Signal Composition Deep Dive: How AIVEX Merges Atomic Signals

The Composition Problem

Running three independent signal modules in parallel is the easy part. The hard part is combining their outputs meaningfully.

Three modules analyzing the same symbol at the same time can easily disagree. News sentiment might be negative. Chart technicals might show a breakout. Fundamental metrics might be neutral. What does the system do?

AIVEX solves this with a structured composition pipeline inside the Signal Engine (Brain module).

Step 1: Confidence Normalization

Each atomic module produces a signal with a raw confidence score. Before composition, all scores are normalized to [0, 1] using per-module calibration curves derived from historical accuracy data.

A News Signal with 0.8 raw confidence is not the same as a Chart Signal with 0.8 raw confidence — they have different base rates and noise characteristics. Normalization makes them comparable.

Step 2: Conflict Detection

Once normalized, the Engine checks for directional conflicts: cases where two or more signals disagree on the direction of the output (bullish vs. bearish).

Conflicting signals are not simply averaged. Instead, the Engine applies a conflict penalty that reduces the composite confidence. Two signals with equal confidence but opposite direction produce a composite confidence near 0 — correctly representing uncertainty.

The formula is:

composite_confidence = weighted_avg(confidences) × (1 - conflict_penalty)
conflict_penalty = 2 × |proportion_bullish - 0.5|

Step 3: Priority Resolution

When a conflict cannot be resolved by confidence weighting alone (e.g., two signals with identical confidence and opposite direction), the Engine applies a configurable priority order:

News Signal (fastest to react to events)
Chart Signal (reflects market participant behavior)
Metrics Signal (slowest-moving, highest signal-to-noise)

This priority order is not hardcoded — it is configurable per deployment.

Step 4: Cooldown Enforcement

After a composite signal is emitted for a symbol, the Engine enforces a cooldown period. During cooldown, new signals for the same symbol are suppressed unless:

The new composite confidence exceeds the previous by more than a configurable margin (default: 0.15)
A significant new event is detected (News Signal urgency score > 0.8)

Cooldown periods are adaptive: they shorten during high-volatility periods and lengthen during quiet markets.

Step 5: Output Packaging

The final composed signal is packaged with:

Symbol and timestamp
Composite confidence score
Contributing modules and their individual confidences
Direction (bullish / bearish / neutral)
Cooldown expiry time

This package passes to the Governor for final gating before API emission.

Why This Matters

Signal composition is where most research platforms fail. They either blindly average signals (losing the information content of confidence) or use hardcoded rules that cannot adapt.

AIVEX's composition pipeline is transparent, configurable, and auditable. Every composition decision is logged, including which signals were considered, which were suppressed by cooldown, and what the final confidence calculation produced.

For research and informational purposes only. Not financial advice.