Options-Implied Crypto Signals

Testing whether spikes in Deribit's BTC and ETH implied-volatility index (DVOL) front-run short-horizon moves in spot crypto returns. A quantitative research project on the lead-lag relationship between options and spot markets.

---

Headline Result

Strategy	OOS trades	OOS Sharpe	Walk-forward mean	WF range
ETH_D3_24h  ·  IV-risk-premium shock, 24h hold	182	3.83	2.76	−7.24 to +14.72

ETH IV-risk-premium shocks show a promising short-horizon directional edge after correcting timing, signal, and metric issues, but the effect is regime-dependent and requires further conditioning before it can be treated as deployable alpha.

---

What This Project Asks

Do sudden DVOL spikes — particularly unusual gaps between implied and realised volatility — carry directional information about spot prices in the next 1–24 hours?

If so, the options market is acting as a forward-looking signal for the spot market, consistent with informed traders expressing views through options before they show up in spot.

Research Pipeline

flowchart LR
    A[Deribit DVOL<br/>hourly · 4.5y] --> C[Signals<br/>D1 – D4]
    B[Binance spot<br/>15 pairs · hourly] --> C
    C --> D[Lead-lag<br/>CCF + Granger]
    C --> E[Event study<br/>CAR]
    C --> F[Backtest<br/>walk-forward]
    C --> G[Cross-asset<br/>15 altcoins]
    D --> H((Headline<br/>result))
    E --> H
    F --> H
    G --> H

Loading

#	Notebook	What it does
01	01_data_exploration.ipynb	Load DVOL + Binance spot, compute realised vol
02	02_signals.ipynb	Generate D1–D4 signals → data/signals/
03	03_lead_lag.ipynb	Cross-correlation, Granger causality, 2 placebos
04	04_event_study.ipynb	Cumulative abnormal returns around signal events
05	05_backtest.ipynb	40 evaluated variants, walk-forward validation
06	06_cross_asset.ipynb	BTC DVOL → 15 altcoins propagation diagnostic

Signal Definitions

Signal	Definition	Notes
D1	log-DVOL z-score > 2σ AND |Δlog DVOL| > 2%	Relative spike — failed Granger
D2	|Δlog DVOL| ≥ 5%	Absolute shock — sparse (~20–25 events/yr)
D3	IV-premium (DVOL − realised vol) z-score > 2σ	Headline signal — denser, stronger
D4	D1 conditions AND range z-score ≥ 1.5σ	Stricter D1 subset — raise threshold for selectivity

24-hour cooldown applied to all signals to enforce event independence.

Key Charts

Event-study cumulative abnormal returns around ETH_D2 signals

Cross-asset propagation heatmap — BTC DVOL → 15 altcoins, by lag

Audit and Fix Cycle

A rigorous code review pass at the end of the project surfaced four genuine bugs that were fixed before publishing the headline result:

#	Bug	Fix
1	Backtest off-by-one — holding_period=H summed H+1 return bars	exit_pos = entry_pos + holding_period - 1 in src/backtest.py
2	Event-study alpha parameter ignored, bands hardcoded to 95%	z = norm.ppf(1 - alpha/2) in src/event_study.py
3	Sharpe annualised on event-trade returns with sqrt(252) — overstated by ~3–4×	Calendar-time hourly equity curve × sqrt(8760) in src/metrics.py
4	D4 identical to D1 in cached signals — range filter too loose	Z-score-based range filter in src/signals.py

These fixes flipped the headline finding from ETH_D2 (sparse, walk-forward catastrophic) to ETH_D3 (denser, walk-forward positive but volatile). The audit-and-fix cycle is documented openly as evidence of research process — the corrected story is both stronger and more honest than the first pass. Each bug has a dedicated regression test in tests/.

Honest Limitations

• Walk-forward dispersion is large. Per-window OOS Sharpes for the headline strategy range from −7 to +15 across 13 windows. The signal works on average but is regime-dependent, not consistently profitable.
• Cross-asset correlations are raw, not beta-adjusted. Altcoins follow BTC, so a portion of "propagation" may be BTC DVOL → BTC spot → alts. Next step: regress alt returns on BTC contemporaneous and lagged returns and re-test on residuals.
• D4 with current calibration filters only ~1% of D1 events. Raise signals.d4.range_multiplier in config.yaml to 2.0–2.5 for a meaningfully stricter subset before drawing inference from D4.
• No live execution simulation. Costs include commission + slippage but not exchange spread, liquidity at scale, funding, or capacity constraints.

Next Research Steps

The natural follow-up is regime conditioning, not more raw signal mining — identify when D3 works (volatility-trending markets? FOMC weeks? bear regimes?) and when it inverts. This is where the project moves from "interesting research signal" to "candidate deployable strategy."

Quick Start

# install
pip install -r requirements.txt

# fetch ~30 MB of parquet data (free, public APIs — no account needed)
python download_data.py

# run all six notebooks in order
for nb in notebooks/0*.ipynb; do
  jupyter nbconvert --to notebook --execute --inplace "$nb"
done

# run the regression test suite
pytest tests/ -v

The tests pin all four audit-found bugs and run in under a second.

Repository Layout

.
├── config.yaml                # All parameters in one place
├── pyproject.toml             # Package metadata + pytest config
├── download_data.py           # Fetch DVOL + Binance spot
├── src/
│   ├── deribit_fetch.py       # DVOL via Deribit API
│   ├── binance_fetch.py       # Binance public archive
│   ├── signals.py             # D1–D4 definitions
│   ├── lead_lag.py            # CCF, Granger, placebos
│   ├── event_study.py         # CAR analysis
│   ├── backtest.py            # Event-driven backtest, walk-forward
│   ├── metrics.py             # Hourly-equity Sharpe, Deflated Sharpe
│   └── cross_asset.py         # Propagation diagnostics
├── notebooks/                 # 01–06, run in order
├── tests/                     # 9 regression tests for audit-found bugs
├── data/                      # Mostly gitignored — large parquets
│   ├── deribit/  crypto/  signals/   (gitignored, fetched by download_data.py)
│   ├── plots/                 # Committed figures
│   └── results/               # Committed headline CSVs
├── archive/                   # Legacy code from the abandoned prediction-market design
├── .github/workflows/         # CI: pytest on every push
├── LICENSE                    # MIT
└── README.md

Data

Source	Symbols	Resolution	Range	Access
Deribit DVOL	BTC, ETH	1 h	2022-01 → 2026-06	Public API, no account
Binance spot	15 USDT pairs	1 h	2022-01 → 2026-06	Public archive, no account

Total downloaded: ~30 MB across 17 parquet files. Cached locally, regenerable via python download_data.py.

References

• Deribit DVOL — Deribit's implied-volatility index, analogous to the CBOE VIX
• Bailey & Lopez de Prado (2014) — The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality
• Granger (1969) — Investigating Causal Relations by Econometric Models and Cross-Spectral Methods
• Kyle (1985) — Continuous Auctions and Insider Trading
• Bollerslev, Tauchen & Zhou (2009) — Expected Stock Returns and Variance Risk Premia

---

Author: Tanvir Chowdhury · MSc Mathematical Trading and Finance, Bayes Business School