Limitations and Disclosures

This page lists the material limitations of the strategy and the assumptions behind the backtest. We surface them so a reader can size the claim.

Headline-result limitations

1. Selection bias on the 4-instrument basket

The headline basket (AAPL + MSFT + WMT + GLD) was selected from a tested universe of 12 instruments based on robust positive excess Sharpe under the multiple-testing acceptance gates. Selection robustness is validated by the Deflated Sharpe Ratio (DSR) with implied-independent-trials correction (PSR = 1.0000 against a noise-floor benchmark for 9 implied independent trials), and the Probability of Backtest Overfitting via combinatorially symmetric cross-validation (PBO = 0.0402, well below the 0.30 acceptance threshold). The selection passes both gates.

2. Risk-free rate assumption

All Sharpe ratios reported as “excess” use a 2.33% annualized risk-free rate, which is the realized average of the CBOE 13-week T-bill yield (IRX) over the OOS window 2018-2024. The OOS period included near-zero rates in 2018-2021 and 4-5% rates in 2022-2024, so the 2.33% within-sample average is the appropriate anchor.

The 2.33% rate is the realized within-sample average. Sharpe is monotonic in (return − rf): a higher rf assumption tightens the excess return, a lower one widens it. The strategy’s Sharpe advantage over the SPX baseline (+0.085) is preserved across rf assumptions because both numbers shift together.

3. OOS sample size

OOS spans 2018-01-01 to 2024-12-31, about 7 years and 1,760 trading days. Effective independent observations at the 30-day forward horizon are roughly 84 non-overlapping windows. The volatility-risk-premium literature (Bollerslev, Tauchen, Zhou 2009) used 17 years of S&P data; ours is 7. Statistical power is bounded by the shorter sample.

Strategy-level limitations

4. Tail risk is real and bounded but not zero

Put-credit-spread max-loss-per-spread is bounded by the wing width: 5 strike points equals $500 per spread per contract before commissions. With realistic per-trade sizing (1% of equity per spread), absolute drawdown in a single day is bounded. But:

  • Multi-day cascade events (e.g., Volmageddon February 2018) can produce simultaneous emergency exits on multiple positions before the halt framework engages.
  • The 4-instrument basket’s correlation rises during stress events. Typical pairwise correlation jumps from 0.26 in normal regimes to 0.6+ in a crisis.

5. Strategy capacity is limited

Open-interest data on our chosen 16-delta strikes shows median per-strike open interest in the hundreds-to-thousands for AAPL, MSFT, and WMT, and in the dozens-to-hundreds for GLD. Scaling beyond ~$5–20M of capital (depending on instrument) likely encounters market-impact friction beyond what our backtest’s bid-ask slippage model captures. This is a retail / small-fund strategy, not an institutional one.

Methodological limitations

6. Iron condor variant tested, put-only shipped

We benchmarked iron condor (put + call wings on the same expiry) and put-only architectures on the same engine and universe. Put-only outperformed iron condor in every paired comparison in OOS 2018-2024 because the post-2020 trending-equity regime systematically destroyed short-call positions. The headline ships put-only as the evidence-based architecture choice. Iron condor results are reported transparently in Variants.

7. Daily-frequency realized volatility, not intraday

The volatility-risk-premium literature finds that 5-minute intraday squared returns produce stronger predictability for VRP-based forecasting than daily close-to-close realized volatility. We use daily realized volatility in the engine’s vol-multiplier computation. Without intraday data, the realized-vol input is coarser than the published-best specification.

8. R² inflation in overlapping regressions

In overlapping multi-period predictive regressions, R² inflates roughly proportionally with horizon × overlap, even in the absence of true predictability. Our 30-day forward window with daily observations creates this pattern.

We mitigate by reporting Hodrick (1992) standard errors as the primary significance measure for any regression we report, rather than naive Newey-West. R² is reported descriptively only.

Backtest-environment limitations

9. Synthetic-pricing fallback when option-chain Greeks are missing

The OptionMetrics chain data has a 7-15% NaN rate on delta and implied volatility per ticker, concentrated at deep-in-the-money or deep-out-of-the-money strikes where Greeks aren’t computed. The engine falls back to Black-Scholes-derived deltas using VIX/100 as an at-the-money implied-volatility proxy when needed. The magnitude is small at our 16-delta strikes (which have populated Greeks in the chain data) but worth flagging.

10. Friction-model calibration is best-effort

Slippage, commission, and gap-aware-fill parameters are conservative best-effort estimates, not directly calibrated to specific broker fee schedules. Real-world results will differ in these dimensions. We err conservative (more slippage in higher-VIX buckets, gap-fills at worst execution).

11. American early-exercise risk on dividend-paying single-stocks

American puts on dividend-paying stocks have positive probability of premature exercise. Our headline basket has three of four names paying dividends (AAPL, MSFT, WMT; GLD is a commodity ETF and does not pay dividends). The engine prices these with European-style Black-Scholes when chain quotes are missing. Magnitude is small at our 16-delta out-of-the-money entry strikes; it grows when the underlying drops materially and the put becomes in-the-money near an ex-dividend date. We do not model the assignment cash-flow precisely; reported P&L represents out-of-the-money-expiration economics.

12. Black-Scholes pricing biases are the source of VRP itself

Black-Scholes (1973) systematically overvalues options on high-variance underlyings and undervalues low-variance, per the empirical test in the original 1973 paper. This bias is the structural source of the volatility risk premium the strategy harvests. We do not “out-predict” Black-Scholes; we systematically take the side of the trade where empirical VRP is positive in the historical record.

Exogenous-factor data limitations

13. Macro factors are point-in-time but not as-of-decision-time stamped

The macro factors used by the engine (VIX, VIX3M, VVIX, SKEW, IRX, TNX, HYG, LQD, SPY, SPX bars) are sourced from an Interactive Brokers TWS feed snapshot frozen at commit time. These are point-in-time by the snapshot date but lack the explicit as-of-decision-time stamping that institutional data providers (Bloomberg, Refinitiv) provide. We use point-in-time pd.Series.asof(today) lookups to enforce no-look-ahead, but small backfill artifacts in the source feed are possible.

14. Per-ticker underlying spot from a separate query

The 4-instrument basket’s underlying daily spots (AAPL, MSFT, WMT, GLD) are sourced from the OptionMetrics IvyDB Securities file (queried via WRDS). This is the same vendor as the option chain data, which ensures security-ID alignment between chain quotes and underlying spot for the same security on the same date. Daily close prices match published exchange records on every spot-checked reference date.

Live monitoring limitations

15. Hoeffding bound is worst-case

The Hoeffding inequality bound holds for any bounded random variable regardless of distribution, but it is loose. Tighter bounds exist (e.g., Bernstein, normal-approximation) for distributions where their assumptions hold. We use Hoeffding for distribution-free guarantees: this means the 50% / 25% / 10% thresholds are conservative. The actual probability of a regime change is at most these values, not exactly these values.

16. μ baseline does not adapt over time

The μ = 0.730 baseline is computed once from the OOS backtest and held fixed in live deployment. Adapting μ live to the trailing 12-month observed win rate would create a moving-goalpost reference and reduce the monitor’s ability to detect drift. The trade-off: occasional false-negatives in genuine regime changes that look “normal” relative to the fixed baseline.

Multi-testing limitations beyond the headline

17. Mode comparison creates a second multiple-testing layer

The variants page includes mode comparisons (halts engaged versus disengaged). Comparing the better of the two modes against the baseline introduces a second selection layer beyond the basket-selection layer corrected via DSR + PBO.

The headline only reports the halts-engaged mode result. The halts-disengaged results are reported on the Variants page for transparency, not as candidates.

Future work: treasury management on undeployed capital

The current backtest accrues idle cash at the realized 13-week T-bill yield (IRX averaged 2.33% over OOS). This is the conservative baseline. A natural next step is active treasury management on the undeployed $200,000 of basket capital that backs the options positions.

Replacing the passive T-bill accrual with an ultra-short Treasury ETF (SGOV, BIL) or a short-duration corporate bond fund (VCSH, BSV) targets an additional 100-200 bps of carry with minimal duration or credit risk. The trade-off is a small increase in drawdown sensitivity during credit-spread blowouts, but the headline -0.12% maximum drawdown leaves substantial room before this becomes a constraint.

The deliberate exclusion is broad-equity exposure (SPY, QQQ) for the undeployed capital. Adding equity beta to the cash sleeve introduces correlation risk: in the stress events the strategy is designed to absorb (Volmageddon, COVID, 2022 bear, 2023 banking), the put-credit-spread portfolio is under maximum pressure exactly when an equity cash sleeve would also be drawing down. This double-correlation would destroy the maximum-drawdown profile that makes the strategy useful as a portfolio defensive sleeve.

Recommended treasury overlay for production deployment: 80% allocation to SGOV (0-3M Treasuries, near-zero duration, currently yielding around 5%) and 20% to short-investment-grade-corporate (VCSH, around 1-year duration, currently yielding around 5%). Expected blended carry roughly 100-150 bps above passive IRX accrual, raising the strategy’s annualized return from 2.41% to approximately 3.5-4%, while keeping the maximum-drawdown floor near -0.12%. This is a low-risk capital-efficiency improvement, not a return-chasing one.