MSFformer, a transformer architecture for time-series forecasting that combines multi-scale feature extraction with pyramidal attention. Its Coarser-Scale Construction Module (CSCM) uses successive feature-convolution blocks to build…
The article presents the Big Bang–Big Crunch algorithm, a population-based method for searching a bounded parameter space. It begins with random candidate points, then repeatedly computes a fitness-weighted center and samples new candidates around it.…
The article describes building a deep belief network initialized with stacked restricted Boltzmann machines, then fine-tuning it with labeled data. For EURUSD on a 30-minute timeframe, it constructs inputs from technical indicators and predicts directional…
The article introduces the Artificial Bee Hive Algorithm as a population-based method for continuous optimization. It maps candidate solutions to food sources and objective-function values to source quality. Agents take on novice, experienced, search, or…
The article explains how FP-Growth can mine associations among binary features in historical trading data. Unlike Apriori, which repeatedly scans the database to evaluate candidate patterns, FP-Growth builds a tree representation and performs subsequent…
This article describes LightGTS, a time-series forecasting framework designed to handle datasets with different sampling scales and recurring periods. Its central method is adaptive periodic patching: estimate or otherwise determine a series’ cycle length,…
The article explains unsupervised learning and applies k-means clustering to trading data. Without labeled target values, clustering groups observations represented as feature vectors by assigning them to nearby centers. The cluster count is a model…
The article reviews the Ilan Expert Advisor, which averages into losing positions using a grid and increasing trade sizes, then closes the basket near its average entry price. It explains why this approach can perform during quiet, sideways markets yet face…
The article describes a Time-MoE forecasting architecture that represents each time step as a token, processes temporal context with transformer blocks, and predicts across multiple horizons. Its focus is the sparse mixture-of-experts component: a router…
The Turtle Shell Evolution Algorithm (TSEA) is a population-based optimization method that arranges candidate solutions in a shell-like structure. It groups solutions vertically by fitness and horizontally by location, with limited capacity in each cell.…
The article presents MetaTrader 5 as an environment for moving an AI trading idea from research into a testable Expert Advisor. It describes using terminal data in Python for analysis and feature preparation, exporting trained models through ONNX for use in…
The article presents orthogonal polynomials as a way to smooth financial price series and extract components associated with averages, trends, and nonlinear shapes. It outlines Legendre, Chebyshev, Laguerre, and Hermite families, describes mapping prices…
African Buffalo Optimization (ABO) is a population-based metaheuristic that represents candidate solutions as buffalo in a herd. Agents exchange information through signals modeled on social behavior, with each update drawing on both the best solution found…
The article presents the Hilbert-Schmidt Independence Criterion (HSIC) as a non-parametric test for dependence between data features and a target. Using kernel matrices, HSIC can detect nonlinear as well as linear relationships and can handle scalar or…
The article explains why directional indicator crossovers can generate repeated losing entries in range-bound markets, then presents a two-layer filter for ADX trades. The first layer replaces a fixed ADXR regime threshold with a gate optimized on validation…
The article describes Deterministic Oscillatory Search, a population-based metaheuristic for optimizing multidimensional objective functions without random numbers. Particles begin at systematically distributed positions and track whether movement improves…
The article proposes generating alternative classification targets directly from OHLC observations instead of relying only on future closing price. It calculates pairwise price midpoints, labels whether each candidate value rises over a chosen forecast…
The article describes Jardine's Gate, a sequential filter intended to decide whether an LSTM-generated cryptocurrency CFD signal should be traded. The six checks cover market structure through compression-based entropy, agreement among multiple LSTM experts,…
The article proposes a financial forecasting architecture inspired by the Hodgkin–Huxley neuron model. It maps market inputs such as prices, volume, indicators, and time features into a hybrid neural system, combining conventional neural layers with…
This article introduces Real-ORL, a framework for studying offline reinforcement learning with trajectories collected from real interactions. Rather than proposing a new learning algorithm, the cited work evaluates existing offline RL methods alongside…
This article studies stacking as a way to combine predictions from an ensemble of extreme learning machine classifiers. The base models produce outputs on separate training and evaluation samples; those outputs become features for a second-level combiner.…
This article explains Supported Policy Optimization (SPOT), an offline reinforcement learning method intended to reduce unreliable value estimates when a learned policy chooses actions that are poorly represented in its fixed training data. SPOT estimates…
The article presents a workflow for preparing labeled market time series and training an N-HiTS forecasting model with PyTorch Lightning and PyTorch Forecasting. It retrieves M15 price data through MetaTrader 5, converts it to a dataframe, and adds time and…