How signals are generated and scored

How signals work

Every tick from the NSE feed flows through three independent layers in our pipeline:

1. Detection — has something statistically unusual just happened?
2. Regime classification — what kind of market are we in right now?
3. Scoring — given the detection and the regime, how confident should we be?

Only signals that pass all three layers reach your dashboard. Everything below that bar is logged for research but never surfaced.

CUSUM detection

We use a cumulative sum (CUSUM) detector to identify shifts in the underlying process. CUSUM is well-suited to financial time series because it looks for a meaningful change in the rolling mean rather than fitting a fixed threshold to noisy data.

In practice, this means CUSUM fires when the recent behaviour of an instrument has departed non-trivially from its preceding window — much faster than a moving-average crossover, and with fewer false positives than naive thresholding. We tune detector parameters per underlying because Nifty and BankNifty have different microstructure.

Regime classification

We classify the current market into one of three regimes:

• Compression — implied volatility and realised range are both low. Coiled-spring conditions.
• Expansion — volatility is rising and price is moving meaningfully. Directional opportunities.
• Trending — sustained directional movement with relatively stable volatility.

The same CUSUM detection is interpreted differently in each regime. A compression-regime breakout carries a very different probability profile than the same move during an established trend.

The signal score

The score (0.0–10.0) summarises how confident the model is. The composition, roughly:

• Trigger strength — magnitude of the CUSUM departure relative to expected noise.
• Regime fit — does the historical base rate of this trigger × this regime support a directional read?
• Amplifiers — supporting evidence (volume, OI shift, options skew).
• Conflicts — contradicting evidence that pulls the score down.

We don't publish exact thresholds because they evolve with model retraining. Use the score comparatively (a 7.4 is meaningfully stronger than a 5.1) rather than as an absolute commitment.

Auto-trade vs manual review

Two modes, configured in your risk profile:

• Manual review (default). Signals appear in your dashboard. Nothing is placed automatically. Useful for your first week — watch how signals behave before delegating execution.
• Auto-trade. Signals that pass your risk profile become live orders at your broker. You must acknowledge the SEBI auto-trade risk disclosure before enabling this for the first time.

Auto-trade respects every limit in your risk profile. The number of positions, the per-position size, the daily loss cap, and the allowed-underlyings list all gate execution server-side.

Reading the live feed

The live signal feed shows, per row:

• Direction — BUY (long) or SELL (short / sell-to-open) on the underlying.
• Instrument — the option contract or future the signal targets.
• Score — the model's confidence as described above.
• Status — whether the signal was acted on, rejected (and why), or is still queued for review.
• Time — exchange timestamp in IST.

Click any row to drill into the trigger detail — which detector fired, the regime at the time, amplifiers, and conflicts. Everything is research-traceable.