Third-Party Research & Methodology Only

This section shares summaries of third-party academic research and descriptions of quantitative models. The content represents the findings of the original researchers, not the opinions or recommendations of Foxholm Financial. Foxholm Financial does not publish hypothetical or backtested performance metrics on its quantitative research pages. All content is restricted to methodology, signal construction, factor logic, and risk architecture. SEC rules require that investment advisers not present misleading performance data, and our methodology-only approach reflects that standard and the firm's fiduciary obligations.

Machine Learning in Finance

Computational Method Signal Generation Cross-Disciplinary

Machine learning in finance refers to the use of algorithms that learn patterns from data to make predictions, classify outcomes, or optimize decisions without being explicitly programmed for each task. It draws on statistics, computer science, and domain-specific financial knowledge.

The field has grown rapidly since the early 2010s as computing power, data availability, and open-source software have lowered the barriers to entry. Applications range from predicting stock returns and detecting credit fraud to extracting sentiment from news articles. While powerful, these methods carry unique risks in financial settings because markets are noisy, non-stationary (meaning their statistical properties change over time), and shaped by human behavior that can shift abruptly.

Definition

Machine learning (ML) is a branch of artificial intelligence in which algorithms improve their performance on a task by learning from data rather than following hand-coded rules. In finance, the "task" is typically prediction (forecasting returns, volatility, or default risk), classification (labeling transactions as fraudulent or legitimate), or optimization (constructing portfolios that balance risk and return).

Core Idea

A machine learning model identifies patterns in historical data and uses those patterns to make decisions about new, unseen data.

The model "learns" by adjusting its internal parameters to minimize errors on a training dataset. It is then tested on separate, held-out data to see whether the patterns it found generalize beyond the sample used to build it. This train-test separation is critical in finance, where overfitting (fitting noise rather than genuine signal) is a persistent danger.

Unlike traditional statistical models, which typically start with a hypothesis about the relationship between variables, many ML techniques are data-driven. They can discover complex, nonlinear relationships that a researcher might not think to specify in advance. This flexibility is both a strength and a vulnerability: the same capacity to find subtle patterns also makes these models prone to fitting random noise in financial data.

Common Methods

Machine learning methods are generally grouped into three categories based on how the algorithm learns from data. Each category has different strengths and typical use cases in finance.

Category	How It Works	Example Techniques	Financial Applications
Supervised Learning	The algorithm learns from labeled examples (input-output pairs) to predict outcomes on new data	Linear regression, logistic regression, decision trees, random forests, gradient boosting, and neural networks	Return prediction, credit scoring, default probability estimation
Unsupervised Learning	The algorithm finds structure in data without predefined labels or target variables	Clustering (k-means), principal component analysis (PCA), autoencoders	Regime detection, asset grouping, dimensionality reduction for large factor sets
Reinforcement Learning	An agent learns by interacting with an environment, receiving rewards or penalties for its actions	Q-learning, deep reinforcement learning, policy gradient methods	Order execution optimization, dynamic portfolio rebalancing, market making

Supervised learning is the most widely used category in quantitative finance. A supervised model might take in hundreds of features (financial ratios, price momentum, macroeconomic indicators) and learn which combinations best predict whether a stock will outperform over the next month. Random forests and gradient boosting machines are popular because they handle nonlinear relationships and noisy data reasonably well.

Neural networks, a subset of supervised learning, use layers of interconnected nodes loosely inspired by biological neurons. Deep learning (neural networks with many layers) has shown strong results in natural language processing and image recognition, but its advantages in financial return prediction remain debated due to the low signal-to-noise ratio in market data.

Applications in Finance

Machine learning is used across many areas of finance. The most common applications fall into four broad categories.

Signal Generation and Return Prediction

ML models can process large numbers of predictive features simultaneously to generate alpha signals (indicators of potential excess return). Traditional factor models might use five or ten variables; ML models can evaluate hundreds or thousands. The challenge is separating genuine predictive power from statistical noise. Gu, Kelly, and Xiu (2020) found that neural networks and tree-based methods improved out-of-sample return prediction compared to linear models, though the improvements were modest in magnitude.

Risk Modeling and Credit Scoring

Banks and insurers use ML to estimate default probabilities, assess creditworthiness, and model portfolio risk. Gradient boosting and random forests can capture nonlinear interactions between borrower characteristics that traditional logistic regression may miss. These models are widely used in consumer lending, commercial credit, and insurance underwriting.

Natural Language Processing (NLP)

NLP techniques extract structured information from unstructured text: earnings call transcripts, Securities and Exchange Commission (SEC) filings, news articles, and social media. Sentiment analysis models can quantify the tone of a CEO's language on an earnings call or flag unusual phrasing in regulatory filings. These text-based signals are often combined with traditional financial data to build composite trading signals.

Fraud Detection and Compliance

Classification algorithms identify suspicious patterns in transaction data. A model might flag a credit card transaction as potentially fraudulent based on the time, location, amount, and merchant type, comparing it against the cardholder's historical behavior. These systems must balance detection accuracy against false positives, since incorrectly flagging legitimate transactions has its own cost.

Challenges and Pitfalls

Financial data presents unique challenges for machine learning that do not arise in many other domains. The most important ones involve overfitting, changing market conditions, and the need for model transparency.

Overfitting and Data Snooping

Overfitting occurs when a model learns the noise in historical data rather than genuine, repeatable patterns. Financial datasets are particularly vulnerable because the signal-to-noise ratio is very low: most of the variation in asset returns is random. A model with thousands of parameters and a limited sample of monthly returns can easily find patterns that look compelling in hindsight but have no predictive power going forward.

Data snooping (also called p-hacking or multiple testing) compounds the problem. When researchers test many model configurations on the same dataset, some will appear to work well by chance alone. Harvey, Liu, and Zhu (2016) documented how the proliferation of published "factors" in academic finance is partly a product of this effect.

Non-Stationarity

Financial markets are non-stationary, meaning their statistical properties (averages, variances, correlations) change over time. A model trained on data from a low-interest-rate, low-volatility environment may perform poorly when rates rise or a crisis occurs. Unlike image recognition, where a cat looks the same in 2010 and 2025, the relationship between financial variables shifts as regulations change, new instruments appear, and investor behavior evolves.

Interpretability and Regulation

Many ML models, especially deep neural networks, are "black boxes" that do not explain why they made a particular prediction. This creates problems in regulated industries. Lending decisions, for example, are subject to fair lending laws that require lenders to explain why an application was denied. Portfolio managers and risk officers also need to understand why a model is recommending a particular trade or flagging a particular risk, not just accept its output at face value.

Known Limitations

Limitations to Keep in Mind

Low signal-to-noise ratio in financial data. Asset returns are dominated by randomness. ML models that excel in domains with strong signals (image recognition, language translation) often struggle with financial prediction because the patterns are weak and unstable.
Survivorship and selection bias. Training data often excludes failed companies, delisted stocks, or discontinued funds, which inflates the apparent performance of strategies tested on that data.
Regime dependence. Models trained on one market regime (bull market, low rates) may fail in a different regime (bear market, rising rates). Financial relationships are not fixed the way physical laws are.
Feedback loops and crowding. When many market participants use similar ML models and similar data, their trades can become crowded. This can erode the very patterns the models were designed to exploit, or create new sources of systemic risk.
Computational and data costs. Training sophisticated models requires substantial computing resources and high-quality data. Alternative data sources (satellite imagery, web scraping, credit card transactions) can be expensive and may raise privacy or ethical concerns.
Lack of causal understanding. ML models identify correlations, not causes. A model might discover that a particular combination of variables predicts returns, but without understanding why, there is no way to assess whether the relationship will persist.

Academic Context

The application of machine learning to asset pricing gained significant academic attention with Gu, Kelly, and Xiu's 2020 paper "Empirical Asset Pricing via Machine Learning," published in the Review of Financial Studies. The authors conducted a comprehensive comparison of ML methods (including neural networks, random forests, and gradient boosting) against traditional linear models for predicting individual stock returns. They found that ML methods, particularly neural networks and tree-based models, produced meaningfully better out-of-sample predictions than linear approaches.

Marcos López de Prado's 2018 book Advances in Financial Machine Learning addressed the practical pitfalls of applying ML to finance. López de Prado emphasized the dangers of backtest overfitting, the importance of proper cross-validation techniques for time-series data, and the need for combinatorial purging to prevent information leakage between training and test sets. The book argued that many published ML-based trading strategies fail in live trading because they were overfit to historical data.

More broadly, the intersection of ML and finance connects to the factor investing literature. Traditional factor models assume linear relationships between a small set of characteristics and expected returns. ML methods relax this assumption, allowing for nonlinear interactions among potentially hundreds of features. Whether this added complexity translates into durable investment performance remains an open question in the academic literature.