Journal of Theoretical Physics & Mathematics Research
ISSN: 3065-8802 | DOI: 10.64030/3065-8802
Terry Ding
Financial machine learning aspires to convert high dimensional market data into repeatable predictive advantage, yet reported backtest performance often fails to translate to live deployment. This paper argues that the resulting “generalization crisis” is less a limitation of computational capacity than a limitation of epistemic and statistical rigor. By synthesizing evidence from over 150 primary sources and drawing parallels to the replication crisis in psychology and to Feynman’s cautionary notion of “cargo-cult” scientific form without corresponding self-correction, we show that several common research practices are systematically misaligned with the causal and non-stationary structure of markets. We identify three recurrent sources of failure: the information-theoretic redundancy of purely price-derived technical indicators, the compression of high-dimensional fundamentals into low-dimensional linear ratios, and the inappropriate application of IID validation protocols to dependent time series. We further analyze failure modes in Large Language Model (LLM) pipelines arising from look-ahead bias and trainingdata contamination. We conclude that progress requires a pragmatic “causal turn” toward Automatic Feature Engineering (AFE), Invariant Risk Minimization (IRM), and Combinatorial Purged Cross-Validation (CPCV), aligning modeling practice with the realities of non-stationarity, dependence, and implementability.