UJDC Quantum-Finance Cockpit Designing out the dead-inventory & earnings-volatility trap
HKU ICB · Individual Assignment
David Kwan, CFO
The Pain
核心财金痛点audited dataEarnings instability + margin swing
Net profit (bars) vs gross margin (line), FY2022–FY2025
The dead-inventory tail
Finished goods by year of origin (audited Note H, 30/9/2025)
Industry Benchmark
上市同业对标genuine market dataHong Kong listed jewellery & watch peers vs UJDC
Market cap, revenue, net margin, equity beta & P/E — retrieved 2026-06-16 from stockanalysis.com & TradingView.
*Indicative value (not listed). †Beta is a listed-peer proxy (UJDC has no traded equity).
Aged-Stock Triage
呆滞存货分流AI action queueWhat the AI recommends the CFO do with the stock pile
Module C segments finished goods by age into an action queue — the dead tail is melt/redesign candidates, not carrying value.
Triage thresholds: ≥3 yrs → melt/redesign · 1–3 yrs → markdown & promote · <1 yr → hold. The CFO approves each batch (human-in-the-loop).
Live Investment Appraisal + ML Optimiser
投资评估与寻优CFO sliders → recomputeCFO control panel
Move the levers, or let the ML optimiser find the best settings. Optimised values appear as ✦ markers.
Decision rule: accept when IRR > WACC.
Project cash flows
Capex (Yr0) then ramped annual benefit; line = cumulative.
Efficient Frontier
有效前沿Markowitz · genuine betasFrontier of HK jewellery stocks
Single-index (Sharpe) model from genuine 1-yr betas; market σ assumed 22%.
Max-Sharpe (tangency) portfolio
Optimal long-only weights across the listed peers.
Risk Engine
风险引擎Monte-Carlo · hedgeMonte-Carlo input-price distribution
6-month gem/gold input price index — accept the distribution, not a point forecast (Quantum mindset).
Margin value-at-risk & hedge
One-sided 95% VaR on UJDC's FY25 precious-input book, before/after hedge.
Explainable-AI driver ranking · attribution ≠ causation.
State-of-the-Art: SHAP & Causal AI
前沿:SHAP 与因果 AIexplainability core(a) SHAP — attribution
Shapley values: a fair share of each prediction to each driver. Answers "what is the model using?"
φᵢ = Σ |S|!(M−|S|−1)!/M! · [f(S∪{i}) − f(S)] · attribution, not causation.
(b) Causal AI — intervention
Effect of an AI markdown policy on margin, confounded by inventory age. Answers "what if we ACT?"
Naive (SHAP-style) association is biased; back-door adjustment recovers the true do(·) effect.
AI Co-pilot
AI 副驾RAG · auto-tuneAsk about UJDC's finances — answers grounded only in retrieved audited data
Anti-hallucination by design (幻觉): the model uses only the data context and will refuse to invent figures. Ask it to optimise and it can auto-adjust the sliders above. The CFO holds the final decision.
Course Lens
课程理论应用Dr Li · M9Machine-Learning Stack & Code
机器学习栈与代码XAI transparencyModels behind the cockpit
Each module, its model and library — the same listing appears in Appendix A of the paper.
Code listings — SHAP, Causal AI, Optimiser
Production-representative code supporting the XAI and optimisation layers (Appendix D of the paper).