Posted in My Daily Journal · 16 October 2026
The losing-streak statistics every systematic trader should know before going live, a null result from a time-series foundation model, and a pre-registered battery that found two trades worth not taking.
What do the losing streaks of a gap-and-go buyer look like over twelve years?
Median losing streak 2, 90th percentile 4, longest 13 in a row (June–July 2022, −₹22k at one lot). Twenty-six streaks of five or more, nine of seven or more. The maximum drawdown was −₹57k: peak in January 2023, trough in May 2024 — 169 trades and 343 sessions — and the old peak was only regained in October 2024, 632 calendar days later. Worst day −₹8.7k, best +₹28.7k, about nine trades a month. If a trader cannot sit through a 500-day drawdown with lots unchanged, the system's edge never reaches them.
You rented a GPU to run a time-series foundation model on the first 30 minutes of each day. Did a 09:45 forecast help the bots?
No, as pre-registered. Across 2,790 sessions the model's direction call for 09:45→15:15 was right 50.9% of the time; the simple 09:15→09:45 momentum baseline did better at 52.3%. Trades where the forecast agreed with the bot's direction earned less than the average trade (₹228 versus ₹258). Every criterion failed. The conclusion on record: our edge is in rules and execution, not in a forecaster — and that was the same answer every "read the day at 09:45" approach gave.
You went through the breakout literature — Zarattini and Aziz on opening-range breakouts, Crabel, Unger — looking for ways to sharpen gap-and-go. What transferred?
Nothing from the exit toolbox. Spot stops at the range midpoint or breakout level, 60–90-minute time stops, break-even after one range width, scaling out half at one width — all lost or tied against the premium trail with a 15:15 exit, with the half-off rule costing ₹125 per trade because the runners are the edge. The current exit design is a plateau. Narrow-range setups (NR7, inside days), which Crabel documents on US futures, were negative on NIFTY options.
So where were the gains?
In not taking two kinds of trade. First, days where the gap was already filled before the breakout — the spot touched the previous close inside the opening range — lost ₹216 per trade over 319 trades, and were 55% of tiny-gap days; without them the rest earned ₹419 per trade with profit-to-drawdown 10.4. A gap that fails inside the range is not a gap-and-go day. Second, breakouts after noon: 92 trades, −₹404 each, negative in both halves. A third signal — rich implied volatility at 09:20 versus its trailing median — carried information too (₹553 per trade versus ₹4 on cheap days) but cut too many trades and was post-hoc in its interactions, so it was logged for its own pre-registered test rather than adopted.
Why were those two filters acceptable when so many others were rejected?
They remove trades without adding risk, they passed in both halves of the sample, they have a plain mechanism, and together they lifted the twelve-year result from +₹3.27L to +₹4.15L (P/DD 9.8) while turning the losing 2025 positive. They were built as switches the operator can turn on, tested against 148 real cached days where every kept trade matched the backtest to within ₹5, and then it was the trader's decision to enable them — not the researcher's.