# RiskHedge

Fail-closed screening for financial risk decisions: accept a cheap risk computation only when a certificate shows that it picks the same hedges as exhaustive evaluation. A QRS@NTU research project; paper by Yingzhi Tang and Yuhe Sui (Quantitative Research Society, Singapore; The Chinese University of Hong Kong, Shenzhen; Nanyang Technological University).

- Overview: https://qrsntu.org/projects/riskhedge/
- How it works: https://qrsntu.org/projects/riskhedge/details/
- Companion paper (OpenReview): https://openreview.net/forum?id=OzOTmkkAWW

Plain-text summary for AI agents and readers, generated from the evidence file behind the pages; the pages are the full account. This is an academic study on synthetic data, not investment advice.

## Design

A risk engine chooses the 3 least risky of 24 hedge packages by expected shortfall over 64 market scenarios. The usual shortcut ranks the packages from cheap approximate prices and accepts the result, so near the cut-off a small error can silently change which hedges are chosen. RiskHedge keeps the cheap computation but accepts its answer only when a certificate shows that exhaustive evaluation would choose the same set: each package has a risk interval, and the set is certified when every chosen package is better at its worst than every rejected package at its best (ties broken by ID). If the intervals overlap, it pays for up to 8 more exact option-pricing primitives and tests again; failed guards (price bounds unavailable, stale for the current input, or a call outside the registered scope) or an exhausted budget fall back to exact evaluation. The guarantee holds when the price bounds are valid for the current input and the call is in scope.

## Key results

- Paper: Certify the Decision, Not the Shortcut: Fail-Closed Screening for Financial Risk Decisions, by Yingzhi Tang and Yuhe Sui; Financial AI Security, Privacy, and Safety Workshop (FinAI-SPS'26), ICAIF 2026, Milan (accepted as a poster; non-archival).
- Pre-specified confirmation on 200 fresh instances from one synthetic Merton option-hedging pipeline, deliberately enriched near the decision boundary: the gate made 0 wrong or unsupported decisions at a median 17.2% of exhaustive pricing work (maximum 0.350 of it); a midpoint ranking with no exact work was wrong on 10/200 (5.0%, Wilson 95% interval 2.7-9.0%), all in the near-tie, wide-bound group (pre-specified Fisher test, p = 0.016).
- 60 injected guard faults were each rejected by the intended guard: 40 fell back exactly and 20 stale-input runs recomputed and certified.
- Post hoc: an uncertified midpoint rule given the same total work as the gate (8,600 against 8,518 primitives) was still wrong on 11; zero errors first appeared at 1.88x the gate's work among the budgets tried (in-sample).

## Limitations

- All evidence comes from one synthetic Merton pipeline with 24 candidates, 64 scenarios and a top-3 choice; there is no real portfolio data. The confirmation population was enriched near the decision boundary, so the rates are study-population rates, not deployment error rates.
- Work counts only the exact primitives paid after the initial price bounds. Building those bounds is not counted and no wall-clock time was measured, so the cost of the cheap certified intervals and any speedup are not established.
- The guarantee needs valid current-input bounds, inherited from the certified pricer and not proved in the paper. Post hoc, corrupted bounds with valid metadata gave 13 false certificates, all of which passed reference-free replay; with a learned surrogate the guarantee did not apply (simultaneous coverage on only 95, 97 and 149 of 200 instances).
- Work-matched comparisons, corrupted bounds, the sampled audit, the refusal path and the learned surrogate are post hoc; the budgets in the work-matched comparison were read off after the results.
- Replay checks the internal consistency of the recorded decision path and was written by the same team; it is not formal verification.

## Not claimed

- No claim of a deployment error rate.
- No claim of a speedup or wall-clock saving.
- No claim of the cost of building the price enclosures.
- No claim of a guarantee for learned surrogates.
- No claim of adversarial robustness or security.
- No claim of privacy.
- No claim of fraud or AML detection.
- No claim of regulatory compliance.
- No claim of results on real portfolio data.

## A question to start with

Explain the idea to a newcomer, then give the three strongest criticisms and one experiment that would settle each.