A night of closing doors — textbook futures rules, a narrative swarm, indicator learners, convex exits — and one that opened

Zero of twenty-one. Opening-range breakouts, Supertrend and EMA stop-and-reverse on 5-minute bars, Bollinger squeezes and fades, floor pivots, Camarilla, gap fades and gap-goes, NR7 and inside-day breaks, last-hour momentum, first-hour fades — none was close to the bar at the pre-registered cost of 0.05% per round trip, and still none at a realistic 0.03%. The trend-following families had gross edge (Supertrend +₹17.6L before costs) at ₹240–350 per trade against a ₹480 round-trip cost: real drift, too thin to cross the toll. The dominant finding was the era split — nearly every cell was positive in 2015–20 and negative in 2021–26. Whatever drift those rules harvested is gone. Family closed; no re-tuning to rescue it.

No — we checked. Supertrend and EMA stop-and-reverse re-run on 15-, 30- and 60-minute bars fell to about one trade a day as expected, but their gross edge shrank with the bar and net stayed deeply negative. The pivot system's profit-to-drawdown of 4.75 comes from its pivot-and-box logic, not from trading less often.

No signal. We built a clean out-of-sample test — a model with a knowledge cutoff before the test window, archived headlines, 366 sessions of 2025–26, baselines scored identically. Morning calls hit 40–41%, below the 45% fair chance under the scoring bands. Evening calls hit 48%, not significant, and the paper profit was ten days' worth: excluding the ten biggest days, one arm made +₹3k and the other −₹41k. Twenty-four personas agreed with four personas 92% of the time — diffusion produced consensus, not information.

Null, null, and instructive. A gradient-boosted model and a transformer on about sixty indicator features at 09:44 hit 51–53% and lost money net; the first 15–60 minutes of one index does not lead the other (they move together within the same minute). The convex-exit battery — thirty-six cells of time, pattern-at-level, contraction and event entries with chandelier trails — had zero passes: the exits reshaped the distribution (payoffs 1.6–3.5) but gross edge per trade was 5–7 points against a 12-point cost. Candlestick patterns at levels carried nothing on 5-minute bars. One lead fell out: the first session after a market holiday, where information accumulates while closed, had a range 11% above normal and positive expectancy in both halves of the sample.

Magnitude is predictable, and that was the night's real finding. A walk-forward model predicting the day's range relative to its 20-day ATR from the first 30 minutes reached a Spearman correlation of 0.43–0.47 with the realised range; days it flagged realised 1.33–1.58× ATR versus 0.93 for the rest. A delayed opening-range breakout taken only on flagged days earned +0.28R per trade at a 1.9 payoff with profit-to-drawdown 5.0, positive in both halves — and the identical pipeline retrained from scratch on BankNifty replicated it on all four model seeds. Ungated, the same entry lost on BankNifty; the gate does the work. Status: a candidate for a pre-registered paper test — about 12–15 trades a year per index, not a daily system.

Neither. Using the forecast to skip the straddle's predicted-expansion days gave +₹2.5L against +₹3.2L for the plain rule — the forecast is not a straddle filter. A quiet-day fade failed at every threshold, a long straddle on predicted-expansion days lost (IV already prices the expansion), and the frequency experiments that followed through the night — a buyer dispatcher, re-entries, a realised-range gate at 10:15, confluence sizing — all ended the same way: every lever that added trades diluted quality. The audited buyer book that survived takes about 49 trades a year at ₹750–800 per trade with profit-to-drawdown between 9 and 12; two to three trades a week at that quality was not available. One pricing lesson from the audit: taking the strike from the morning's reference rather than from the spot at entry had flattered an ATM cell by a factor of three — strikes must come from the price at the moment of entry.

Not for direction (5 of 64 cells significant, about chance, none repeating across years), not as a filter, and only marginally for magnitude — and that marginal gain was really implied-volatility change, which our premium data already carries. The data purchase was declined.