An execution-quality report card for options traders, and an autonomous agent that trades to earn a good one.
Submission: one-page write-up — AI logic, risk gates, and Alpaca
infrastructure. Alpaca paper trading account PA32CGA2U1DY.
Retail options traders lose more to the spread than to being wrong. Bryzgalova, Pavlova and Sikorskaya (Journal of Finance, 2023) put aggregate retail options losses at $2.10bn over Nov 2019 – Jun 2021 — against $6.4bn of trading costs over the same period. Their words: "the bulk of the losses comes from the indirect costs of trading."
And nobody is required to show those costs. Rule 605 mandates execution-quality disclosure — effective spread, price improvement, E/Q — for NMS stocks, and 17 CFR 242.600 defines an NMS stock as "any NMS security other than an option." Equities got execution-quality disclosure in 2001. Options never did.
A guided tour for reviewers → — the five things worth looking at, with the exact command to reproduce each one.
The explorable walkthrough → — what a strike actually costs, why the numbers are trustworthy, and the working behind them. Seven interactions over real measured data; plain English on the surface, the full statistics in the appendix.
The agent's decision log → — every decision it made, published as it ran, refusals included. Hash-chained, so editing any earlier entry breaks every hash after it.
A measurement method, six findings, and an agent that acts on them.
The method. Alpaca's paper engine fills a marketable order at the true NBBO independent of your limit price — verified here with limits spanning $1.00 producing fills spanning $0.01. Every fill is therefore an exact reading of the real market, which makes ground truth free.
The findings, all measured on live markets:
| Result | |
|---|---|
| Free feed accuracy | Midpoint approximately unbiased; spread error flips sign with liquidity — 0.67× true width on liquid contracts, 1.94× on mid-liquidity |
| Fair value from the underlying | Slope 0.89, t = 24.1, R² = 0.726 at a 40-second horizon on SPY 0DTE — and the signal only clears the feed's noise floor at ~20 seconds |
| Execution timing | Null. +$0.15/contract, t = +0.31, CI [−$0.82, +$1.12] — excluding the published ~$2.50 effect |
| Contract selection | ~80× larger than timing. Deep-ITM calls cost 102–116× more in spread than buying the equivalent shares outright |
| Size-weighted mid | Null. The microprice does not beat the plain mid against true NBBO (n=64; RMSE $0.229 vs $0.141; paired p = 0.30), and on gate-passing contracts the two are indistinguishable |
| What the refusals cost | Every veto is journalled with the quotes it refused and settled against the underlying's close, so the gates are scored rather than asserted |
The agent. Defined-risk SPY verticals, atomic multi-leg, flat overnight, behind a pre-trade risk kernel whose gates map to SEC Rule 15c3-5.
tools/pricing.py Black-Scholes, implied vol, greeks on an intraday clock
(Alpaca returns none at 0DTE)
tools/alpaca_io.py resilient HTTP, append-only journal, verified flatten
tools/risk_kernel.py 14 pre-trade gates, 23 unit tests + 10,000 random cases
tools/executor.py the only component that can move an order
tools/*_study.py the experiments behind the findings above
tools/opportunity_cost.py settles every refused trade against the close
tools/statlib.py the statistics, written out so nothing needs installing
config/risk.json frozen risk limits
PRE-REGISTRATION.md rules fixed before the first trade
site/ source for the public explorable
tools/build_site.py regenerates docs/index.html from results/
app.py Streamlit operator view (live data during a session)
./tools/run_tests.sh
Runs the unit tests, ten thousand randomised cases against the risk kernel's invariants, the arithmetic behind the refusal ledger, and recomputes the analyses from the recorded data. Standard library only — no install, no network, no account.
Paper trading only. Simulated results are hypothetical and do not represent actual trading. Nothing here is investment advice.