The term structure did not tell us when to sell

The hypothesis

Our scoring model runs on fixed scales and never asks whether today is a good day to sell a put at all. That is correct for a score — a 78 has to mean the same thing in March as in October — but it quietly became true of the whole pipeline.

A 0.20-delta strike is sold on the premise that it carries roughly 80% odds. That premise is a claim about the distribution of the next 45 days, and the market publishes its own opinion of that distribution twice: once at 30 days and once at 90. When the near-dated index trades above the far-dated one, the market is saying it expects the next month to be worse than the quarter.

If that is information, a put sold into an inverted curve should be breached more often. We tested it over twelve years and 55,470 entries.

The answer was no, and it pointed the wrong way

Excluding 2020 changes nothing (29.1% against 36.7%). The middle flat band came in highest of the four at 36.7%, which is not a monotone anything. There is no signal here in either direction.

Why — and it is mechanical, not mysterious

The strike is placed off trailing realized volatility. So a calm market gets a tight strike and a frightened market gets a wide one, automatically:

Hold the cushion still — compare only entries in the 5–10% band — and the gap collapses to 32.5% against 34.7%, which is nothing at all. There is no residual information in the term structure once strike placement has had its say.

The live scan should be better protected than this test, not worse: it places strikes off the vendor’s own implied volatility, which reacts to a change in regime faster than a thirty-day trailing estimate does.

Two ways to get this answer wrong

Using the entry day’s close. The scan fires at 10:30 ET and the day’s closing index level does not exist yet. Bucketing an entry by the close of the session it was entered on is a look-ahead of exactly the kind that makes a regime signal look prophetic, and it is one line to get wrong.

Treating a missing reading as a regime. If either index is absent for the prior session the trade leaves the split entirely, rather than falling into a bucket it was never measured into.

We kept the split anyway

The classification still runs on every backtest. If a later change to strike selection ever stops absorbing the regime, the row that would show it is already there — and a measurement is much cheaper to keep than to rebuild after the fact.

Nothing in the product gates on this today, and nothing should.