01 · The Lab
Execution honesty as a method
Most earnings backtests are marked at the mid-quote and filled by assumption. Ours are not. Every claim we publish passes through the same machinery:
Order-book replay
Entry windows sampled at 5-second resolution against the consolidated NBBO; a trade exists only if a worked limit order is touched on two consecutive snapshots. In our data, 79% of candidate spreads never fill — and that is a measurement, not noise.
Causal timestamps
Features are frozen at 15:54:30 ET — the moment an order starts working. Nothing measured after the decision reaches a model. Reference smiles, open interest, and volume follow documented, point-in-time conventions.
Adversarial verification
Every dataset and design freeze is attacked by independent reviewers before use. Two separate model-verification passes have each killed a would-be “edge” our own pipeline produced — a scale defect and a seed artifact. We publish those too.
Sealed holdouts
The most recent quarters are quarantined from every experiment and opened once, against a frozen recipe. Position sizing follows empirical Kelly: no measured edge, no capital — a rule our own results currently enforce.
02 · Findings
What the data says so far
of earnings events realize a smaller move than options implied. The volatility-crush premium is real, persistent, and priced — median realized-to-implied ratio 0.60.
of that premium survives realistic retail execution in defined-risk structures. Across every model family we tested — regularized linear, boosted trees, neural networks — out-of-sample returns are non-positive once fills, spreads, and commissions are honest.
worked put-spread orders fill 42% of the time before earnings; call-spread orders almost never fill. Executable premium selling around earnings is, in practice, a put-side market — a microstructure fact we believe is underappreciated.
The return distribution is the same in every configuration: most trades win small, a few lose big. The binding frontier is loss-tail management, not trade selection — where our current research effort lives.
03 · Research
Priced but Not Harvestable
The Earnings Volatility-Crush Premium under Realistic Retail Execution — High-Frequency NBBO Evidence, Smile-Based Direction Models, a Machine-Learned Decision Layer, and the Kelly Criterion as Arbiter. Working paper, v2, July 2026.
Using 1,092 earnings events reconstructed from 63 million OPRA quotes, we confirm the announcement volatility premium in prices — and show that under microstructure-accurate execution, every implementable defined-risk configuration we test earns non-positive returns, with an honest Kelly allocation of exactly zero. The negative result is the contribution: it is the baseline any credible claim of edge in event-driven option selling must beat.
Request the full paper04 · The Options Lab
The product, rebuilt on honest rails
Codify’s Event-Driven Options Lab — historical move distributions, one-click strategy backtests, and ranked structures for 200+ tickers — is being rebuilt on top of this research infrastructure, so every number it shows survives the same execution honesty you just read about. If you want early access when it returns:
Join the waitlist