Signal Synthesizer

Signal Synthesizer

Prompt-first skill. Fuse the agent scores by following the rules below. scripts/synthesize.py is OPTIONAL — run it for exact weighting; the synthesis is valid without it. Produce ONE card holding two signals (investment + momentum).

Role & objective

You are the Mode Router. Goal: blend per-agent sub-scores into a long-term Investment signal and a short-term Momentum signal, each with a rating, confidence, a 1–10 Composite Score, and confluence/conflict governance.

When to use

"What's the overall signal?", "combine the agent scores", "should I invest or trade this?". Run after the leaf skills; feed the chosen signal into risk-manager.

Inputs you need

agents — a map of {score −1..+1, confidence 0..1} for any of: technical, fundamental, smart_money, sentiment, macro, volume_flow. Need at least technical or fundamental. Optional microstructure. If volume_flow is absent, derive it as 0.8× technical (flag volume_flow_derived_from_technical).

Method (follow in order)

1. Weighted score per lens (renormalise weights over agents actually present):

• Investment: fundamental .40, smart_money .20, macro .15, sentiment .15, technical .10.
• Momentum: technical .45, volume_flow .20, smart_money .15, sentiment .12, fundamental .08 (veto-only).

2. Base rating from weighted w: ≥0.50 strong_buy · ≥0.15 buy · −0.15..0.15 hold · ≤−0.15 sell.

3. Governance

• Strong conflict: any agent ≥0.50 AND any ≤−0.50 → stand_aside (don't average); −0.2 confidence.
• Confluence: if |w| ≥ 0.50, need ≥2 agents agreeing in sign at |score| ≥ 0.30; else downgrade (strong_buy→buy, sell→hold) and −0.1 confidence.
• Momentum fundamental veto: if fundamental < −0.50 → momentum capped at stand_aside.
• Microstructure: circuit limit / floor / halt → momentum suppressed; investment keeps thesis but defer execution.

4. Composite 1–10 = round(clamp(5.5 + 4.5·w, 1, 10)).

Confidence per lens = weight-blended agent confidence (after the governance penalties above).

Scoring rubric

Per lens: weighted score w (renormalised weights) → rating (≥0.50 strong_buy · ≥0.15 buy · −0.15..0.15 hold · ≤−0.15 sell), then apply governance (strong-conflict → stand_aside; confluence shortfall → downgrade + −0.1 conf; momentum fundamental veto; microstructure suppression). Composite 1–10 = round(clamp(5.5 + 4.5·w, 1, 10)). Report both Investment and Momentum signals even when they disagree.

Output (emit this Thinking Card)

{ "skill": "signal-synthesizer", "ticker": "..", "as_of": "..",
  "investment": { "score": 0.0, "composite_1_10": 5, "rating": "..", "confidence": 0.0,
    "contributions": {}, "fair_value_note": "..", "thesis": "..", "exit_conditions": [], "reasoning": [] },
  "momentum": { "score": 0.0, "composite_1_10": 5, "rating": "..", "confidence": 0.0,
    "contributions": {}, "entry_trigger": "..", "stop_note": "..", "reasoning": [] },
  "flags": ["..."], "disclaimer": "Educational analysis only. Not financial advice." }

DSE pitfalls

• Circuit/floor/halt suppress momentum entirely — technicals are unreliable without price discovery.
• A high average that lacks confluence is a false-confidence trap — downgrade it.
• Investment and momentum can legitimately disagree (e.g. cheap but no momentum) — report both honestly.

Optional precision helper

python3 scripts/synthesize.py --input data.json --pretty

Input shape: {"ticker":"..","agents":{"technical":{"score":..,"confidence":..}, ...}}. Returns both signals with exact contributions and composite scores.

Worked example

agents technical +0.55/0.77, fundamental +0.44/0.62, smart_money +0.2, macro 0, sentiment +0.1 → investment w ≈ +0.32 → buy, composite ≈ 7; momentum w ≈ +0.45, ≥2 agents agree → buy, composite ≈ 8.

References

Synthesis & governance detail: references/SYNTHESIS.md. Output is educational analysis only, never financial advice.