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
The article introduces feedforward neural networks through the perceptron, explaining how input, hidden, and output layers pass information forward. Each neuron combines weighted inputs and applies an activation function; the author describes sigmoid and…
The document explains HyperDet3D, a point-cloud object detector that uses a hypernetwork to adapt Transformer decoder parameters to scene context. It separates learned scene-invariant knowledge, shared across training scenes, from scene-specific knowledge,…
The article extends a reinforcement learning trading system based on random decision forests. It addresses overfitting by selecting price increment features and tuning regularization against out-of-bag classification error, then uses Monte Carlo sampling in…
The article explains Feature Aggregated Queries (FAQ), a Transformer technique for video object detection, and considers how temporal features could help models interpret changing market states. Rather than averaging randomly initialized queries from…
The article explains statistical causality in time series and presents transfer entropy as a way to measure whether past information from one series improves understanding of another series’ future. It contrasts this approach with Granger causality, which…
This article ports causal trend-scanning features to MQL5 and replaces repeated window summation with three running statistics per horizon. Once each horizon is initialized, its sums update in constant work per bar, reducing the computation described from…
This article surveys the ALGLIB numerical analysis library as adapted for MQL5, describing its scope and major updates from version 3.5 to 3.19. The material highlights methods relevant to quantitative work, including interpolation and fitting, optimization,…
This article explains how to filter financial time series in the frequency domain: preprocess a finite series, apply a fast Fourier transform, modify frequency components with a filter response, then invert the transform and undo preprocessing. It describes…
The article demonstrates a MetaTrader 5 workflow for using neural network modules (NNMs) in an automated EURUSD trading system. It describes a module containing several subnetworks trained on different segments of a time series, with a signal line derived…
This article explains why regressions between unrelated non-stationary time series can appear convincing. It generates two independent random walks, regresses one on the other, and shows that fit statistics and coefficient significance can falsely suggest a…
The article describes HimNet, a neural forecasting architecture that adapts its recurrent graph-processing weights to temporal and spatial context. Trainable embeddings represent time patterns and individual tickers or venues, then query compact…
The article describes a brute force method for searching price patterns and generating trading robots from the selected parameters. It represents a signal as a multivariable polynomial expansion whose inputs are candle price changes across several bars.…
This article traces MQL and MetaTrader through early platform releases to the modern development environment. Its central theme is the steady expansion of what traders can do inside the platform: automate strategies, test them on historical data, create…
The article explains covariance matrix adaptation evolution strategy (CMA-ES), an evolutionary optimizer that samples candidate solutions from a distribution whose mean, step size, and covariance adapt over successive generations. Successful candidates…
The article presents practical uses of Kohonen self-organizing maps in algorithmic trading. For selecting expert-advisor settings, it recommends looking for broad plateaus of acceptable objective values rather than isolated backtest maxima, treating plateau…
The article applies self-organizing Kohonen maps to two tasks: identifying robust expert-advisor settings from optimization results and forecasting time series. For parameter selection, it argues that traders should seek broad regions, or plateaus, of…
The article explains how N-BEATS forecasts time series and uses constrained model blocks to represent components such as trend and seasonality. Each block maps an input sequence through a lower-dimensional representation and produces both a forecast and a…
The article introduces formulaic alphas as algebraic representations of trading signals. It argues that expressing signals as formulas can make large collections of machine-generated ideas easier to inspect, combine, and audit. This is presented as a…
The article presents an architecture for passing market data from an MQL5 Expert Advisor to a language model through a Python server. The EA acts as a data collector and receives a structured response, while the server calculates technical indicators,…
The article describes ETARE, a proposed trading system that combines an LSTM model, reinforcement learning, and an evolutionary population of strategies. Strategies carry neural network weights, which can be recombined or mutated; periodic selection removes…
The article explains the Ecological Cycle Optimizer (ECO), a population-based metaheuristic that models candidate solutions as ecosystem roles. The population is divided into producers, herbivores, carnivores, and omnivores: the best candidates act as…
This article explains Fully Parameterized Quantile Functions (FQF), a distributional reinforcement-learning method for estimating reward distributions in Q-learning. Earlier approaches use fixed reward categories or predetermined quantiles, which require…
The article describes the Multi-Agent Self-Adaptive (MASA) framework for dynamic portfolio management. It assigns return generation to a TD3 reinforcement-learning agent, risk adjustment to a separate optimization-based controller, and market analysis to an…
The article describes a live drift monitor for an ONNX trading model running in MetaTrader 5. It compares a frozen reference sample with a rolling live sample for each input feature, using the one-dimensional Wasserstein-1 distance: sort equal-sized samples,…