How our predictions are computed

Design philosophy

We deliberately run a rule-based, fully auditable model rather than an opaque machine-learning one. Every score can be decomposed into the exact signals that produced it, each shown on the asset page with its weight and direction. When a model explains itself, you can disagree with it intelligently — which is the whole point of research tooling.

Inputs

• Price history — up to five years of daily NSE candles per stock; NAV history for mutual funds. A minimum of ~60 sessions of history is required before a prediction is generated at all.
• Return statistics — trailing 1-day, 1-week, 1-month, 6-month and 1-year returns, volatility, and the 12-1 momentum factor (12-month return excluding the most recent month).
• Technicals — moving-average placement, trend slopes and drawdown measures.
• Fundamentals (stocks) — revenue/profit/EPS growth, P/E, P/B, ROE, RoCE, debt/equity, operating and profit margins, dividend yield, market cap.
• Fund quality (mutual funds) — scheme category, AUM band and NAV consistency from the AMFI registry mirror.
• News sentiment — recent headlines per asset, scored for polarity and recency.

From signals to score

The model starts from a neutral 50 and adds or subtracts weighted points from each signal family, then clamps the result to 0–100. Examples of the rules:

• Momentum — a strong trailing-horizon return (e.g. +15% over the horizon window) adds up to +8 points; weak momentum subtracts up to 7. Momentum's weight tapers for ultra-short horizons, where it is mostly noise.
• Trend & technicals — price above long moving averages and constructive 12-1 momentum add smaller increments; drawdowns beyond typical ranges subtract.
• Valuation & quality — reasonable P/E against growth, healthy ROE/RoCE and low leverage add; expensive or deteriorating profiles subtract. Fundamentals only move the score when the data is actually present — missing fields never count for or against.
• News — a run of positive headlines nudges the score up, negative ones down, with recent items weighted more.

The blended score is labelled from Very Weak to Very Strong and displayed with the reasons that produced it, highest weight first.

Scenario ranges

For each horizon we estimate three outcomes — bear, base and bull. The base case annualises the asset's own historical return profile with mean-reversion toward market norms; the bear and bull cases widen the band by that asset's realised volatility, so a shaky small-cap gets a much wider range than a stable large-cap. Ranges are estimates, not targets — see the guide to reading predictions.

Refresh cadence

Candles are ingested after each NSE session completes; predictions for all covered assets and horizons regenerate daily before the 09:15 IST open (the UI promises "refreshed by 9 AM"). News sentiment refreshes each morning. Each prediction carries a "data as of" timestamp so you always know what it is based on.

Known limitations

• Historical statistics say nothing about tomorrow's news, management fraud, regulatory shocks or macro events — the categories that most often break trend-based reasoning.
• Thinly traded scrips with short or gappy histories get wide, low-confidence ranges or no prediction at all.
• Fundamental data depends on public aggregators and can lag filings by weeks.
• The model is calibrated qualitatively, not backtested against live trading outcomes; treat scores as a research starting point, never as edge.

Questions about the model? Write to [email protected] — substantive questions get answered and often improve these pages.