Hedging-Partner Recommender
A stock recommender that keeps the investor's taste but, among the stocks they would plausibly buy, prefers the ones that lower the risk of the portfolio they already hold. Each example is a real case, either a retail customer's purchase from the FAR-Trans dataset or a fund manager's quarter-end book from SEC Form 13F filings: the actual holdings, the collaborative-filtering (CF) favourite, and the hedging partner our risk gate picks from the same top list. Class project for Deep Learning Recommender Systems (HUFS, fall 2026).
Setup
- Retail data. FAR-Trans (CC BY 4.0): buys and sells of a European bank's retail customers, 2018–2022, with daily closes. Holdings are rebuilt from the transactions, valued at the last close before the purchase day, and include funds and bonds. The closes are not adjusted for corporate actions, so a daily move beyond the Athens Exchange's ±30% limit (e.g. Attica Bank's 2021 rights issues) counts as no change, and price histories are back-adjusted the same way.
- Institutional data. SEC Form 13F holdings at quarter ends, within the 300 most widely held US common stocks (CUSIP → ticker via OpenFIGI, Yahoo Finance adjusted closes). 13F has no MiFID profile, so the risk level is the quartile of the holdings' predicted volatility.
- Taste. WMF, trained on FAR-Trans purchases before 2022-02-09 (94 Athens-listed stocks) or on 13F holdings up to 2021Q4 (300 US stocks); the stocks the investor does not hold are ranked by its score.
- Risk. Ledoit–Wolf covariance of the 250 daily log returns before the purchase day (13F: up to the quarter end). Adding a stock at weight a changes the portfolio variance to (1−a)²σ²P + 2a(1−a)σP,i + a²σ²i; the default weight is 1/(n+1) for n holdings.
- Gate. The top-5 list may only change for clear gains: a stock ranked 6–10 by taste can enter if it cuts predicted volatility by at least 2%, replacing one that does not. The hedging partner is the list's most risk-reducing stock other than the CF favourite.
- Check. “기간별 리스크 변화” is realized, not predicted: the volatility over the following trading days with the stock added at the same weight, against holding nothing new.
- Examples. Selected, not representative, three per risk level with the largest predicted gap between the CF favourite (which would raise risk) and the partner. FAR-Trans: test purchases (May–Sep 2022) by customers with 3–8 priced holdings, partner cutting risk by 5%+. 13F: test quarter ends (Jun, Sep 2022) with 10+ holdings. A manager holds 45 universe stocks at the median, so one stock at weight 1/(n+1) seldom moves risk by the gate's 2%; the gate acts mostly for smaller books (13 holdings at the median among the cases with a partner).