Project · Machine Learning · FinTech
Personalized Equity Discovery
A hybrid recommendation engine for retail equity discovery, modelling the personalization problem at a platform like Robinhood or Fidelity. The catalog holds 103 US-listed equities and ETFs. Five investor cohorts — each with distinct sector affinities and portfolio size distributions — are generated synthetically from published retail brokerage behavioral statistics.
Describe your investment style, preferred sectors, and risk appetite — the engine matches you to a similar synthetic investor and returns personalised recommendations in under 100 ms. Evaluation uses a temporal hold-out (last 20% of the timeline) to prevent look-ahead bias. “Investors similar to you also viewed” — not investment advice.
Configure & Run
How It Works
Implicit Feedback
No explicit ratings exist. Trades (weight 4), watchlist adds (2), and detail views (1) compose a confidence-weighted interaction matrix. All CF models treat unobserved items as weak negatives — the standard approach for brokerage and streaming platforms.
iALS & BPR
Implicit ALS learns user & item embeddings via weighted alternating least squares (Hu et al., 2008). BPR directly optimises ranking via pairwise SGD — the probability that an interacted security ranks above an unobserved one. BPR often outperforms iALS on NDCG when the interaction matrix is extremely sparse.
Content-Based & Hybrid
Each security is represented by a 14-dim feature vector: one-hot GICS sector (11 dims), market-cap tier, dividend yield tier, and ETF flag. The hybrid linearly fuses CF and CB scores (α=0.7), handling cold-start users and newly-listed securities that CF cannot rank.
Evaluation & Bias
Temporal split prevents look-ahead bias. Serendipity measures personalisation value beyond the popularity baseline. Intra-list diversity penalises sector concentration — a recommender surfacing AAPL, MSFT, NVDA, GOOGL, META to every user has near-zero discovery value regardless of precision.
FINRA Suitability
Under FINRA Rule 2111, broker-dealers have suitability obligations for investment recommendations. The critical distinction: discovery surfacing (“investors similar to you also viewed”) is a product feature; personalised advice (“you should buy this”) is regulated. All outputs use discovery language.