Machine learning for investment decisions
Practical guides on applying ML models, data pipelines, and quantitative methods to portfolio analysis and strategy research.
Where to start - and where to go next
These guides are written for learners who already understand the basics of finance or programming and want to connect those two worlds in a structured way. Each one focuses on a specific problem rather than a broad topic.
How a supervised learning model reads price history
Most tutorials skip the part where raw OHLCV data becomes a training matrix. This guide works through feature construction step by step - lag windows, rolling statistics, and the normalisation choices that affect model stability more than architecture does.
- Choosing a look-back window without leaking future data
- When to use z-score vs min-max scaling on financial series
- Structuring a walk-forward validation split
Portfolio variance isn't the whole story
Standard deviation treats upside and downside volatility equally. This guide covers semi-deviation, CVaR, and how ML models can be trained to minimise tail exposure rather than overall variance - with a worked Python example using real index data.
Sourcing alternative data on a constrained budget
Satellite imagery and credit card transaction feeds are out of reach for most learners. This guide maps the free and low-cost data sources that still carry genuine signal - earnings call transcripts, SEC filings, and public sentiment APIs - and explains how to preprocess each one.
Gradient boosting on financial tabular data
Tree-based models often outperform neural networks on structured financial data, but they require careful hyperparameter control. This guide covers early stopping, feature importance interpretation, and the specific overfitting patterns that appear when training on short time series.
Signal decay and why backtests mislead
A strategy that looked strong in 2019 data may carry no useful signal by 2023. This guide explains how to measure signal half-life, detect regime shifts in your training data, and build a simple monitoring layer that flags when a model's edge has likely eroded.
From model output to a realistic order
Predicted returns don't translate directly into position sizes. This guide covers Kelly-fraction position sizing, slippage assumptions for different asset classes, and how to structure a simple paper trading loop that logs execution quality alongside model predictions.