robust-eventstudy

Dependence-robust inference for heavy-tailed cross-asset event studies.

This package is the inference toolkit extracted from the analysis pipeline of:

Farzulla, M. "Do Cryptocurrency Markets Differentiate Infrastructure from Regulatory Shocks? A Multi-Moment Event Study with Dependence-Robust Inference." Under review at Digital Finance. Preprint DOI: 10.21203/rs.3.rs-8323026 Pipeline repository: studiofarzulla/crypto-event-study

The problem

Event studies on asset panels almost always violate the independence assumptions their tests import. A handful of assets (N ≈ 5–10) see the same events, their returns are heavily cross-correlated (crypto: ρ̄ ≈ 0.7), and their innovations are heavy-tailed (fitted Student-t ν ≈ 3–4). Under those conditions:

• pooling asset-event observations and bootstrapping them treats correlated observations as independent — p-values collapse toward zero;
• t-tests across per-asset coefficients are pseudoreplication — in the source paper the naive p = 0.0008 became p ≈ 0.32 under correct inference;
• parametric bootstraps with Gaussian innovations understate tail mass and bias bootstrap p-values downward;
• wild (sign-flip) bootstraps are near-degenerate for variance-equation coefficients, because ε² is sign-invariant — their tiny p-values are an artifact, not power.

robust-eventstudy packages the estimators and tests that survive these critiques, for both moments of an event study — and keeps the known-bad alternatives available, clearly labelled, so their optimism can be demonstrated on your data rather than asserted.

What's in it

The GARCH engine is gjr-garch-x (≥ 0.3.0), whose seeded multistart estimation reproduces the research pipeline's fits to ~4 decimal places (verified in this package's golden tests).

Install

pip install robust-eventstudy            # once published
pip install robust-eventstudy[speed]     # + numba: ~10x faster GARCH refits

From source (both packages, until 0.3.0/0.1.0 are on PyPI):

pip install "gjr-garch-x @ git+https://github.com/studiofarzulla/gjr-garch-x@master"
pip install "robust-eventstudy @ git+https://github.com/studiofarzulla/robust-eventstudy@main"

Python 3.11–3.13. (3.14 is untested upstream: parts of the scientific stack still misbehave there.)

Quickstart: reproduce a paper number

The paper's returns-leg headline — infrastructure vs regulatory events show no distinguishable CAR difference (Δ = +7.19pp, block-bootstrap p = 0.283) — from the committed data, in ~10 lines:

import pandas as pd
from robust_eventstudy import block_bootstrap_diff, event_level_cars, im_test
from robust_eventstudy.io import load_coingecko_prices, simple_returns

# price CSVs + events.csv from github.com/studiofarzulla/crypto-event-study (data/)
returns = {s: simple_returns(load_coingecko_prices(f"data/{s.lower()}.csv"))
           for s in ["BTC", "ETH", "XRP", "BNB", "LTC", "ADA"]}
events = pd.read_csv("data/events.csv")
infra = event_level_cars(returns, events[events.type == "Infrastructure"].to_dict("records"))
reg = event_level_cars(returns, events[events.type == "Regulatory"].to_dict("records"))
res = block_bootstrap_diff([e["mean_car"] for e in infra], [e["mean_car"] for e in reg])
print(f"diff = {res['diff']*100:+.2f}pp, p = {res['p_two']:.4f}")  # +7.19pp, p = 0.2828

The variance-leg inference of record (t-copula null-imposed bootstrap, p = 0.322 at B=2000) runs through GarchXBootstrap:

from robust_eventstudy import EventStudyDesign, GarchXBootstrap

design = EventStudyDesign(returns_pct, events, extra_regressors=sentiment,
                          contrast=("Infrastructure", "Regulatory"),
                          start_date="2019-01-01")
boot = GarchXBootstrap(design.build())            # Student-t copula by default
obs = boot.fit_observed(seed=12345)               # multistart per-asset fits
null = boot.run_null(b=2000, n_jobs=8, seed=12345 + 10_000)
print(obs.multiplier, null.p_one_sided)

See tests/test_golden.py for the full reproduction of the committed pipeline results, including exact seeds and data preparation.

What this package will tell you that others won't

The design goal is inference that doesn't overclaim. Expect:

• Null-imposed bootstrap p-values that include the estimator's finite-sample bias instead of assuming it away.
• Effective degrees of freedom reported alongside design-effect corrections — with N=6 assets and ρ̄=0.69, df_eff ≈ 1.1, and a t(1.1) tail is a very different animal from a normal one.
• Warning labels on the seductive shortcuts: the Gaussian-innovation bootstrap and the wild bootstrap for variance coefficients are implemented, documented as biased, and kept for comparison.
• Minimum detectable effects, because a null result is only informative relative to what the design could have detected.

Roadmap (v0.2)

• SizeStudy: simulate the whole inference ladder under a true null on your own panel (the paper's Table 8 / c10 machinery, generalized) — empirical size for each test rather than nominal claims.
• Pre-event CAR anticipation diagnostics on the returns leg.
• Regime/break-control dummies in EventStudyDesign (the paper's c8 battery).

Citation

If you use this package, please cite both the software (see CITATION.cff) and the paper (DOI 10.21203/rs.3.rs-8323026).

License

MIT. Test fixtures under tests/data/ are copied from the crypto-event-study repository (MIT) at commit 1baf97c; price histories derive from CoinGecko and sentiment aggregates from GDELT.