Modern Machine Learning Tools in Finance: A Critical Perspective
A survey argues most machine learning cannot sustain trading profits because markets adapt. It maps an arc of discovery, early gains, crowding, crash, and decay, citing a 58 percent post publication alpha drop, the August 2007 quant crisis, and COVID regime whiplash. Contributions include a bridge between finance and ML, architecture level stress tests, and a validation playbook. The core claim is reflexivity limits any model. Models shine in niches but often fail across breaks.
What it examines
The paper bridges machine learning and finance to explain why ML trading fails. Using the Adaptive Market Hypothesis and reflexivity, it reviews modern non-stationary tools (PatchTST, non-stationary transformers, neural HMMs, RL) and documents strategy decay. It aims to assess limits, compare methods, and propose realistic validation.
What it concludes
Findings: no model reliably beats adaptive markets; success decays as strategies change markets. The authors recommend continuous learning, fast regime detection, dynamic strategy mixing, meta-learning, and adversarial robustness. Uses include risk management, execution, and niche or less efficient markets. Limits: costs, data, novelty. Future work: evolutionary, multi-agent, causal, rigorous benchmarks.
Evidence objects
Survey delivers verdict: adaptive markets erode ML profits via reflexivity, with a lifecycle of discovery, early gains, crowding, crash, decay; evidence includes 58% drop, August 2007 quant crisis, COVID whiplash.
key_findings bullet 1 · key_findings · validation V0
Three contributions: bridge aligning finance and ML; a stress test of PatchTST, non-stationary transformers, neural HMMs, TFT, MoE, online/meta-learning, RL; and a validation blueprintregime tests, adaptation latency, costs/impact, buy-and-hold benchmark.
key_findings bullet 2 · key_findings · validation V0
Findings: models succeed in niches but fail at structural breaks; reflexivity constrains design. Strengths: continuous learning, regime recognition, meta-learning, adversarial robustness. Limits: subjective scoring, scarce tests, underexplored markets, execution edges.
key_findings bullet 3 · key_findings · validation V0
This critical survey uniquely bridges finances AMH/reflexivity with machine learning limits in trading, reframing failures as fundamental rather than technical. It synthesizes non-stationary architectures (PatchTST, non-stationary transformers, neural HMMs, RL), integrating disparate literatures. Though offering no new algorithms or experiments, its perspective clarifies boundaries, aligns research agendas, and tempers expectations.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
This paper discusses why machine learning approaches consistently fail to generate profitable trading strategies, viewing this challenge through the lens of
Source row: 1373 · abstract type: snippet