Machine Learning for Investing: A Survey
By 2026, the application of Machine Learning (ML) to investing has shifted from "phenomenological" pattern matching to Scientific Causal Discovery. The focus is no longer just "What will the price be?" but "What is the causal mechanism driving the return?"
1. Causal Factor Investing (The 2025 Paradigm)
The "Factor Zoo"—the explosion of thousands of reported market factors—has been largely debunked as a "Factor Mirage" due to p-hacking and selection bias.
A. The 7-Step Causal Protocol
Following the research of Marco Lopez de Prado (2025), leading quant shops have adopted a Causal ML workflow:
- Causal Graphing (DAGs): Mapping variable interdependencies using Judea Pearl's framework.
- Do-Calculus: Using graph surgery to select control variables and avoid "collider bias."
- Double Machine Learning (DML): Using ML to estimate treatment effects while controlling for high-dimensional confounders.
2. Advanced Portfolio Optimization: HRP 2025
The most significant theoretical breakthrough in 2025 was the formal validation of Hierarchical Risk Parity (HRP).
A. Overcoming Markowitz's Instability
In Risk Magazine (Jan 2025), Lopez de Prado and colleagues provided the first analytical proof that HRP is significantly less noisy than classical Mean-Variance Optimization (MVO).
- The update: We now have derived analytical values for the "noise" of allocation weights coming from estimated covariance matrices.
- Finding: HRP's tree-based clustering approach effectively regularizes the covariance matrix, preventing the "Optimizer's Curse" (where small errors in expected return lead to massive, unstable bets).
3. Position Sizing: Bayesian Neural Networks (BNNs)
Traditional ML provides point estimates (e.g., "The return will be 5%"). Advanced 2026 models use Bayesian Neural Networks to provide a probability distribution:
Where W are the weights and D is the data.
- Uncertainty-Aware Sizing: Quants use the Epistemic Uncertainty (model uncertainty) to size positions. If the model predicts a high return but has high variance in its weights, the position is automatically downsized.
4. Supply Chain Risk: Graph Neural Networks (GNNs)
2025 research has successfully applied GNNs to model systemic financial risk.
- Adjacency Matrices: Companies are nodes; edges represent supply chain dependencies, co-ownership, or board intersections.
- Risk Propagation: If a tier-2 supplier in Taiwan faces a "Hormuz Shock" precursor, the GNN propagates the risk through the graph, identifying vulnerable tech firms in the US before their quarterly earnings are impacted.
5. Execution and Tactical Allocation (RL)
Reinforcement Learning (RL) has moved beyond toy models to dominate Execution and Tactical Allocation.
- SAPPO (Sentiment-Augmented PPO): Kirtac & Germano (July 2025) demonstrated that agents using LLM-derived sentiment in their advantage function achieve Sharpe Ratios of 1.90 vs. 1.55 for traditional DRL.
- Hierarchical RL: A "Master" agent decides the asset allocation (Medium-Term), while "Sub-agents" execute the trades at the micro-level to minimize market impact.
6. Case Study: The 2026 Iran War Shock
During the February 2026 kinetic escalation, traditional Markowitz-style optimizers failed due to the sudden "Correlation Breakdown" where all assets dropped in tandem.
The ML Alternative: Portfolios using Hierarchical Risk Parity and Minimum Sentiment Connectedness (MSC) maintained their risk diversification. By identifying "Systemic Chokepoint" sentiment 72 hours before the Hormuz escalation, these models automatically shifted from high-beta tech to defensive commodities and energy, mitigating the -15% index drop.
See Also:
