Posted in My Daily Journal · 22 September 2026
Shorting small caps intraday, dealer-gamma regimes on NIFTY, a portable TradingView system, a 30-session test of AI discretion, and the postmortem of a bot that silently stopped running.
A Reddit claim: short small-cap stocks from the open to the close for ₹50–60k a month per ₹5L. Does intraday decay exist?
The direction is real — small-cap intraday decay mirrors the overnight drift, with the most volatile names losing 35–40 basis points a day net of a 12.5-point cost, while large caps go the other way — but the edge is smaller than a realistic small-cap spread of 40–60 basis points round trip, and the short-sale list and upper-circuit trap make it uncollectable. The claim-test's own compounding block produced a ₹75-trillion artifact (hindsight top-ten times 5× daily compounding), which is how you know an instrument has broken.
Does "dealer gamma exposure" predict NIFTY ranges or pinning, as US order-flow educators argue?
On NIFTY it is implied volatility in a costume. Range prediction correlated at −0.16 against IV's +0.54; the pinning claim was absent (trendiness correlation +0.01); it added nothing as a straddle filter. A retail-seller market does not have the US dealer-gamma structure. One lead did fall out: the straddle earns far more on high-IV mornings (+₹392 versus +₹16 per day in the top tercile) — queued as a filter test.
A popular TradingView system claims portability across timeframes and ten profitable years in ten. Both true?
Portability yes — profit factor 1.40, 1.44 and 1.39 on 10-, 30- and 60-minute bars untouched, because ATR-scaled exits are a real mechanism. Ten of ten only gross of costs on spot; net of slippage and charges it is seven of ten on every timeframe, red in 2022, 2025 and 2026, with costs eating roughly half the gross.
Can an AI day-trade with discretion better than the mechanical systems?
We ran the test instead of arguing. Each morning before 09:15 an AI model is given a fixed, reproducible context — the US close, NIFTY's five-day trend, the previous close, India VIX, whether it is expiry day — and logs one call: buy the ATM call, buy the ATM put, or stay in cash from 09:20 to 15:15. The call is logged before the open so there is no hindsight, and it is scored against the straddle bot's real recorded premiums as a buyer, with stopped legs capping the winners. Our own pre-registered prediction was that it would not clear the bracket. Thirty sessions; verdict at the end. The first genuine call was "cash — low VIX chop, no clear trend".
A bot stopped running for six weeks and no one noticed. How?
Its two scheduled entries had vanished from the schedule when the schedule was restored from a template that had never been updated to include them. The bot had no bug and its journal was complete up to the day it stopped; it simply was never invoked again, while an expired position sat in its state file. An end-of-day watchdog finally flagged the stale state. The fixes were procedural: the template is re-snapshotted after every change, and every bot now has a heartbeat that is checked daily. The lesson for anyone running automation: a system that is silent and a system that is dead look identical from the outside unless something counts the heartbeats.