OpenBracket — for investors

Status: In active development. Nothing here is a recommendation. No real-money trades happen yet.

What it does, in one sentence

Every weekday morning at 8:15 AM IST — about an hour before the Indian stock market opens — OpenBracket writes a ranked list of NIFTY 100 stocks with a low-to-high bracket showing where each one is likely to close, and how confident the model is in that bracket.

Why an hour before open?

Because by 8:15 AM India time, a lot has already happened that the Indian market hasn't priced in yet:

• The US market closed about six hours earlier.
• Asian markets (Singapore, Hong Kong, Tokyo) have been trading for an hour or two.
• Overnight news, oil prices, and currency moves are all available.
• Indian financial press has had time to write the morning's stories.

All of that information is theoretically embedded in tomorrow morning's opening price — but only theoretically. In practice, the first hour of NSE trading is messy. The goal of the forecaster is to make a calmer, evidence-based call before the open, so you're not reacting to the noise of the first 15 minutes.

What you'd see each morning

A JSON file (and eventually a clean web view) that looks roughly like this:

{
  "for_session": "2026-05-19",
  "predictions": [
    {
      "rank": 1,
      "ticker": "RELIANCE",
      "predicted_move_pct": 2.84,
      "confidence": 0.74,
      "expected_range": "+1.4% to +3.9%",
      "top_drivers": [
        "Brent crude up 3.1% overnight",
        "Refinery margin upgrade in the morning press",
        "Foreign-investor 5-day inflow trending positive"
      ],
      "hint": "BUY"
    }
  ]
}

The pieces:

• predicted_move_pct — the model's best single-number guess for today's close vs yesterday's close.
• confidence — calibrated, not vibes. A 0.74 confidence means: on past days where the model said 0.74, it was right 74% of the time. Boring, verifiable.
• expected_range — the 10th-to-90th percentile of where the close will probably land. Wider range = less certain.
• top_drivers — the three biggest reasons the model made that call. So you can sanity-check it. If the top driver is "yesterday's price moved a lot" and nothing else, you'll know to be skeptical.
• hint — a one-word summary (BUY / SELL / HOLD), but it is a hint, not an order. The user always decides.

How accurate does it have to be?

To be useful, the model needs two things:

1. Direction hit rate above 52% on the days when it's most confident. (Coin-flip is 50%. Anything over 55% is the goal.)
2. Honest confidence intervals. When the model says "80% sure this stock will close between +1% and +3%", the actual close should fall in that range 80% of the time. Not 95% (over-confident), not 60% (under-confident). Exactly 80%.

Both are measured continuously on held-out data the model has never seen, and any drop is visible on the live page.

What signals does it use?

Eight different "information layers" feed the model:

1. Global markets overnight — US close, Asia open, currency, oil, gold, volatility index.
2. Cross-stock signals — when the sector leader moves, the rest of the sector tends to follow. So do related companies (e.g., Tata Motors and Tata Steel).
3. Indian financial news — sentiment-scored articles from ET, Mint, Moneycontrol, Business Standard, LiveMint, lagged so it only uses what was actually public.
4. Big-money flows — foreign-investor and domestic-institution buying/selling, futures open interest, block deals.
5. Calendar events — earnings dates, board meetings, dividends, RBI policy days, regulatory filings.

Three more layers (alternative data, earnings-call NLP, social sentiment) are on the roadmap but deferred — they're heavier work and the team wants to see the first five layers prove out first.

What's the catch?

A few honest ones:

• It's research-grade software. No app yet, no broker integration, no SMS alerts. You'd read a JSON file (or a web page generated from it).
• A 12-month diary on the research side ended with "no measurable edge." The forecaster is the redesign informed by that result. It may also find no edge — that's a real possibility, and you'll be the first to know if it does.
• No real money is in flight. Paper-trading first. One month minimum. Real money only after the hit rate stays above 52% for a full month.

When can I use it?

OpenBracket is building piece by piece — each information layer ships and gets evaluated before the next is added. Realistic landing for v1 paper-trading is mid-to-late 2026.

If you want to follow the build, the orchestrator state page updates every 10 minutes from the live database. Or read the journey to understand the choices that led here.