Posted in My Daily Journal · 1 November 2026
A Sunday spent checking what the Diwali muhurat session does to a pivot system, and running a weekly option-buying bot through hundreds of simulated days of broker failures.
Diwali muhurat trading is a one-hour Sunday session. Should an overnight pivot system trade it?
No, and ours doesn't. In all seven muhurat sessions since 2019 the index opened above the standing pivot high — a festive gap-up that faded in six of the seven — so a pivot rule would have bought in the first minute every time. Marked at the next session's open that was +₹43k over seven years; by the next close it was −₹33k. Noise. The real finding was elsewhere: the six 10-minute bars of a muhurat session change the next morning's pivot high in six of seven years, because the spike becomes the new pivot. A system that takes bars from a broker's history feed should filter them to the regular session, or its first trading day after Diwali works from a pivot the backtest never had.
What is the point of a stress test the night before a bot goes live, if the logic was already verified?
Logic review finds what you think to look for; fault injection finds what you didn't. We drove the bot through 215 simulated days three times with a fake broker that randomly rejected orders, timed out, filled with unknown fate and went unreachable, and checked after every day that the book matched the broker, that no position was ever net short, that paper days produced no real orders and that quantities were always whole lots. It found one real hole: a buy whose fate stayed unknown — treated as real, correctly — followed by a stop exit that sold the book's quantity while the broker held nothing. On a real account that sells an option you don't own and leaves you short. The fix is the rule the older bot already had: before any sell or roll, read the broker's actual holding and sell at most that; holding zero closes the book on record with no order.
Did anything else come out of the fault injection?
A roll — selling the expiring contract and buying next week's — was re-buying for real on a day the operator's switch said paper, because a position kept its venue for its whole life. A roll is new real buying, so now it follows the switch and stops if anything is halted. Both changes passed the original 29 checks, a transition test from a paper position into a live one, four roll-day scenarios and three chaos runs with zero violations.
You keep saying "pre-registered". Why?
Because the easiest way to fool yourself with backtests is to choose the metric after you have seen the result. Every test here is written down first — the rule, the data, the pass bar and the prediction — and the result is recorded whether it passes or fails. Most fail. The few that pass have earned it.