---
title: When Backtests Lie: A Governance-Grade Validation Framework for Machine-Learning Value-at-Risk Models under Regime Shifts
url: https://www.ml-quant.com/papers/ssrn/7652318/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-10-23
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: SSRN 7652318
source_url: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7652318
featured: 2026-10-23
citations: unknown
topic: Risk, Credit & Banking
---


# When Backtests Lie: A Governance-Grade Validation Framework for Machine-Learning Value-at-Risk Models under Regime Shifts

The study shows that machine-learning VaR models passing standard backtests breach risk limits two to nine times worse in stressed regimes, revealing false comfort in current regulatory validation.

- Source: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7652318
- Identifier: SSRN 7652318
- Released: 2026-10-22
- First featured: Quant Letter No. 136 (2026-10-23): https://www.ml-quant.com/issues/2026-10-23/
- Citations (Semantic Scholar): not tracked
- Published in: not yet
- Topic: Risk, Credit & Banking
- Authors: Niraj Neupane

## Related

- [E-Backtesting Expected Shortfall without Value-at-Risk Forecasts under Time-series Models](https://www.ml-quant.com/papers/arxiv/2610.17681/): Sequential backtests for Expected Shortfall using only observed losses and ES forecasts control false rejections, with time-series extension under AR-GARCH structure proving convergent under grid refinement.
- [A note on subadditivity of value at risks (VaRs): A new connection to comonotonicity](https://www.ml-quant.com/papers/doi/10-1017-jpr-2025-31/): The research reveals a new characteristic of value at risk (VaR), stating that its subadditivity holds for any confidence level only if the loss random variables are comonotonic.
- [Quantiles under ambiguity and risk sharing](https://www.ml-quant.com/papers/arxiv/2412.19546/): The study introduces Choquet Expected Shortfall, a new class of risk measures, and provides optimization algorithms and examples using financial data.
- [Time-Series Foundation AI Model for Value-at-Risk Forecasting](https://www.ml-quant.com/papers/arxiv/2410.11773/): The research highlights the superior performance of a time-series model, TimesFM, in forecasting Value-at-Risk, with fine-tuning further enhancing the results.
- [Optimal insurance design with Lambda-Value-at-Risk](https://www.ml-quant.com/papers/arxiv/2408.09799/): The paper studies optimal insurance solutions using the Lambda-Value-at-Risk model, revealing that a truncated stop-loss indemnity is ideal under certain conditions and discusses the effect of model uncertainty.
- [Adaptive Multilevel Stochastic Approximation of the Value-at-Risk](https://www.ml-quant.com/papers/arxiv/2408.06531/): The paper introduces a multilevel stochastic approximation algorithm that adaptively selects the number of inner samples to compute the value-at-risk of a financial loss, improving the previous scheme's complexity.