Temporal Drift in Accounting Concept Networks
Overview
A graph ETL and embedding alignment pipeline processing multi-quarter SEC XBRL financial statements. It converts raw filings into Pointwise Mutual Information (PMI) weighted concept co-occurrence networks, trains quarterly node2vec embeddings, and applies Procrustes rotation alignment to quantify structural drift over time.
Approach
- Engineered a 5-table canonical database schema and modular parser translating raw SEC XBRL datasets into standardized graph edges.
- Constructed PMI-weighted concept co-occurrence graphs for each quarter and executed orthogonal Procrustes alignment across sequential node2vec embedding spaces.
- Implemented a 5-metric evaluation suite (embedding displacement, edge overlap, degree distribution shape, vocabulary churn, and community detection).
Results
Metrics — sourced from the repo (README / model card / test output)
- Embedding drift, Q1→Q2 vs. Q2→Q3
- 0.431 → 0.373
- Vocabulary churn, Q1→Q2 vs. Q2→Q3
- 23.0% → 12.1%
- Centrality → stability regression
- R² = 0.25, n = 3,935 concepts
- Zero-drift sanity check
- <0.1 vs. 0.43 real
- Test suite
- 23 test files, built test-first
5 independent metrics agree drift is decelerating
real drift measured at ~4–5x the alignment-noise floor
Proved systemic decelerating structural drift across sequential financial quarters (embedding distance 0.431 → 0.373) and revealed counter-intuitive drift patterns in core GAAP concept pairs, validated against a zero-drift sanity check built into the test suite.