Built a 200k-edge market knowledge graph to filter false dip-buy signals

Details

External ID

46416991

Source

HN

Company

—

Product

—

Website domain

—

Launched

Dec. 29, 2025

Cohort

—

Upvotes

5

Upvotes percentile

0.10400763358778627

Tags

—

Fetched at

Sept. 7, 2026, 9:25 p.m.

Updated at

Sept. 7, 2026, 9:25 p.m.

Description

I’ve been experimenting with a graph-based approach to a classic trading problem: why most dip-buying strategies can’t tell the difference between a temporary overreaction and a genuine structural collapse.Most systems treat a −5% move the same regardless of context. My hypothesis was that where a company sits in the market’s structure matters more than the price move itself.The engineering ideaI built a knowledge graph of the U.S. public markets with ~207k edges across ~21 relationship types, organized into four layers:Operational: supply-chain relationships (SUPPLIES_TO, PRODUCES)Flow: ETF and institutional ownership plumbingSocial: board interlocks (SHARES_DIRECTOR_WITH)Environmental: geography / competitionFor each layer, I compute centrality scores using PageRank-style methods (with inverse-degree weighting to avoid ETF super-nodes dominating).These structural features are then combined with basic price/volume context and fed into a tree-based model (XGBoost) to rank stocks after sharp drawdownsWhat surprised meWhen I validated the rankings out-of-sample (2024–2025, using Alphalens to avoid look-ahead issues): * Operational and Flow edges provided most of the lift * Social edges (board interlocks) added much less than I expected * Graph features roughly doubled ranking quality versus price-only baselines This wasn’t obvious to me going in — I expected “social” connections to matter more.Why I’m postingI’m in the process of turning this from a research notebook into a production dashboard, and before I lock in the graph schema I’d love feedback from people who’ve built large graphs in other domains. In particular: * Have you seen board-interlock / social edges be predictive elsewhere? * Are there graph normalization tricks you’ve found essential at this scale? * Any pitfalls you’ve hit when mixing heterogeneous edge types?Happy to answer questions about the graph construction, centrality calculations, or validation setup.

• Source listing
• Website (no URL on record)