Project H. Where AI meets Indian F&O.
A research engine that ranks options trades using a calibrated stack of machine learning, real-time news correlation, options-flow, and event-driven signals. This site is the public log — how the system thinks, not what it holds.
One composite score, fourteen ablation-toggleable factors.
Each factor lifts or trims the base LambdaRank conviction multiplicatively. None of them get to overrule the model — they move it. Every one of them is independently switchable so we can measure the contribution of each block, every trading day.
LambdaRank base
Hybrid LightGBM ranks the F&O universe daily on 147 features. Output is the entry probability — every boost below operates on top of this rank.
News Intelligence Engine v6
Dual-head model on a 31K-headline corpus + macro propagation chains. Outputs intraday + daily + weekly conviction scores per name.
Pre-market news magnitude
Catalyst-weighted expected-move predictor combining urgency, event type, and ATM IV. Replaces a binary boost with a calibrated lift.
Broker consensus alignment
5-source broker signal (NSE OI spurts, analyst upgrades, Trendlyne ratings). Lifts composite when consensus agrees with LGB direction.
Quarterly fundamentals
Screener.in P&L scraper — EPS beat, revenue beat, QoQ profit growth. Lifts names that just printed strong fundamentals.
Real-time candle patterns
11 pattern recogniser on 5-min + 15-min bars. Hammer, engulfing, three-soldiers, breakout-bar — gated by confidence threshold.
Live OI flow
Per-symbol OI buildup direction from the Fyers options chain. Tracks where dealers are positioning ahead of an underlying move.
Volume surge detection
Z-scored 5-min volume vs 20-day baseline. Surfaces names where unusual size is hitting the tape before the price reaches it.
Put-Call Ratio regime
Contrarian Nifty PCR overlay on call/put bias — high PCR (puts dominant) shifts portfolio toward CALL-heavy and vice versa.
Ownership overlay
Trendlyne pledge % + promoter QoQ change + insider 30d net + bulk-deal 5d net. Static per-symbol ownership signal.
Durability / Valuation / Momentum
Trendlyne's proprietary 0-100 three-score average mapped to ±10% multiplier. Filters out structurally weak names from the top-decile.
SmartOptions daily insights
SmartOptions algorithmic flags — 40-50 stocks/day surfaced for unusual activity (earnings beat, breakout, OI spike). Binary +5% lift when flagged.
30-day insider net buy
Trendlyne insider trades (SAST filings) — net acquisition vs disposal ratio over 30 days. Smart-money directional signal with low false-positive rate.
Insider × Bulk confluence
Fires only when BOTH insider 30d net buy AND bulk-deal 5d activity exceed ±20% magnitude AND agree in direction. ±15% boost — designed-rare, high-precision.
Latest entry
The AI/ML approach to options selection — why we picked ranking over classification
Picking the right machine-learning objective for short-horizon options trading is non-obvious. Classification asks 'will this stock go up?' — the wrong question. Ranking asks 'which of these 146 stocks will outperform the others?' — the right one. Why the distinction matters and how it changes everything downstream.
More from the engineering log
Why news correlation matters for short-horizon options trades
Most retail trading systems read the chart and ignore the wire. The thesis behind Project H's news layer: pre-market and intraday news flow leads price action by minutes — sometimes hours — and that lead is exactly the window where short-horizon options strategies live or die.
Project H — Building an AI-driven options trading engine
An algorithmic trading research project for the Indian F&O market. Combines machine learning, news correlation, and event-driven signals to make systematic CALL/PUT decisions across the Nifty 500 derivatives universe.