Strategic Impact Mapping: A Directed-Network Framework for S&P 500 Crisis Regimes

Reproducible code for an end-to-end directed-network pipeline that turns the daily S&P 500 return panel into a sparse directed "strategic impact map" and a single composite stress reading, evaluated across twenty regime snapshots over 1985–2024.

Status: author preprint, submitted to IEEE Transactions on Network Science and Engineering (TNSE). The compiled manuscript and its 75-page companion appendix are in paper/ (paper.pdf, appendix.pdf).

---

What this is

A single panel-scale workflow that integrates four ingredients prior S&P 500 work uses only in part, run per snapshot on the survivorship-free, historical- constituent CRSP panel (point-in-time crsp.msp500list membership, PERMNO- indexed; 1,442 PERMNOs, 655–889 alive per snapshot, 10,080 trading days; the 2008-era distress exits LEH/BSC/MER/FNM/FRE/WAMU enter with their full pre-bankruptcy trajectory):

Stage	Does	Key methods
1. A-DCC GARCH	time-varying conditional correlations → window-average R̄	GARCH(1,1)-t; globally pooled (a,b,g) on a 100-asset always-alive subset
2. Graphical LASSO	sparse partial-correlation graph	adaptive EBIC; n/p-tiered identity shrinkage + constrained-BIC fallback (recovers a non-empty graph at 19/20 windows)
3. Lead/follower	direct each surviving edge	two-test cascade: lagged Frisch–Waugh–Lovell partial correlation and bivariate Granger F
4. Network metrics	four topological readouts	ER clustering |Z_C|, PageRank Gini/Herfindahl, Louvain modularity + GICS sector purity, MAN triads (FFL/MR/SIM)
5. NSI	composite stress score in [0,1]	weighted sparsity + hub-concentration + mean-ρ + motif-shift

Twenty snapshots: 7 crisis peaks (Oct 1987 Black Monday, Oct 1997 Asian, Oct 1998 LTCM, Apr 2000 Dot-com, Sep 2001 9/11, Oct 2008 GFC, Mar 2020 COVID), 6 non-crisis stress windows, 2 recoveries, 5 calm baselines.

Two elements are the stated novelty: (i) the integration itself, and (ii) the two Stage-2 numerical modifications above. A formal inference layer (exact permutation tests over all C(20,7)=77,520 label assignments, cluster- and block-bootstrap confidence sets, Benjamini–Hochberg/Yekutieli + Bonferroni-20 multiplicity control) and an 836-check numerical-invariant assertion suite (805 PASS / 31 expected n/p<1 near-PSD + GARCH-boundary FAILs) back every reported number.

Headline findings (honest summary)

• PageRank Gini concentration is the strongest crisis-vs-non-crisis discriminator (Cohen's d = 1.04; the only channel clearing the uncorrected one-sided permutation test, p = 0.022) — but no channel survives multiplicity correction (smallest Benjamini–Hochberg q = 0.152).
• Q4 (mutual-dyad fraction, d = 0.68) and Q3 (cross-sector edge fraction) move in the predicted direction; Q1 is a non-randomness sanity check.
• The NSI ranks Oct-2008 GFC first but does not separate crisis from non-crisis at the regime-mean level (motivates a weight redesign, F1).
• Snapshot NSI is positively but not significantly associated with the CBOE VIX-continuity benchmark (Pearson r = 0.321; cluster-bootstrap 95% CI includes zero); a rolling counterpart is sign-inconsistent across horizons.

This is a methodology-and-characterization paper for a network venue: the contribution is the integrated framework and its rigorous, openly-disclosed evaluation (including negative results), not a working crisis predictor.

Repository layout

.
├── run_pipeline.py          # orchestrates Stages 1–5 -> results/snapshots/stage{1..5}_results.pkl
├── src/                     # pipeline
│   ├── stage1_data/         #   A-DCC GARCH + CRSP panel construction
│   ├── stage2_precision/    #   Graphical LASSO + shrinkage tiers + constrained-BIC fallback
│   ├── stage3_direction/    #   lead/follower cascade + sensitivity sweep
│   ├── stage4_network/      #   ER / PageRank / Louvain / MAN-motif metrics + dyad-preserving null
│   ├── stage5_nsi/          #   NSI composite, rolling NSI, VIX continuity, volume-weighted NSI
│   ├── _assertions/         #   836-check numerical-invariant suite
│   └── utils/               #   A-DCC core, shared helpers
├── tools/                   # data fetch (WRDS/FRED/Refinitiv), multipanel sweep, diagnostics
├── tests/                   # pytest suite (212 tests)
├── paper/                   # manuscript PDFs (preprint) + figure/table/inference generators
│   ├── paper.pdf, appendix.pdf
│   └── generate_figures.py, _inference.py, _generate_*.py, _nsi_*.py
└── data/README.md           # how to obtain CRSP (data is NOT shipped — license)

Install

Python 3.9.6 (pinned in requirements.txt / environment.yml):

pip install -r requirements.txt        # or: conda env create -f environment.yml

Core stack: numpy 2.0.2, scipy 1.13.1, pandas 2.2.3, scikit-learn 1.6.1, networkx 3.2.1, statsmodels 0.14.6, arch 7.2.0, python-louvain 0.16 (+ matplotlib 3.9.4, pyarrow 21.0.0). A Dockerfile is provided.

Data

Not included — the CRSP US Daily Stock File license prohibits redistribution. See data/README.md for the expected schema and how to rebuild it (WRDS subscription required for CRSP; the VIX continuity series is public via FRED). The derived Stage-1…5 caches, the multipanel sweep summary, and the assertion-suite report will be deposited on Zenodo (DOI on acceptance).

Reproduce

# 1. fetch data (needs WRDS/CRSP access — see data/README.md)
python tools/fetch_crsp_sp500_full.py --username <wrds_user>
python -m tools.build_crsp_volume

# 2. run the five-stage pipeline (~30 min on Apple M5)
python run_pipeline.py

# 3. 836-check numerical-invariant assertion suite
python -m src._assertions.invariants

# 4. panel-size / liquidity sweep  (N in {50..500} x {coverage, ADV, deciles})
python -m tools.run_multipanel --all --workers 8

# 5. Stage-3 (alpha, tau, lag, c_floor) sensitivity grid
python -m tools.run_stage3_sweep --csv out.csv

# 6. tests
python -m pytest tests/

# 7. regenerate the paper's figures / tables / inference macros from the caches
python paper/generate_figures.py
python -m paper._inference
python -m paper._generate_multipanel_tables

Determinism: NumPy default_rng streams are pinned (seeds 2026 / 42); the Stage-1 multi-start MLE and the sklearn coordinate-descent solver are deterministic but non-seeded (host-BLAS dependent) and reproduce within rank order across re-runs. Full per-module seed table and per-stage run-times are in the appendix "Reproducibility" section.

License & citation

• Code (src/, tools/, tests/, paper/*.py, run_pipeline.py): MIT — see LICENSE.
• Manuscript PDFs: © 2026 the author, posted as a preprint; the version of record will be under IEEE copyright on acceptance.
• Please cite via CITATION.cff.

Author

Nezih Arhan Tözenbilek — Department of Computer Engineering, Hacettepe University, Ankara, Turkey.