Do Cryptocurrency Markets Differentiate Infrastructure from Regulatory Shocks?

A Multi-Moment Event Study with Dependence-Robust Inference

Replication materials for the merged multi-moment paper (supersedes the earlier single-moment "Infrastructure vs Regulatory Shocks" and the companion "Same Returns, Different Risks").

Finding. Cryptocurrency markets do not statistically differentiate infrastructure from regulatory shocks at either moment — returns (block-bootstrap p ≈ 0.28) or conditional variance (Student-t-copula bootstrap p = 0.322) — once inference accounts for cross-event/cross-asset dependence and heavy tails. The apparent volatility asymmetry reported in the earlier version was an inference artefact. The lead contribution is a methodological cautionary tale: an inference ladder and a Monte-Carlo size study showing how naive event-study inference manufactures significance that correct inference dissolves.

Repository structure

• paper/ — merged manuscript (main.tex, main.pdf) + response/cover letters
• code/ — verified analysis pipeline (c1–c14, tarch_x_manual.py, tarch_x_fast.py)
• code/src/ — CAR engine for the returns leg (ConstantMeanModel in event_study.py + config.py), inherited from the retired standalone returns paper; imported by c11
• data/ — shared sample: 6 assets (BTC, ETH, XRP, BNB, LTC, ADA), 50 events (events.csv + events_reclassified.json), GDELT sentiment
• results/ — committed outputs (CSVs + per-analysis FINDING docs)
• _archive/ — superseded prior-version materials; do not cite
• springer-submission/ — frozen original single-moment submission, retained as record (superseded)

Script → paper map

script	produces
c1_build_candidate_pool	candidate-event pool + drop-out census
c2_relaxed_threshold / c2b_two_asset_point	scope-condition multiverse (curated vs mechanical)
c3_bai_perron	structural breaks (descriptive)
c5_pseudoreplication_test	inference-ladder rungs 2–3 (event-level / cluster)
c6_garchx_clustered	design-effect / correlation-weighted (rung 4)
c7_ccc_garchx_bootstrap	Gaussian-copula bootstrap (rung 5)
c9_tcopula_bootstrap	Student-t-copula bootstrap — inference of record (rung 6)
c8a / c8h_break_controls	structural-break regime controls
c8b/c8c/c8d/c8e	anticipation / winsorisation / constant-mean / persistence
c8f_weekly_granger_fdr	weekly sentiment-leads-volatility (+ BH-FDR)
c10_size_study	Monte-Carlo size-distortion study
c11_returns_block_bootstrap	first-moment returns null (rung 7)

Reproduce

Tested on Python 3.11–3.13. (Python 3.14 is not yet supported by the pinned pandas==2.3.1, whose datetime C-extension segfaults there — use 3.11–3.13.)

python3.13 -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
python code/descriptive_stats.py       # Table 1 (fast sanity check)
python code/c9_tcopula_bootstrap.py    # inference of record (variance)
python code/c11_returns_block_bootstrap.py   # first moment (returns)
python code/c10_size_study.py          # Monte-Carlo size study

Scripts read the committed data/*.csv and write to results/; all seeds are fixed, so a clean-clone run regenerates the committed CSVs (e.g. c11 rewrites results/c-gate-returns-unified-results.csv — git diff should come back clean).

Returns leg (c11) in full: the gate runs A/B — the numbers used in the paper (block-bootstrap p = 0.283 on the 6-asset basis) — need only the committed CSVs and run from a clean clone as-is. The script's optional smoke test additionally replays the retired returns paper's original 4-asset sample, which requires a Binance daily-kline cache that is not committed (derived data, *.parquet is gitignored). Rebuild it first with

python code/fetch_binance_cache.py

which fetches from the public Binance klines API (no key needed), writes data/cache/*.parquet, and verifies the rebuilt returns series against SHA-256 fingerprints of the original run. Without the cache, c11 skips the smoke test with a notice and still produces the gate results.

Note on the prior version

This repository previously hosted the single-moment "5.7×, p = 0.0008" result. That estimate did not survive dependence-robust inference; the point estimate is unchanged but it is no longer statistically distinguishable from zero. This paper reports the corrected dual null and the inference lesson openly, as self-correction. Superseded materials are in _archive/ and springer-submission/.

License

• Code (code/, including code/src/): MIT — see LICENSE.
• Manuscript (paper/) and the FINDING docs in results/: CC BY 4.0.
• Data: events.csv / events_reclassified.json are author-curated (CC BY 4.0); price CSVs derive from CoinGecko public data and gdelt.csv from the GDELT Project — both redistributed here as small research extracts with attribution. The Binance-derived smoke-test cache is not redistributed; rebuild it locally with code/fetch_binance_cache.py.

Citation

See CITATION.cff.