Most option sellers have a feeling about "good days to sell". This article is about testing one such feeling properly: a filter that only allows short premium when recent volatility is high. It passed more tests than almost anything else in our research, and it is still listed on our Projects page as retired on evidence. Both halves of that story are worth understanding.

The rule

The filter is deliberately simple. Each morning, compute NIFTY's average realised volatility over the previous 20 sessions. If that number sits in the top third of its own history, short premium is allowed that day; otherwise, stay out.

Two details matter more than they look:

  • Only past data is used. The 20-session average stops at yesterday, and the "top third" threshold is recalculated each day from history available at that point. Nothing from the future leaks in.
  • It is computed on every trading day, not just traded days. An early version of the test accidentally built the volatility series only from sessions where a trade was placed. That quietly stretched "20 sessions" across much more calendar time and distorted the results. The rule we kept afterwards: build any regime variable on the full daily series, then attach it to trades.

Why it might work

The mechanism was measured, not assumed. On qualifying days, a short iron condor collected more credit while the average trade's worst move against it was smaller. More premium for less travel.

That fits a broader pattern in options markets: when implied volatility is elevated after a turbulent stretch, it tends to overstate what happens next by more than usual. Sellers are paid best when others are most nervous.

The risk finding was arguably more interesting than the return finding. The condor's worst sessions did not happen on already-volatile days. They happened when volatility expanded suddenly from a calm base — days the filter classified as "normal". Inside the filter, the worst losses were much smaller, and so was the peak-to-trough drawdown.

The tests it passed

Because the result looked good, it was attacked hard:

  • Placebo by rotation. The filter's on/off labels were shifted in time against the results hundreds of times, preserving how qualifying days cluster together. The real result sat far outside the placebo distribution.
  • Live parameter choice. Instead of fixing the 20-day lookback in hindsight, a walk-forward procedure re-chose the lookback at every step using only past data.
  • Cost stress. It survived tripling the modelled slippage.
  • Dropping the best days. It survived removing its five best sessions.
  • Bootstrap and multiple-testing adjustment. Block-bootstrap confidence intervals excluded zero, and a deflated performance statistic, accounting for the number of variants tried, cleared 0.9 on the main specification — strong, though short of the usual 0.95 bar.

By the standards of most published backtests, that is a strong result.

Why it was still retired

Three findings outweighed everything above.

1. The threshold didn't transfer between instruments. NIFTY and BANKNIFTY preferred opposite ways of measuring the same "top third" threshold — against all prior history, or against a rolling window — and the version that worked on one did not hold up on the other. A rule whose result depends on how you measure its own threshold is not a validated rule.

2. It almost never trades. In a quarterly walk-forward on BANKNIFTY, choosing the settings each quarter from past data only, there were five consecutive calendar years with no qualifying sessions at all. On NIFTY there were only a handful of out-of-sample quarters to test, and in some of them the filter never fired.

3. The evidence comes down to about two events. On both instruments, qualifying days clustered in a few volatility episodes. NIFTY and BANKNIFTY are highly correlated, so an episode that shows up in both is one event seen twice, not two confirmations. Strip that away and the out-of-sample evidence is roughly two market episodes.

What it actually is: an avoidance rule

The most revealing number from the walk-forward was where the advantage came from. Compared with selling premium every day, roughly 75–80% of the filter's advantage came from losses avoided by staying out, not from profits earned while in.

That reframes the whole result. The filter is not a way to make money from short premium. It is evidence that selling premium continuously is what loses: the market pays sellers in specific high-volatility windows and charges them the rest of the time.

What to take away

  • A filter can pass placebos, cost stress and walk-forward parameter checks and still not be a strategy. Transferability, trade frequency and the number of independent episodes behind the result matter as much as any p-value.
  • "When to stay out" can be a real and useful finding even when "when to get in" isn't.
  • If you sell premium every day by habit, the more honest question is not which filter to add, but whether the days you are trading are ones the market is actually paying for.