Volatility Course Project: Build a Conditioned Premium Screen

The final output of this volatility project is a daily table, though the grade comes from the research underneath it. You will combine implied versus realized volatility, IV rank, and vol-of-vol, then show exactly what each condition contributes and where the screen fails.

Our volatility trading learning roadmap sequences the wider subject. This high-risk page specifies one course project and its deliverables, with no claim that a high reading automatically creates a profitable trade.

State three separate questions

Begin with the hypothesis that a combination of cross-sectional premium, own-history stretch, and volatility stability may identify a useful research cohort. Keep the language conditional because the project exists to test the claim rather than assume it.

Each measure answers a different question. Implied versus realized asks how much volatility is priced relative to recent delivery. IV rank asks where current implied volatility sits inside the name's trailing range. Vol-of-vol describes how unstable the volatility state has been.

• Premium: calculate both IV minus HV and IV divided by HV.
• Stretch: retain IV rank and IV percentile as distinct fields.
• Stability: use vol-of-vol as a condition or sizing input.
• Outcome: define forward reversion or realized movement before screening.
• Failure: reject thresholds that collapse outside the research sample.

Build the research panel

Start with a point-in-time optionable universe and timestamp every input. Join implied volatility, historical volatility, IV rank, IV percentile, vol-of-vol, liquidity fields, sector, and known catalyst dates. Preserve rows excluded by data quality or liquidity rules.

Avoid dividing by tiny realized volatility without a floor or diagnostic because the ratio can explode mechanically. Inspect stale observations and corporate events. A stock ahead of earnings can look richly priced for a perfectly legible reason that the model should record.

Create a cohort ladder instead of jumping to the final screen. Measure the broad universe, then premium only, premium plus stretch, and the full conditioned cohort. Report how many observations survive and which names or periods dominate each step.

Keep the transformations readable. The panel should include raw IV, raw HV, ratio, spread, rank, percentile, vol-of-vol, date, and universe membership before any boolean gates appear. A reviewer should be able to choose one row, recompute every derived field, and explain exactly why it entered or missed the final cohort.

Measure without optimizing the answer

Pin broad threshold families before calculating outcomes. If you examine many cutoffs, record the entire search and reserve a later period for the chosen rule. Prefer a stable region where nearby thresholds tell a similar story over one impressive peak.

Report mean, median, quantiles, hit rate, maximum adverse movement where available, and results by year, sector, and catalyst status. The final screen is only as credible as the cohort history that supports it.

Treat sizing as a separate research question. Vol-of-vol may describe ride quality, though it does not forecast direction by itself. Test whether smaller weights in unstable names improve the distribution without presenting the choice as settled fact.

Add turnover and persistence diagnostics. A screen that replaces every name daily creates a different implementation problem from one whose ranks remain stable for weeks. Measure how often names enter and leave, how scores change around catalysts, and how much of a hypothetical holding period is spent in the selected state. These diagnostics turn a cross-sectional snapshot into something a portfolio process can evaluate.

1. Pin the hypothesis, outcome, and threshold families.
2. Build the timestamped cross-sectional panel.
3. Measure the baseline and every incremental gate.
4. Reserve a final period for the selected screen.
5. Attack catalysts, tiny denominators, tails, and concentration.
6. Publish the current table with data and risk caveats.

Attack the screen

Replace IV rank with percentile, change the realized-volatility window, remove the largest contributors, and split scheduled catalysts from ordinary days. Increase trading frictions and require a minimum sample per name. Record which changes alter the conclusion.

Look for common exposure. A screen full of one sector or a single earnings week is not a diversified volatility cohort. Also check whether the data were revised and whether the optionable universe uses information available at the historical date.

The Stacking the Filters lesson demonstrates the incremental research structure. The course project extends that structure into a complete panel, held-out test, attack memo, and current screen.

Submit a screen and a defense

The final bundle includes the research contract, data dictionary, cohort ladder, validation report, attack memo, current screen, and reproducible Python. Add a one-page note explaining why each displayed column exists and which row deserves the most skepticism.

The grader should choose an unannounced date and ask the student to rebuild the screen. That test exposes hard-coded outputs, hidden manual exclusions, and stale assumptions. The student should be able to trace one selected name from raw response through every transformation, then explain why a nearby rejected name failed the gates. Reproducibility at the row level is the practical defense of the whole artifact.

The quant course projects hub contains related deliverables. A passing project does not need a positive edge or an attractive final cohort. It needs transparent measurement, honest failure analysis, and a screen whose limitations are as visible as its ranks.