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 develops binary classifiers to estimate whether a currency pair will rise or fall. It explains how logit and probit models map a linear combination of lagged price changes to a probability, and describes forming and standardizing a feature set…
The article introduces Kohonen self-organizing layers as a way to group input vectors into regions of similar patterns. It explains winner-take-all selection, in which the neuron with the closest weight vector responds, and adjusts that winner toward the…
The article describes an MQL5 Wizard signal class that combines an Unscented Kalman Filter (UKF) with a Capsule Network. The UKF estimates a hidden, potentially nonlinear price state from noisy observations, aiming to avoid some lag associated with…
The article demonstrates parameter-efficient LoRA fine-tuning of a pre-trained GPT-2 model on financial data, following an earlier installment on full-parameter fine-tuning. It explains how to configure LoRA with the PEFT library, including adapter rank,…
The article presents Anarchic Society Optimization (ASO), a population-based optimization method inspired by decentralized social behavior. Each candidate solution retains its current, previous, and personal-best positions while the population tracks a…
The article explains how the Mantis time-series classification model turns financial data into a fixed sequence of tokens for transformer analysis. It describes convolutional feature extraction, division into patches, and pooling to summarize each patch. The…
The article explains a deep deterministic policy gradient agent for a trading workflow that combines supervised-learning outputs with reinforcement learning. Its agent uses actor and critic networks, matching target networks, separate optimizers, and a…
The article explains the UT Bot Alerts concept and describes an MQL5 implementation designed for automated use. It derives a trailing-stop distance from ATR multiplied by a user-selected factor: a wider distance generally reduces signal frequency, while a…
The article explains transfer learning as reusing a trained model, or its early layers, as the starting point for a model aimed at a new task. Reusing learned weights can reduce training effort compared with starting from random initialization. The author…
The article explains random forests as ensembles of decision trees trained on resampled data and randomized feature subsets. For classification, it combines tree outputs by majority vote; for regression, it averages predictions. It also describes extending a…
The article builds a Kolmogorov-Arnold Network (KAN) that represents each connection as a learnable univariate curve. It uses cubic B-splines as basis functions, with curve coefficients fitted through ridge-regularized least squares. The network predicts the…
The article outlines a proposed foreign-exchange triangular arbitrage system that combines three currency pairs with neural-network forecasts. It describes the basic conversion cycle and the idea of comparing implied exchange rates, then presents a workflow…
The article connects monoid actions, which let operations transform values in a separate set, to a machine-learning system that forecasts changes in price-bar range for adjusting trailing stops. It discusses feature-importance methods, including tree-based…
This article describes updates to an MQL5 trading assistant that separates user interface components into a reusable include module and adds automated signal handling. The proposed flow checks for signals on new bars or by user request, parses AI responses…
This article introduces deep Q-learning as a reinforcement learning method that uses a neural network to estimate action values from an environment’s states. It explains that the agent learns from state, action, reward, and next-state transitions, and…
This article explains Atmosphere Clouds Model Optimization, a population-based metaheuristic inspired by cloud formation and movement. It maps a multidimensional search space into regions, each with humidity and pressure values. Clouds are created in…
This article explains Atmosphere Clouds Model Optimization, a population-based metaheuristic inspired by cloud formation and movement. It maps a multidimensional search space into regions, each with humidity and pressure values. Clouds are created in…
This article introduces Brain Storm Optimization (BSO), a population-based search method intended to locate several promising solutions in a multimodal optimization problem. Candidate solutions are grouped by similarity using K-means; the algorithm then…
The article describes a money management method that uses a Fenwick tree to organize recent volume data and compare volume across portions of a lookback window. It outlines four proposed sizing modes—linear, conservative, aggressive, and mean-reversion—and…
This article outlines a neural-network approach to forecasting financial prices, emphasizing data preparation and input representation. It recommends training on quote changes or log returns rather than raw price levels to reduce dependence on price scale…
This article compares feature importance methods for tree based trading models, with particular attention to correlated or duplicated predictors. Mean decrease impurity divides credit among interchangeable features, while mean decrease accuracy can…
The article explores a double moving average crossover approach that uses a higher timeframe to establish directional bias and a lower timeframe to time entries. It contrasts trend-following entries, which wait for a lower-timeframe crossover in the…
The article explains the Structured Component Neural Network (SCNN), an architecture for forecasting multivariate time series by separating persistent trends, seasonal patterns, short-term changes, co-evolving behavior, and residual noise. It describes a…
This article describes reward decomposition for reinforcement learning, with an implementation in a Soft Actor-Critic variant using the DICE method. Rather than learning one value function from a composite reward, the agent learns a value estimate for each…