# Crossover-to-Crossover — strategy prospectus & test plan **Date:** 2026-09-07 · **Author:** Claude, assessing Nitin's proposal. **Status:** proposal. Nothing built. This is the design + how to prove it. ## The idea, in one line Always be in the market in the direction of the last confirmed EMA 9/20 crossover; flip on the next confirmed opposite crossover; no intermediate strategy exit. The crossover is treated as a regime change, not a scalp trigger. ## Verdict: worth building — as a SEPARATE, MEASURED strategy It attacks the exact disease the six-week analysis found: the production strategy captured only ₹66 of a ₹254 peak on 2026-09-07, and gave back ₹4,621 of MFE across the prior week, because it exits on geometry long before the move ends. "Hold until the next crossover" is the cleanest possible expression of "let the winner run." Nitin's instinct to build it ISOLATED — separate strategy, compared on the same data, production untouched — is exactly right and matches this project's whole discipline. ## Three hard truths before a line of code ### 1. This platform buys OPTIONS, and options bleed. (The big one.) The proposal describes "Long/Short" as if holding the index. Production does not short and does not hold the index — a bullish signal BUYS A CALL, a bearish signal BUYS A PUT. So "hold Long crossover-to-crossover" means **holding a long call whose time value decays every minute (theta)**. - The current strategy is in a trade ~3–13 minutes. This strategy is in a trade from one crossover to the next — often 20–60 min, sometimes hours. - A long option held through a flat/choppy stretch bleeds theta even when direction is right and price is unchanged. The longer hold that captures a big trend is the SAME mechanism that bleeds you in a range. - The flip (close call, open put) is one round-trip in cost — the same cadence as now — so cost-per-flip is not worse; cost-per-HOUR-held is new, and only the excursion data can say whether the bigger captured moves outweigh it. This is not a reason not to test it. It is the reason the test must be on OPTION premium P&L, never on index points — an index backtest would look brilliant and lie to you. ### 2. Always-in-the-market pays for every false crossover. On a 1-minute chart the 9/20 whipsaws — look at your own 13:00–13:20 zone, several near-crosses in a tight range. "No intermediate exit" also means "no filter refuses a bad crossover" — every cross flips the position, both directions, all day. On a trending day (your 12:20→13:00 down-leg, then 13:00→14:00 up-leg) this is gold. On a chop day it is a series of round-trip fees on false signals. Your own hypothesis list already names this ("performs worse in sideways/choppy") — the test must REPORT the trending-vs-chop split, not average it away. ### 3. "Hold until next crossover" still needs a hard safety stop. A gap or a crash can run against a held position for a long time before any opposite crossover forms. You already carved out "mandatory risk controls" — good. Keep exactly one strategy-level safety: a catastrophic hard stop (wide, e.g. a fixed % of premium) that exists ONLY to cap a disaster, not to manage the trade. Everything else — regime veto, m4, trail, target, spike — is OFF for this strategy by design. Name that one stop explicitly so the test knows the difference between "the strategy held" and "the safety net caught a catastrophe." ## How to test it — the tools we deliberately KEPT are for exactly this We spent two weeks removing research machinery and keeping two things: Historical Replay and shadow_mode. This proposal is what they were kept for. The test path, in order: 1. **Historical Replay first (zero risk, zero broker calls).** Register `crossover_to_crossover` as a second strategy, run it over the SAME historical 1m corpus the production strategy sees, priced on OPTION premium (using the captured option candles where present; state coverage honestly where not). Produce Nitin's full comparison table — trades, win rate, avg win/loss, MFE, MAE, holding time, profit factor, max drawdown, long vs short (call vs put), trending vs chop, premature-exits- avoided, reversals. n stated before any verdict; Inconclusive allowed. 2. **Shadow mode second (live, still zero orders).** If replay is promising, run it in shadow alongside production for a stated number of sessions — it evaluates the same live crossovers, records what it WOULD have done, places nothing. Compare shadow vs production on identical live signals. 3. **Paper trading only if 1 and 2 both hold**, and even then as its own registered strategy the user selects — never replacing ema_crossover_v2. ## Architecture (keeps production untouched) - A new registered strategy name (`crossover_to_crossover`), its own manifest, its own minimal filter set (only what's mandatory), scoped to its own name via the applies_to mechanism the filter-scope audit pins. It inherits NONE of ema_crossover_v2's exit policies by default. - It reuses the same crossover DETECTION as ema_crossover_v2 (the confirmed 9/20 cross) so the two differ ONLY in the held/exit behaviour — that is what makes the comparison clean. Same entry trigger, opposite exit philosophy. - It is a STATE MACHINE: flat → long-call → (opposite cross) → put → (opposite cross) → call … The state persists across ticks; only a confirmed opposite crossover or the safety stop changes it. ## Where this sits in the roadmap This is the "let winners run" hypothesis the Phase 1 context study was going to raise, made concrete and testable. It does NOT jump the queue of structural fixes (H-7 risk envelope is still next for production). It rides the SAME measurement infrastructure — Phase 0a's excursion columns, the replay corpus — so it is cheapest to build AFTER Phase 0a lands. Build order: Phase 0a → this strategy in replay → read the comparison → decide shadow. ## The one honest expectation This may well win in trends and lose in chop, and the net depends entirely on how NIFTY's days split between the two — which the backtest will show and no one can promise in advance. The value of the test is the ANSWER, not the assumption. If it captures materially more of the move net of theta and whipsaw, it becomes a real candidate. If it bleeds in chop more than it gains in trends, we learned that for the price of a backtest instead of a month of paper losses.