Posted in My Daily Journal · 26 October 2026
Before sizing any real-money plan we had every claim in the week's handoff re-derived from raw data by independent reviewers. Nothing was refuted, three numbers moved, and one 2026 result turned out to rest on a single trade.
Why re-check numbers that were already in the ledger?
Because a handoff written by the same process that produced the numbers inherits its blind spots. Five independent reviewers re-priced 222 chain trades from the raw option files and recomputed every figure the go-live plan depended on. The chain's edge was confirmed to the rupee (2021–26 about ₹2,061 per trade at five lots, 2023–26 ₹605–740, win rate 38–42%). Three figures changed: the RBI-meeting dates used for a calendar test were wrong, the stress drawdown was understated, and a filter's count was 2 wins in 36 rather than 1 in 35. None of them changed a decision, but all three would have been wrong on the record otherwise.
What was the single most useful thing the review found?
That the whole of 2026's year-to-date for the intraday chain rested on one trade. On 19 March 2026 one data feed printed 288.25 where the other printed 286.15 against a trail threshold of 288; on one feed the position was stopped for a ₹2,923 loss, on the other it rode to the close for +₹10,371 — a ₹66,500 swing at five lots. Excluding that trade, the year sat between ₹54 and ₹718 per trade depending on the feed. Lesson for anyone running a backtest against a live feed: measure the feed divergence on the real legs, not just the model's optimism — ours was ₹230 per lot per trade in 2026, bigger than the model issue we had been worrying about.
Did anything about costs change?
Yes, and it matters for every intraday system. From 1 April 2026 STT on option sales rose from 0.10% to 0.15% and on futures from 0.02% to 0.05%. On the chain that is about ₹35 per trade at five lots; on index futures it removes roughly 7 points per round trip — more than the entire net of some futures systems we had measured. Every engine now carries a date-gated STT so older numbers stay comparable.
Was the old "stretched market" filter real?
Confirmed, narrowly. On medium-gap days (0.2–0.5%) when the previous close sat 2.5% or more above its 20-day average, the gap-and-go lost on 34 of 36 trades since 2021, about −₹9,400 per trade at five lots, with a permutation p below 0.0005 and no random-removal null matching the gain. Reversing the trade instead of skipping it looked good only because of 2021 — so skip, don't flip. Caveat on the record: only ten of those trades fall after 2023.
Any seasonality worth sizing on?
No. Only January passed even a loose month test (6 of 6 years), and a shuffle null produces about one "passing" month by chance with probability 0.69. Month-switching beat flat on profit-to-drawdown both ways, but the null matched it 6.4% of the time and the whole effect was January 2024. No month sizing.
What else did the exchange change that a bot has to know about?
Two things. The Diwali Muhurat session fell on a Sunday with Tuesday holidays around it, moving the weekly expiry to a Monday. And the Closing Auction Session was in force from August: continuous trading in F&O stocks ends at 15:15, the auction runs to 15:35, derivatives trade to 15:40, and the index is frozen from 15:15 until the auction result. A further circular effective 7 December moves continuous trading back to 15:30 with the auction 15:30–15:40 and index derivatives to 15:45. Any rule that reads the index after 15:14 has to be date-gated to the regime it was tested in.