The Nonstationarity-Complexity Tradeoff in Return Prediction¶
Bibliographic record. Follow the original-source link for the publication.
| Field | Value |
|---|---|
| Primary domain | Financial Model Validation |
| Other domains | — |
| Methods | Financial Ml, Research Methods |
| Facets | — |
| Authors | Agostino Capponi, Chengpiao Huang, J. Antonio Sidaoui, Kaizheng Wang, Jiacheng Zou |
| Published | 2025-12-29 |
| Source | SSRN Financial Economics Network |
| Identifiers | doi:10.2139/ssrn.5980654, ssrn:5980654 |
| URL | Open original source |
Editorial synthesis¶
Why it matters¶
The paper links model complexity to how much training and validation history remains useful after distribution shifts. It motivates financial model-selection research, not a validated solution for option, quantile or ranking targets. (full_text:S33, full_text:S37, full_text:S98, full_text:S160, description:S6)
Main author claims¶
- The authors report: Under independent-batch and boundedness assumptions, the authors bound squared-error excess risk using approximation, estimation and distribution-shift terms. The tradeoff does not establish a universal preference for simpler models or shorter windows. (
full_text:S33,full_text:S98,full_text:S99,full_text:S160,full_text:S220) - The authors report: ATOMS compares candidates using pair-specific validation lookbacks and a randomized tournament. Candidates span model classes, hyperparameters and training windows. Expected linear comparison count does not include fitting the candidate grid. (
full_text:S37,full_text:S267,full_text:S269,full_text:S270,full_text:S286,full_text:S321) - The authors report: Across 17 U.S. industries, Table 3 reports zero-forecast OOS R² of 0.049 versus 0.043 for the strongest fixed-window benchmark, January 1990–November 2016. That is +0.006, about 14% relatively, not +14 percentage points; not replicated here. (
full_text:S40,full_text:S41,full_text:S387,full_text:S388)
Data, method, or discussion scope¶
The study combines 17 industries and 129 predictors, September 1987–November 2016, using daily models and monthly evaluation. Twenty past-month splits are not independent market histories. Q4 follows SSRN posting on December 29, 2025; the read 109-page revision is dated August 13, 2026. (full_text:S324, full_text:S335, full_text:S376, full_text:S377, description:S7, description:S8)
Main limitations¶
Theory relies on independence/boundedness; practical M′ uses a historical percentile. Costs use 4e-4/252 per turnover, not 4 bp per trade. Sharpe averages within-month daily-return ratios. Feature lags/vintages were not code-audited: PIT safety is unverified, not disproved. Preprint; no replication or independent proof verification. (full_text:S98, full_text:S99, full_text:S160, full_text:S466, full_text:S468, full_text:S471, full_text:S724, full_text:S727, full_text:S734, full_text:S745, full_text:S916)