Quant Signals as Probabilistic Evidence — Never Advice

NataPulse runs a quantitative signal pipeline that reads market candles, produces probabilistic model output — forecasts, volatility estimates, anomaly scores — and then, deliberately, refuses to publish most of it. A signal reaches the public event stream only when a selective promotion policy finds corroborating evidence, and every event that does get through carries the same fixed disclaimer: probabilistic quantitative signal, not financial advice. This article explains how the pipeline works, what the promotion policy actually requires, and why the separation between model output and published evidence is the point of the design.

Model output is not evidence

The core design decision is architectural, not editorial: the quantitative system is separated from the public event stream. A time-series model, given a window of candles, will always produce an output — a bullish or bearish bias with some strength, an expected volatility, an anomaly score. That output is a hypothesis about recent price behavior, nothing more. Treating every model run as a publishable market event would flood users with directional noise dressed up as insight.

So NataPulse stores every generated signal internally, but publishes none of them by default. The stored signal and the public event are different objects with different standards: the first records what the model said; the second asserts that something evidentially notable happened. Crossing from one to the other requires the promotion policy described below. The public site never shows a directional price-forecast widget — only promoted anomaly and volatility events, each shipped with its disclaimer.

From candles to signals, with a quality gate

The input is intentionally narrow: OHLCV candles only — open, high, low, close, volume. Crypto candles come from exchange APIs; equities, FX and commodities from market-data providers. The inference engine is Kronos, an open-source time-series model that NataPulse runs on its own infrastructure, so the raw forecasts never depend on a third-party inference service.

Before any inference runs, each candle window gets a quality snapshot: ready, degraded, stale, incomplete or invalid. Windows that fail the gate suppress generation or mark the resulting signal as degraded, and degraded signals do not pass into reports or, in the normal case, promotion. The principle is stated plainly in the product documentation: a strong-looking signal computed on poor input is not strong evidence.

Promotion: one strong trigger, or two weak ones

The promotion policy is where the philosophy becomes enforceable code, and it is auditable by construction — every decision records which triggers fired and why. In qualitative terms:

Strong triggers include a high anomaly score, a substantial volatility spike, and directional agreement across multiple timeframes.
Weak triggers include an elevated (but not high) anomaly, a moderate volatility change, a strong non-neutral forecast, and confirming volume from the market itself.

A signal is promoted to a public market_quant event only when at least one strong trigger fires, or at least two weak triggers corroborate each other. The rule with the most consequence is the one about forecasts: a directional forecast is only ever a weak trigger, so a bare model prediction — however confident it looks — can never promote an event on its own. It needs independent support, such as a volatility spike or volume confirmation, before users ever see it.

A promoted event can attach to an existing cluster for the same asset as one piece of quantitative evidence; a single quant signal never spawns a persistent narrative by itself.

What users see on Market Desk

On Market Desk, quant signals appear as structured evidence rather than calls to action: signal type (forecast, volatility, anomaly), timeframe, strength, confidence, data-quality status and generation time — each dimension kept distinct, because strength describes the model output, confidence describes support for it, and data quality describes the input. Promoted events also surface on Live Pulse and in Event Explorer, and generated reports can include a “Quant Read” section built from the same gated signals. The read-only quant API returns a disclaimer field in every response envelope.

The documented workflow for interpreting a signal starts with data quality, not direction — and ends with the reminder that promotion means the policy found sufficient evidence, not that the event is confirmed true or profitable.

The same discipline at index scale: 503 constituents, every close

The quant pipeline is not the only place this philosophy applies. Since June 2026 a companion pipeline ingests daily end-of-day candles for all 503 S&P 500 constituents, fetched after the U.S. close via Polygon’s official grouped-daily API in a single request per day. It feeds a separate market-anomaly stream rather than the quant model itself, but the publishing rule is identical: events are emitted only on anomalies — gaps of 3% or more, volume at 2.5 times the 20-day average, daily moves of 7% or more, or new 252-day highs and lows — and those anomaly events are what downstream clustering and the Emerging Narratives engine consume each day.

The pattern is the same at every layer: compute broadly, publish selectively, and let the disclaimer travel with the data. NataPulse produces research evidence. What to do with it is not a question the pipeline answers.