Documentation
How Gamma Score works
Every number on the dashboard maps to a definition here. Start with Glossary for the plain-English version. Drill into Detailed Formulas, ML, or Backtest for the math and methodology.
Read this first.
Every number on the dashboard means something specific. This page explains every term in plain English, so you never see a metric you can\'t interpret. The Detailed Formulas tab has the math; this tab has the intuition.
The three numbers that matter
Score
−100 to +100
Direction and strength. Positive = up. Negative = down. Magnitude = how strong the edge is. Same scale across every stock and every window, so comparisons are fair.
Conviction
0% to 100%
How confident the model is. 90% conviction at +30 beats 30% conviction at +90. Always read score and conviction together.
Target return
Predicted %
The ML model\'s point estimate of return over the horizon (1d / 5d / 20d). On Today\'s Picks this is the headline number; on Markets Live it powers the rankings.
How a pick is generated — start to finish
1
Capture the raw data
Two Kite WebSocket connections stream ticks all session. We persist option-chain snapshots, spot prices, and futures snapshots into Postgres. NSE bhavcopy backfills historical EOD bars going back 2+ years.
2
Compute features
At every window-close (30m, 1h, full-day) and at 4:00 PM EOD, we compute ~40 features per (stock × window): technical indicators, cross-section ranks, option-chain sentiment, regime context. Stored in
features_eod.3
Run 15 rule-based sub-scores
Each sub-score tries to capture one specific pattern: opening-range breakouts, volume anomalies, IV skew direction, mean reversion, etc. Each returns a value in [-1, +1] with a conviction.
4
Combine into Composite Gamma
Conviction-weighted average of the 15 sub-scores. This is the rule-based headline number you see on Markets Live and the Signals page.
5
Run the ML model in parallel
The same feature set (plus regime + option-chain features) is fed to a LightGBM predictor trained weekly on the laptop. It outputs a predicted return % for 1d, 5d, and 20d horizons. Stored in
predictions_eod.6
Today\'s Picks pulls from both
The Picks page ranks by ML prediction (primary) and shows the rule-based score as a second opinion. When they agree, conviction is higher.
Where the signals come from
Price / volume
Rolling indicators like RSI, MACD, Bollinger bands, range %, body %.
Volume profile
Up-vol vs down-vol split, volume z-score vs 20-day baseline.
Cross-section
Where this stock ranks vs the F&O universe today, vs its sector, vs NIFTY.
Option chain
PCR (oi + volume), IV skew, max pain distance, unusual OI z-score.
Greeks
Delta, gamma, vega, theta computed via Black-Scholes from live option prices.
Multi-timeframe
Aligned momentum across 5-day and 20-day horizons.
Market regime
NIFTY trend, breadth, dispersion, NIFTY vol — context for "good signal in chaotic market".
Intraday momentum
Late-day return, opening-range volatility, overnight gap, where in the day's range the close landed.
Dictionary of terms
Gamma ScoreA single number from -100 to +100 saying how bullish or bearish we think the next move is.
Gamma Score
A single number from -100 to +100 saying how bullish or bearish we think the next move is.
Positive means we expect the price to go up over the prediction horizon. Negative means down. Zero means no edge. The bigger the absolute number, the stronger the signal. Made up of 15 sub-signals combined.
ConvictionHow sure we are about the score, expressed 0-100%.
Conviction
How sure we are about the score, expressed 0-100%.
A score of +80 with 30% conviction is a guess. A score of +20 with 90% conviction is a high-confidence call. Always look at score AND conviction together — never one alone.
WindowA specific time slice of the trading day used to compute features.
Window
A specific time slice of the trading day used to compute features.
Examples: "open_30m" = first 30 minutes after 9:15. "open_1h" = first hour. "full_day" = full session. Different windows expose different patterns — an opening-range breakout in the first 30 min is different from a late-day momentum push.
Conviction-weightedWhen we combine multiple scores, low-conviction ones count less.
Conviction-weighted
When we combine multiple scores, low-conviction ones count less.
Twenty signals saying "+50 at 20% conviction" carry less weight than one signal saying "+50 at 90% conviction". This keeps weak signals from drowning out strong ones.
Cross-sectionalA measurement that compares this stock to all other F&O stocks today.
Cross-sectional
A measurement that compares this stock to all other F&O stocks today.
If RELIANCE is up 1% but the median F&O stock is up 0.3%, RELIANCE is in the top quartile cross-sectionally. Tells you "is this move special or is the whole market moving?"
Composite GammaThe headline rule-based score combining all 15 sub-signals.
Composite Gamma
The headline rule-based score combining all 15 sub-signals.
For each underlying we compute 15 separate scores (each tries to capture one specific pattern). Composite Gamma blends them with weights chosen by their conviction. This is the rule-based score; the ML model is a separate, often more accurate signal.
ML PredictionOur LightGBM model's estimate of next-day, next-5-day, or next-20-day return %.
ML Prediction
Our LightGBM model's estimate of next-day, next-5-day, or next-20-day return %.
Trained on 2+ years of historical data covering ~210 F&O stocks. Sees the same features the rule-based scores see, plus option-chain data and market-regime context. Retrained weekly from fresh data.
IC (Information Coefficient)A 0-1 score for how well our predictions rank correctly.
IC (Information Coefficient)
A 0-1 score for how well our predictions rank correctly.
IC = correlation between predicted return and actual return across all stocks on a day. 0.05+ is considered tradeable for equities. We target 0.05-0.10 on the 5-day horizon and 0.10+ on the 20-day horizon. (An earlier 1-day horizon was retired — it underperformed a simple rule-based baseline.)
Walk-forward backtestA backtest that re-trains the model every step forward in time — no peeking at the future.
Walk-forward backtest
A backtest that re-trains the model every step forward in time — no peeking at the future.
Train on data up to day X. Predict day X+1. Then add day X+1 to the training data and predict day X+2. Repeat. This simulates exactly what would have happened if you ran the model live, with zero look-ahead bias.
Look-ahead biasThe classic backtesting mistake of using information you wouldn't have had at the time.
Look-ahead bias
The classic backtesting mistake of using information you wouldn't have had at the time.
Example: training on data through December 2024 and then "predicting" October 2024. The model has seen the future. Our walk-forward setup makes this impossible — predictions for every test day are made by a model that only saw data BEFORE that day.
Sharpe ratioReturn per unit of risk. Higher is better. Above 1 is good, above 2 is rare.
Sharpe ratio
Return per unit of risk. Higher is better. Above 1 is good, above 2 is rare.
Mathematically: (strategy return - risk-free rate) / strategy volatility, annualised. Captures whether high returns come from skill or just from taking more risk.
DrawdownThe biggest peak-to-trough fall in the strategy's equity curve.
Drawdown
The biggest peak-to-trough fall in the strategy's equity curve.
If the strategy goes 100 → 150 → 110 → 140, the drawdown is 150 → 110 = 27%. Tells you the worst pain you'd have felt holding the strategy. Lower is better.
Hit rate / win rateWhat fraction of trades made money.
Hit rate / win rate
What fraction of trades made money.
A 60% hit rate means 6 of 10 trades were profitable. By itself this doesn't mean much (you can have a 90% hit rate and still lose money if losers are big). Read it alongside average winner vs average loser.
IV (Implied Volatility)The volatility level baked into an option's price.
IV (Implied Volatility)
The volatility level baked into an option's price.
Solved from the Black-Scholes formula given the option's observed market price. High IV = market expects big moves. Low IV = market expects quiet.
PCR (Put-Call Ratio)Total put open interest divided by total call open interest.
PCR (Put-Call Ratio)
Total put open interest divided by total call open interest.
High PCR (>1.2 on Indian markets) often = bearish positioning (a lot of people own puts). Low PCR (<0.7) = bullish positioning. Useful sentiment gauge but contrarian signals matter at extremes.
Max PainThe strike price at which most option holders lose money at expiry.
Max Pain
The strike price at which most option holders lose money at expiry.
Markets tend to "pin" toward max pain as expiry approaches (we see this empirically every Thursday on NIFTY weekly). Distance from spot to max pain is a small but real near-expiry signal.
OI buildupHow open interest is changing alongside price action.
OI buildup
How open interest is changing alongside price action.
Price up + OI up = long buildup (bullish). Price up + OI down = short covering. Price down + OI up = short buildup (bearish). Price down + OI down = long unwinding. Tells you what kind of participant is driving the move.
Paper tradeA simulated trade with fake money on real market data.
Paper trade
A simulated trade with fake money on real market data.
You learn the strategy and feel the wins/losses without risking real capital. We start every user with paper money so they can validate the picks before subscribing.
Realised P&LProfit/loss locked in by closing positions. Excludes positions you still hold.
Realised P&L
Profit/loss locked in by closing positions. Excludes positions you still hold.
The incentive program uses realised P&L only — you have to actually close a winning position to count, not just hold an unrealised gain that could vanish tomorrow.
What we don\'t claim
✗
We don\'t promise returns.
✓ We measure edge honestly via walk-forward backtests. Sharpe, drawdown, hit rate are all on the Backtest page.
✗
We don\'t do tips or stock recommendations.
✓ We\'re not SEBI-registered as an investment advisor. We surface patterns; you decide whether to act.
✗
We don\'t hide bad days.
✓ Every published backtest is no-look-ahead and shows the worst drawdown alongside the best return.
✗
We don\'t use your trades to inform anyone else\'s picks.
✓ The platform is for analysis, not for routing flow. Your paper trades are yours alone.
GammaScore · F&O analytics + paper trading · not a SEBI-registered advisory.
Past performance is not indicative of future results.