Thirty trades prove nothing

Telling a real 55% win rate apart from a coin flip takes hundreds of trades. At thirty, the 95% interval around a measured win rate is roughly plus or minus 18 points, which spans everything from a losing system to a good one.

CI = p̂ ± 1.96 √(p̂(1−p̂)/n)

Why to be skeptical of the analysts

• Lopez-Lira and Tang. A GPT-4 long-short strategy reached a high annualised Sharpe, but returns decline as adoption rises and the edge does not survive realistic transaction costs.
• Kirtac and Germano, 2024. A GPT-3 sentiment model reached 74.4% accuracy on US news with a Sharpe of 3.05 after 10bps of costs.
• Shobayo and others, 2024. Plain logistic regression at 81.83% accuracy beat GPT-4 at 54.19% on the Nigerian all-share index. A simple baseline beat the large model.

All of that evidence is from other markets. There is no published evidence that these edges exist for Indian intraday equities. The edge here may be zero. That is why the models never touch a number, why every probability is calibrated against real outcomes, and why the cost gate can veto any of them.

What this system is honest about

• Opening-range breakout is a weak, decaying edge. Realistic win rates sit between 40% and 60%, and the value comes from reward-to-risk and trend-day capture, not from hit rate.
• Every probability shown is a calibrated estimate with a measured error, not a forecast.
• A candle that touches both the stop and the target is graded as a loss, because minute data cannot say which came first.
• No order is ever placed by this software. Every order is reviewed and confirmed by you inside Kite.