Data and Literature

Data sources

The strategy reads from three external data providers, all snapshotted to local parquet files at commit time. No live network connections occur during a backtest.

1. OptionMetrics IvyDB US (option chains)

Daily option chains 2012-2025 with bid, ask, impl_volatility, delta, gamma, vega, theta, volume, open_interest for the 12 instruments tested. Acquired from WRDS as a single CSV (strat2.2.csv.gz, 2.4 GB compressed, 68.4M rows) and split via scripts/convert_optionmetrics_to_parquet.py into per-ticker parquets at data/processed/options_by_ticker/{TICKER}.parquet. The engine prices entries and exits at real chain quotes (not Black-Scholes fallback) wherever available.

2. OptionMetrics IvyDB Securities (per-ticker underlying OHLC)

Daily underlying open/high/low/close/volume for the 4 headline-basket tickers (AAPL, MSFT, WMT, GLD) and 6 ablation tickers (TLT, GOOGL, JNJ, KO, PG, JPM, PEP). Acquired from WRDS as wheelstrat_data.gz (760 KB) and split via scripts/convert_wheelstrat_data.py into per-ticker parquets at data/raw/{TICKER}.parquet. Spans 2012-01-03 to 2025-08-29. The IvyDB Securities file uses the same secid system as the option chains, so the join between underlying spot and chain quotes for the same security on the same date is exact.

Daily closes match published exchange records on every spot-checked reference date (e.g. AAPL 2018-04-16 closes at $175.82, matching the public record).

3. Interactive Brokers TWS feed snapshot (macro factors)

Frozen snapshots of CBOE volatility indices (VIX, VIX3M, VVIX, SKEW), Treasury yields from FRED via TWS (IRX, TNX), credit-stress proxies (HYG, LQD), broad equity tape (SPY, SPX bars). Stored at data/raw/{TICKER}.parquet. Read by pd.Series.asof(today) at trade-decision time so no look-ahead bias is possible. Used for the halt framework’s Layer 1 hard tail-event triggers, the ML stress head’s feature matrix, the friction model’s VIX-conditional slippage, and the risk-free baseline (IRX averaged over OOS = 2.33%).

Literature foundation

The strategy and writeup are grounded in the following published references. None of these are course-internal materials, all are peer-reviewed.

Variance risk premium

Carr, P., and Wu, L. (2003 working paper, 2009 Review of Financial Studies). Variance Risk Premia. Documents the structural gap between option-implied and realized variance across U.S. equity indices. Provides the empirical t-statistics on the SPX cluster that anchor the project’s universe selection (SPX strongly supported, NDX weaker, RUT extrapolated, TLT and GLD are novel cross-asset extensions).
Bollerslev, Tauchen, Zhou (2009). Expected Stock Returns and Variance Risk Premia. Review of Financial Studies. Establishes the 63-trading-day quarterly horizon as empirically optimal for VRP-based forecasting and motivates the Hodrick-style standard errors used for any overlapping regression in this work.
Black, F., and Scholes, M. (1973). The Pricing of Options and Corporate Liabilities. Journal of Political Economy. The original closed-form pricing of European options. The empirical biases the paper documents in OTC option pricing (BS overvalues high-variance underlyings, undervalues low-variance) are the structural source of the volatility risk premium the strategy harvests.
Merton, R. C. (1973). Theory of Rational Option Pricing. Bell Journal of Economics and Management Science. Theorem 13 on early exercise of American puts on dividend-paying stocks frames the pricing approximation disclosed for the wheel-name positions.

Multiple-testing correction

Bailey, D. H., and López de Prado, M. (2014). The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality. Journal of Portfolio Management. Equation 9 of this paper provides the implied-independent-trials adjustment N̂ = ρ̄ + (1 − ρ̄) · M used to deflate the headline Sharpe against the 12-trial selection-bias noise floor.
Bailey, D. H., Borwein, J. M., López de Prado, M., and Zhu, Q. J. (2015). The Probability of Backtest Overfitting. Journal of Computational Finance. Defines the Probability of Backtest Overfitting via combinatorially symmetric cross-validation. Uses S = 16 partitions, C(16, 8) = 12,870 logit combinations to compute PBO = 0.0402 on the project’s per-instrument return matrix.

Live monitoring

Hoeffding, W. (1963). Probability Inequalities for Sums of Bounded Random Variables. Journal of the American Statistical Association. The bound P[X̄ − μ ≥ t | H₀] ≤ exp(−2 t² N) for any bounded random variable underpins the live-monitoring framework’s threshold semantics.
Egger, J., and Vestal, R. (2025). Trader-Application Form of Hoeffding’s Inequality. FinTech 533, Duke MEng FinTech. Source of the trader-form reformulation (Eq. 1.4) and the threshold semantics this project’s live monitor uses.

Calibration

Niculescu-Mizil, A., and Caruana, R. (2005). Predicting Good Probabilities with Supervised Learning. Proceedings of the 22nd International Conference on Machine Learning. Provides the sample-size threshold (n_pos minority class ≈ 200) that the auto-select between Platt scaling and isotonic regression uses in src/models/calibration.py.

Time-series statistics

Hodrick, R. J. (1992). Dividend Yields and Expected Stock Returns: Alternative Procedures for Inference and Measurement. Review of Financial Studies. Defines the variance estimator for overlapping multi-period predictive regressions that the project’s regression diagnostics use, in preference to naive Newey-West standard errors.

Data limitations and how they shape the writeup

Limitation	Where the writeup acknowledges it
OptionMetrics has 7-15% NaN on Greeks at deep-OTM/ITM strikes	Limitations
Per-instrument bars from a separate WRDS query (not from the chain file itself)	Limitations
TWS-sourced macro factors are point-in-time but not as-of-decision-time stamped	Limitations
Daily-frequency RV in the engine, not 5-minute intraday	Limitations
OOS sample only 7 years long	Limitations
RUT and NDX dropped because their underlying OHLC was not available locally	Limitations

--- title: "Data and Literature" --- ## Data sources The strategy reads from three external data providers, all snapshotted to local parquet files at commit time. No live network connections occur during a backtest. ### 1. OptionMetrics IvyDB US (option chains) Daily option chains 2012-2025 with `bid`, `ask`, `impl_volatility`, `delta`, `gamma`, `vega`, `theta`, `volume`, `open_interest` for the 12 instruments tested. Acquired from WRDS as a single CSV (`strat2.2.csv.gz`, 2.4 GB compressed, 68.4M rows) and split via `scripts/convert_optionmetrics_to_parquet.py` into per-ticker parquets at `data/processed/options_by_ticker/{TICKER}.parquet`. The engine prices entries and exits at real chain quotes (not Black-Scholes fallback) wherever available. ### 2. OptionMetrics IvyDB Securities (per-ticker underlying OHLC) Daily underlying open/high/low/close/volume for the 4 headline-basket tickers (AAPL, MSFT, WMT, GLD) and 6 ablation tickers (TLT, GOOGL, JNJ, KO, PG, JPM, PEP). Acquired from WRDS as `wheelstrat_data.gz` (760 KB) and split via `scripts/convert_wheelstrat_data.py` into per-ticker parquets at `data/raw/{TICKER}.parquet`. Spans 2012-01-03 to 2025-08-29. The IvyDB Securities file uses the same `secid` system as the option chains, so the join between underlying spot and chain quotes for the same security on the same date is exact. Daily closes match published exchange records on every spot-checked reference date (e.g. AAPL 2018-04-16 closes at $175.82, matching the public record). ### 3. Interactive Brokers TWS feed snapshot (macro factors) Frozen snapshots of CBOE volatility indices (VIX, VIX3M, VVIX, SKEW), Treasury yields from FRED via TWS (IRX, TNX), credit-stress proxies (HYG, LQD), broad equity tape (SPY, SPX bars). Stored at `data/raw/{TICKER}.parquet`. Read by `pd.Series.asof(today)` at trade-decision time so no look-ahead bias is possible. Used for the halt framework's Layer 1 hard tail-event triggers, the ML stress head's feature matrix, the friction model's VIX-conditional slippage, and the risk-free baseline (IRX averaged over OOS = 2.33%). ## Literature foundation The strategy and writeup are grounded in the following published references. None of these are course-internal materials, all are peer-reviewed. ### Variance risk premium - **Carr, P., and Wu, L. (2003 working paper, 2009 Review of Financial Studies).** *Variance Risk Premia.* Documents the structural gap between option-implied and realized variance across U.S. equity indices. Provides the empirical t-statistics on the SPX cluster that anchor the project's universe selection (SPX strongly supported, NDX weaker, RUT extrapolated, TLT and GLD are novel cross-asset extensions). - **Bollerslev, Tauchen, Zhou (2009).** *Expected Stock Returns and Variance Risk Premia.* Review of Financial Studies. Establishes the 63-trading-day quarterly horizon as empirically optimal for VRP-based forecasting and motivates the Hodrick-style standard errors used for any overlapping regression in this work. - **Black, F., and Scholes, M. (1973).** *The Pricing of Options and Corporate Liabilities.* Journal of Political Economy. The original closed-form pricing of European options. The empirical biases the paper documents in OTC option pricing (BS overvalues high-variance underlyings, undervalues low-variance) are the structural source of the volatility risk premium the strategy harvests. - **Merton, R. C. (1973).** *Theory of Rational Option Pricing.* Bell Journal of Economics and Management Science. Theorem 13 on early exercise of American puts on dividend-paying stocks frames the pricing approximation disclosed for the wheel-name positions. ### Multiple-testing correction - **Bailey, D. H., and López de Prado, M. (2014).** *The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality.* Journal of Portfolio Management. Equation 9 of this paper provides the implied-independent-trials adjustment N̂ = ρ̄ + (1 − ρ̄) · M used to deflate the headline Sharpe against the 12-trial selection-bias noise floor. - **Bailey, D. H., Borwein, J. M., López de Prado, M., and Zhu, Q. J. (2015).** *The Probability of Backtest Overfitting.* Journal of Computational Finance. Defines the Probability of Backtest Overfitting via combinatorially symmetric cross-validation. Uses S = 16 partitions, C(16, 8) = 12,870 logit combinations to compute PBO = 0.0402 on the project's per-instrument return matrix. ### Live monitoring - **Hoeffding, W. (1963).** *Probability Inequalities for Sums of Bounded Random Variables.* Journal of the American Statistical Association. The bound P[X̄ − μ ≥ t | H₀] ≤ exp(−2 t² N) for any bounded random variable underpins the live-monitoring framework's threshold semantics. - **Egger, J., and Vestal, R. (2025).** *Trader-Application Form of Hoeffding's Inequality.* FinTech 533, Duke MEng FinTech. Source of the trader-form reformulation (Eq. 1.4) and the threshold semantics this project's live monitor uses. ### Calibration - **Niculescu-Mizil, A., and Caruana, R. (2005).** *Predicting Good Probabilities with Supervised Learning.* Proceedings of the 22nd International Conference on Machine Learning. Provides the sample-size threshold (n_pos minority class ≈ 200) that the auto-select between Platt scaling and isotonic regression uses in `src/models/calibration.py`. ### Time-series statistics - **Hodrick, R. J. (1992).** *Dividend Yields and Expected Stock Returns: Alternative Procedures for Inference and Measurement.* Review of Financial Studies. Defines the variance estimator for overlapping multi-period predictive regressions that the project's regression diagnostics use, in preference to naive Newey-West standard errors. ## Data limitations and how they shape the writeup | Limitation | Where the writeup acknowledges it | |---|---| | OptionMetrics has 7-15% NaN on Greeks at deep-OTM/ITM strikes | [Limitations](limitations.qmd) | | Per-instrument bars from a separate WRDS query (not from the chain file itself) | [Limitations](limitations.qmd) | | TWS-sourced macro factors are point-in-time but not as-of-decision-time stamped | [Limitations](limitations.qmd) | | Daily-frequency RV in the engine, not 5-minute intraday | [Limitations](limitations.qmd) | | OOS sample only 7 years long | [Limitations](limitations.qmd) | | RUT and NDX dropped because their underlying OHLC was not available locally | [Limitations](limitations.qmd) |