Why news correlation matters for short-horizon options trades
Charts are downstream of news
Every meaningful intraday move in an Indian F&O stock has a cause. Most causes leave a trail in the news flow before they leave a trail on the chart:
- A corporate filing on the BSE bulletin board at 9:05 AM that traders haven't priced in yet
- A regulatory headline overnight that re-rates an entire sector at 9:15 open
- A management guidance change buried in an investor-presentation deck released the prior evening
- An institutional allocation report that telegraphs FII/DII direction
- A commodity-price shock 4 hours ahead in another time zone that hasn't propagated to the local sector ETFs
Charts catch up. By the time the price chart "tells you" something, the news wire and the smart money have both already known. The arbitrage isn't in the chart — it's in how fast you can ingest, classify, and weight the news that's about to hit the chart.
This is the central thesis of Project H's news layer. The market is efficient over weeks. It's much less efficient in the first 60 minutes of a trading session, when news from the prior 18 hours is still digesting.
What "news correlation" actually means here
Three practical jobs the news layer has to do:
1. Aggregate broadly, not deeply
A single news source is a single failure mode. Project H pulls from a wide mix — exchange filings, real-time wire services, broker action signals, social-sentiment streams, macro event calendars — and treats redundancy as a feature, not a bug. When two independent sources agree on a direction, the signal strengthens. When they disagree, the system is smart enough to abstain rather than guess.
2. Score for materiality, not just sentiment
Not every headline matters equally. "Reliance reports earnings" needs different weight from "Reliance acquires nano startup for ₹50cr." Sentiment-only models miss this — they read both as positive and weight them similarly. Materiality scoring is harder: it asks "would this news, if you'd seen it 30 minutes before the market priced it, have changed your position?" That question is the whole game in news-driven trading.
Modern LLMs are unusually good at materiality classification — better than older NLP approaches like FinBERT in our experience — and that capability is what makes the news layer competitive on a single-developer budget.
3. Cross-validate against price
Every news signal eventually has to be tested against realised forward returns. A source that scores headlines beautifully but doesn't predict price is decoration. The audit pipeline runs continuously: per-source information coefficient (IC) versus next-day returns, across multiple time windows. Signals that don't earn their place get cut. Signals that do get weighted up.
What makes this work for short-horizon options
Options have two characteristics that matter:
-
Theta decay — every minute a position is held, premium leaks. So you can't be "right but slow." Being right at 2 PM about a position you took at 9:30 doesn't matter if the option premium decayed faster than the move developed.
-
Convexity — when you're right with timing, the payoff is non-linear. A 2% directional move in the underlying can be a 50%+ move in the option premium.
Together these mean timing is the alpha. Not stock selection alone — selection plus the right time-of-day to enter and exit. The news layer's job is to lean the timing distribution: stocks where pre-market news suggests a directional move in the first hour get a directional bias; stocks with no fresh news catalysts get a different (often more conservative) treatment.
Why this is hard with retail tools
Most retail platforms either:
- Show you news as a side panel, decoupled from the trading layer
- Sell you a sentiment score with no path to inspect or modify the inputs
- Offer technical-analysis-only systems that ignore the wire entirely
Project H takes a different stance: the news layer and the AI ranking layer are first-class peers, designed to talk to each other. A news event that arrives at 9:08 AM can shift the ranker's CALL/PUT split by the time the 9:30 entry fires. That round-trip — from headline to position-sizing decision — happens automatically, and the system explains itself in the post-trade log.
The tradeoff: this requires owning the engineering, not just consuming a SaaS. Which is why Project H exists as an open engineering log. Anyone trying to build something similar can read what's been built, what worked, what didn't, and skip the same dead ends.
What we're publishing next
- The AI/ML side of the system — model architecture, validation methodology, what works for short-horizon directional ranking
- The risk-management mechanics — why exits matter as much as entries
- The 30-day paper-trading results once the live arc starts producing signal
The trading engine is the what. The news layer is one of the whys it works. The blog is the how we got there.