SYNCRISK Quantlets

Replication code for "SYNCRISK: A Label-Free Audit of Effective Diversity in AI Risk-Scoring Panels" (Daniel Traian Pele).

SYNCRISK measures residual synchronization (Excess-AFSI) among LLM-based financial-risk scores — the cross-model agreement that survives after shared reaction to common observable information is removed out of sample. Each Quantlet below reproduces one reported table or figure from the released, B-stabilized results files. Every Quantlet ships a .py script, a .ipynb notebook, a .csv of its output, and (where the deliverable is a figure) a .png.

Quantlet	Reproduces	Output
SYNCRISKdecay	Decay chain (Table 3, Figure 2): Raw → C1x → C2x → C3x by arm	csv, png
SYNCRISKladder	Cleaning-ladder specification and orthogonality summary	csv
SYNCRISKfamily	Family-block decomposition (within-commercial / within-open / cross-family)	csv
SYNCRISKstress	Stress diagnostic and stress-minus-normal premium (upper bounds)	csv
SYNCRISKcore	High-coverage 13-asset core robustness check	csv
SYNCRISKtier	Exploratory frontier-tier extension (Appendix G)	csv
SYNCRISKeq7	Descriptive co-movement projection — the market-link null	csv
SYNCRISKoverlap	Effective diversity as a curve: K_eff(J) from the matched-overlap sweep, with the empirical panel (J=0.57, K_eff=2.1) marked on it	csv, png
SYNCRISKtimeseries	Per-week Excess-AFSI time series	csv, png
SYNCRISKsynthetic	Full pipeline end-to-end on a synthetic panel — no licensed data	csv
SYNCRISKpanel	Realized panel construction and coverage (30 US equities × 244 weeks)	csv, png

Layout (self-contained)

Every Quantlet folder is independently self-contained: it ships its own input data, code, and results side by side, and runs on its own even if copied out of this bundle. Each quantlets/SYNCRISK*/ directory holds:

• the .py script and matching .ipynb notebook (code),
• the input data file(s) it reads (data) — bundled locally, so no ../../data reach-out,
• the .csv/.png it produces (results),
• the two pipeline Quantlets (SYNCRISKtimeseries, SYNCRISKsynthetic) also carry a local copy of the syncrisk/ library they import.

Quantlets/
├── README.md, Metainfo.txt
├── src/syncrisk/          canonical index/cleaning library (dev source of truth)
├── data/published/        canonical released, B-stabilized result files (provenance)
├── data/synthetic/        canonical synthetic-panel generator (no licensed inputs)
└── quantlets/SYNCRISK*/   the eleven Quantlets, each self-contained:
                           script + notebook + its input data + csv/png output

The top-level src/ and data/ remain the canonical source of truth; each Quantlet folder carries its own copy of exactly the files it needs, so a reviewer can download a single Quantlet folder and run it in place.

To run one: cd quantlets/SYNCRISKdecay && python SYNCRISKdecay.py (Python 3, pandas/numpy/scikit-learn/lightgbm). Each script reads its input data from its own folder and writes its .csv/.png output beside itself.

Reproducibility

• All Quantlets render from the released results files bundled under data/published/ (e.g. upg_Bstab_results.json); reproduction requires no model calls and no licensed data.
• SYNCRISKsynthetic runs the entire standardization → cleaning → Excess-AFSI pipeline on synthetic scores, so it is fully self-contained and is the recommended entry point for reviewers.
• The underlying price/news inputs are from a single commercial feed (EOD Historical Data) and are licensed, so they are not redistributed; the package ships the elicited and derived files needed to reproduce every reported number.
• Fixed seeds throughout (random_state = 0).

Status

The repository is withheld during double-blind review. On acceptance the collection is deposited on the QuantLet platform with persistent identifiers, with accompanying slides on Quantinar.