Posted in My Daily Journal · 18 October 2026
The weekly ITM buyer's simulation now reproduces the bot's own log, a stock-option door that looks good on the underlying dies at the option's break-even spread, BankNifty and option selling are measured as a possible third leg, and a gap filter suggested in the morning is withdrawn by evening with four reasons why.
Does the simulator finally match what the bot actually did?
Yes. The rebuilt live-faithful simulation of the weekly ITM buyer produces 445 trades since 2015 at an average of ₹752 per trade, about ₹3.35 lakh net at one lot, with a worst drawdown of ₹27.9k, a profit-to-drawdown ratio of 12 and profits in 8 of 9 years. Replayed against the bot's own log it matched 444 of 445 trades, which is the first time the paper record and the simulation have been the same object. The live plug was then rebuilt after re-reading ten earlier live mistakes, run through fifteen simulated-broker failure scenarios, and placed behind a 60-session paper gate before it may touch the live book.
Stock options looked promising on the underlying. Do they survive once the option itself is priced?
On the underlying, the ≥1% door earned +0.07–0.10% per trade in 2015–22 and +0.085% out of sample in 2022–26, so the signal is real. Emulated as an option position it loses ₹259 to ₹684 per trade, because a stock option's spread and time value eat a move that small. The direction splits that looked meaningful flipped sign between datasets, which is the signature of noise, and a VIX feature added a Spearman correlation of 0.01–0.02 without making the gated days any bigger, so neither was adopted.
So we bought real stock-option premiums. What did they show?
The archive covers 55,126 contracts across 91 symbols, later widened to 206 symbols and about 935 million rows, and it validates against intrinsic value at expiry to within 0.004–0.01%, with a simple rule that a bar counts only if it has volume or open interest. On 21 liquid names the door made 197 trades at +₹84 per trade assuming a 0.25% spread, but its break-even spread is 0.39%; the day-ORB variant lost ₹346 per trade, and later ₹278 across 27,510 trades. The gap-down door went from +₹727 on 90 trades to +₹198 on 90 names to −₹142 on 117 names, a textbook case of a small-sample winner diluting toward zero as the universe grows.
One filter — NIFTY below yesterday's low at 09:44 — turned the stock door around. Why not trade that?
With the filter on, the stock door earned ₹733 per trade against −₹678 with it off (264 versus 85 trades), and the same split reproduced on the underlying in both eras, +0.186% against +0.051%. But on the full archive the break-even spread fell from 1.33% to 0.30% per side, below the kill line we had pre-registered before looking, so the stock-option door is not tradeable at real spreads. A filter that lifts a losing book to a marginal one is a finding, not a strategy.
What about BankNifty as the third leg, on the days the NIFTY book sits idle?
Nine candidate mechanisms for a third leg were asked to clear the pre-registered bar on idle days and none did; cascading into them was outright negative, and the only piece banked was a filter, index beyond yesterday's extreme, with a day-clustered t of 3.6. BankNifty on its own gap days is good: 373 trades, ₹913 per trade, profit-to-drawdown 10.7, 7 of 7 years positive. The problem is overlap: only 12% of NIFTY-idle days see a BankNifty gap of 0.5% or more, and once monthly time value and a 0.5% spread are charged the idle-day cell loses ₹414 per trade with a stacked profit-to-drawdown of 0.09. A ceiling of roughly 50 trading days a year for the buying book appears to be real rather than an artefact.
And option selling?
Selling the 9:20 straddle on idle days lost about ₹20 a day over 1,753 days once 3 points of slippage were charged, so there is nothing there. Expiry day is different: +₹497 per expiry day, profit-to-drawdown 13.6, 9 of 12 years positive, worst day −₹7,489, and the edge lives only on the day of expiry (+₹497 at zero days to expiry against −₹90 one day out). That gives a fleet shape of a buyer active about 50 days a year plus an expiry seller active about 48.
You suggested restricting the live buyer to gaps of 0.5% or more, then withdrew it the same evening. What went wrong?
Four things, each enough on its own: the comparison window was an artefact, the simulation omitted filters the live bot actually applies, the two sides were run at different lot counts, and slippage was charged at 0.5% where the measured figure is about 0.1% and unbiased, roughly 1.2% of premium as pure noise. The gap buckets tell the real story — under 0.2% earned ₹1,146 per trade, 0.2–0.5% ₹274, 0.5–1% ₹2,025 and above 1% ₹1,722 — and 5 of the live book's 6 trades in its best fortnight came on gaps under 0.5%, so the filter would have removed the trades that made the money. The standing rule from this: never judge a live bot against a simulation unless the window is complete, the environment is its own, the spread is the measured one, and the live journal has been read first.