Skip to content

SIGNAL_STRATEGY.md

Latest commit

 

History

History
542 lines (404 loc) · 14.6 KB

SIGNAL_STRATEGY.md

File metadata and controls

542 lines (404 loc) · 14.6 KB
Raw

brandOS Signal Strategy

Strategic positioning and phased implementation for brand signal intelligence.

Executive Summary

The Gap: Paid platforms (Brandwatch, Meltwater, AlphaSense) excel at collecting and displaying signals. They stop at "here's what happened" — leaving users to manually interpret, decide, and act.

Our Position: brandOS goes further — from signal detection through to proposed actions, human approval, execution, and outcome tracking. We own the interpretation-to-action layer.

One-liner: "Brandwatch shows you what happened. brandOS tells you what to do about it."

Positioning

Market Map

INTERPRETATION DEPTH
      ^
      |
 High |    [brandOS]  ← Agent synthesis + action proposals
      |
  Mid |         [Brandwatch, Meltwater, Sprout]  ← Dashboards
      |
  Low |    [GDELT, NewsAPI, Finnhub]  ← Raw feeds
      |
      +-------------------------------------------------> Cost
           Free       $100s/mo     $1000s/mo    Enterprise

What We Do

Capability Differentiation
Multi-source fusion News + filings + jobs + trends in one schema
Weak signal detection Surface anomalies before they trend
Agent synthesis LLM agents produce analysis, not charts
Action proposals "Based on X, we recommend Y" with rationale
Human approval gates High-stakes decisions require sign-off
Decision audit trail What was proposed, approved, executed, outcomes
Outcome learning Results improve future recommendations

What We Don't Do

Skip Reason
Historical archives Incumbents' moat; we focus on real-time + forward
Dashboard UI CLI-first, API-second; UI is distraction
Enterprise compliance SOC2/GDPR expensive; target SMB first
Social publishing at scale Hootsuite owns this; we integrate
Influencer databases Different product
Ad campaign management Out of scope

Signal Taxonomy

Sources

SIGNAL SOURCES
├── MARKET (what the market does)
│   ├── Search trends ─────── Google Trends, pytrends
│   ├── Stock movements ───── yfinance, Finnhub
│   ├── Economic indicators ─ FRED
│   └── Consumer spending ─── (future: alt data)
│
├── COMPETITIVE (what competitors do)
│   ├── SEC filings ───────── edgartools (8-K, 10-K, 10-Q)
│   ├── Patent filings ────── USPTO PatentsView
│   ├── Job postings ──────── LinkedIn, Greenhouse boards
│   ├── Press releases ────── news APIs
│   └── Product changes ───── pricing, features tracking
│
├── REPUTATION (what people say)
│   ├── News mentions ─────── GDELT, Google News RSS
│   ├── Social mentions ───── Reddit (PRAW), Twitter
│   ├── Reviews ───────────── G2, Capterra, app stores
│   └── Forums ────────────── technical communities
│
└── WEAK SIGNALS (early indicators)
    ├── Anomalous mention velocity
    ├── Sentiment shift patterns
    ├── Cross-source correlation spikes
    └── Emerging entity co-occurrence

Value Matrix

Signal Type Detection Actionability Competition
News mentions Easy Medium Crowded
Social sentiment Easy Medium Crowded
SEC filings Medium High Moderate
Patent filings Medium High Low
Job postings Medium High Low
Search trends Easy Medium Moderate
Weak signals Hard Very High Empty

Focus areas: SEC filings, patent filings, job postings, weak signal detection.

Implementation Phases

Phase 0: Foundation (Weeks 1-2)

Goal: Solidify signal schema and storage.

Schema Additions

Extend existing Signal model:

class Signal(BaseModel):
    # Existing fields from ROADMAP.md...

    # Weak signal detection
    velocity: float | None         # mentions/hour delta
    novelty_score: float | None    # how "new" is this
    correlation_ids: list[str]     # related signal IDs

    # Outcome tracking
    action_taken: str | None
    outcome: dict | None

Storage

  • SQLite for MVP
  • Alembic migrations
  • 90-day hot retention, archive to Parquet

CLI Baseline

brandos signals sources list
brandos signals fetch --brand acme
brandos signals show <id>
brandos signals stats --brand acme

Deliverable: Signal storage + existing Google News provider working.

Phase 1: Source Expansion (Weeks 3-5)

Goal: Add high-value, low-competition sources.

1.1 SEC EDGAR

# signals/providers/sec_edgar.py
from edgartools import Company

def fetch_competitor_filings(ticker: str, form_types: list[str], since_days: int) -> list[Signal]:
    company = Company(ticker)
    filings = company.get_filings(form=form_types)
    # Extract key sections, summarize
    ...

Extract:

  • Risk factor changes (10-K section 1A diffs)
  • Material events (8-K items)
  • Revenue segment changes
  • Executive departures

1.2 GDELT

# signals/providers/gdelt.py
import gdelt

gd = gdelt.gdelt(version=2)

def fetch_gdelt_mentions(query: str, timespan: str = "1d") -> list[Signal]:
    results = gd.Search(query, table='gkg', coverage=True)
    ...

Extract:

  • Global mention volume
  • Tone/sentiment by region
  • Source diversity
  • Theme taxonomy

1.3 Job Postings

Track competitor hiring patterns:

  • Role distribution (engineering vs sales)
  • Seniority levels
  • Location expansion
  • Tech stack mentions

Deliverable: 3 new providers feeding unified schema.

Phase 2: Enrichment (Weeks 6-8)

Goal: Transform raw signals into scored intelligence.

2.1 Domain-Specific Sentiment

# signals/enrichment/sentiment.py
from transformers import AutoModelForSequenceClassification

class FinancialSentiment:
    """FinBERT for financial text, VADER for social."""

    def __init__(self):
        self.model = AutoModelForSequenceClassification.from_pretrained("ProsusAI/finbert")

    def score(self, text: str, source: str) -> float:
        if source in ["sec_edgar", "finnhub", "news"]:
            return self._finbert_score(text)
        return self._vader_score(text)

2.2 Entity Extraction

# signals/enrichment/entities.py
import spacy

nlp = spacy.load("en_core_web_sm")

def extract_entities(text: str) -> dict:
    doc = nlp(text)
    return {
        "organizations": [e.text for e in doc.ents if e.label_ == "ORG"],
        "people": [e.text for e in doc.ents if e.label_ == "PERSON"],
        "products": [...]  # Custom NER
    }

2.3 Velocity Tracking

# signals/enrichment/velocity.py
def calculate_velocity(brand: str, window_hours: int = 24) -> dict:
    recent = get_signals(brand, since=f"{window_hours}h")
    baseline = get_average_rate(brand, days=30)

    return {
        "current_rate": len(recent) / window_hours,
        "baseline_rate": baseline,
        "velocity_ratio": len(recent) / window_hours / baseline,
        "is_anomalous": (len(recent) / window_hours) > baseline * 2
    }

Deliverable: All signals have sentiment, entities, velocity scores.

Phase 3: Weak Signal Detection (Weeks 9-12)

Goal: The differentiator. Surface signals before they're obvious.

3.1 Anomaly Detection

Statistical first (no training):

# signals/detection/anomaly.py
from scipy import stats

def detect_anomalies(signals: list[Signal], threshold: float = 2.5) -> list[Signal]:
    velocities = [s.velocity for s in signals]
    z_scores = stats.zscore(velocities)
    return [s for s, z in zip(signals, z_scores) if abs(z) > threshold]

Then Isolation Forest for multivariate:

from pyod.models.iforest import IForest

def detect_multivariate(features: np.ndarray) -> np.ndarray:
    clf = IForest(contamination=0.05)
    clf.fit(features)
    return clf.predict(features)

3.2 Changepoint Detection

# signals/detection/changepoint.py
import ruptures as rpt

def detect_changepoints(series: np.ndarray, min_size: int = 5) -> list[int]:
    algo = rpt.Pelt(model="rbf", min_size=min_size)
    algo.fit(series)
    return algo.predict(pen=10)

3.3 Cross-Source Correlation

The key insight: when multiple independent sources mention the same thing.

# signals/detection/correlation.py
def detect_emergence(entity: str, window_hours: int = 48) -> dict:
    signals = get_signals_mentioning(entity, since=f"{window_hours}h")
    sources = set(s.source for s in signals)

    # 3+ independent sources = emerging signal
    return {
        "entity": entity,
        "sources": list(sources),
        "is_emerging": len(sources) >= 3,
        "confidence": min(1.0, len(sources) / 5)
    }

Deliverable: Automated weak signal alerts with confidence scores.

Phase 4: Agent Synthesis (Weeks 13-16)

Goal: Connect detection to agent architecture from AGENTS.md.

4.1 Signal Digest Agent

# agents/digest.py
from pydantic_ai import Agent

digest_agent = Agent(
    'anthropic:claude-sonnet-4-20250514',
    system_prompt="""Brand intelligence analyst. Given signals, produce:
    1. Executive summary (3 bullets)
    2. Key threats
    3. Opportunities
    4. Recommended actions with rationale
    Cite signal IDs.""",
    result_type=SignalDigest
)

4.2 Threat Assessor

Extends existing ThreatAssessor from AGENTS.md with signal detection:

# agents/threat.py
threat_agent = Agent(
    'anthropic:claude-sonnet-4-20250514',
    system_prompt="""Assess competitive threats. Focus on:
    - Competitor launches
    - Pricing changes
    - Executive moves
    - Patent filings
    - Hiring patterns

    Rate: LOW / MEDIUM / HIGH / CRITICAL
    Cite evidence.""",
    result_type=ThreatAssessment
)

4.3 Action Proposer

The differentiator — from analysis to recommendation:

# agents/action.py
action_agent = Agent(
    'anthropic:claude-sonnet-4-20250514',
    system_prompt="""Propose specific actions from signal analysis.
    Actions must be:
    - Concrete (not "monitor situation")
    - Time-bound
    - Assigned to function (content, product, PR, legal)

    Format each:
    - Action: [specific]
    - Rationale: [citing signals]
    - Urgency: LOW/MEDIUM/HIGH
    - Owner: [function]""",
    result_type=ActionProposal
)

Deliverable: Agent pipeline producing actionable recommendations.

Phase 5: Decision Loop (Weeks 17-20)

Goal: Close the loop with approval and learning.

5.1 Rejection Learning

Extend existing approval workflow:

# workflows/approval.py
def on_enter_rejected(self, reason: str):
    log_decision(action=self.action, status="rejected", reason=reason)

    # Improve future proposals
    update_action_model(
        action_type=self.action.type,
        was_approved=False,
        reason=reason
    )

5.2 Outcome Tracking

# core/outcome.py
class ActionOutcome(BaseModel):
    action_id: str
    executed_at: datetime
    metrics_before: dict
    metrics_after: dict
    success_score: float
    learnings: str

5.3 Feedback Loop

Use outcomes to improve relevance:

def update_relevance_weights(outcomes: list[ActionOutcome]):
    for outcome in outcomes:
        signals = get_signals_for_action(outcome.action_id)
        delta = 0.1 if outcome.success_score > 0.5 else -0.1
        for signal in signals:
            update_source_weight(signal.source, delta)

Deliverable: Full decision loop with learning.

Success Metrics

Phase 0-1

  • 5+ signal sources integrated
  • 1000+ signals/day capacity
  • < 5 min latency source → storage

Phase 2-3

  • Sentiment accuracy > 80%
  • Anomaly precision > 70%
  • Weak signals surfaced 24-48h before mainstream

Phase 4-5

  • Recommendation acceptance > 60%
  • Signal → action latency < 1 hour
  • Outcome tracking coverage > 90%
  • Relevance scores improve over time

Resources

Open-Source Tools

Tool Purpose Link
GDELT Global news, 15-min updates gdeltproject.org
edgartools SEC filing analysis github.com/dgunning/edgartools
FinBERT Financial sentiment github.com/ProsusAI/finBERT
FinGPT Financial LLM github.com/AI4Finance-Foundation/FinGPT
pyod Anomaly detection github.com/yzhao062/pyod
ruptures Changepoint detection github.com/deepcharles/ruptures
spaCy NLP/NER spacy.io
PRAW Reddit API praw.readthedocs.io

Research

Topic Source
Weak signal mining Springer: Systematic Literature Review
Organizational sensemaking ScienceDirect: Digital Strategic Agility
Brand health metrics Ehrenberg-Bass Institute
Multi-source aggregation arXiv: AIMM-X

Paid Landscape (for reference)

Platform Focus Moat
Brandwatch Social listening Historical archive
Meltwater Media monitoring Publisher relationships
AlphaSense Market intelligence Expert transcripts
RavenPack Financial NLP Hedge fund integrations

CLI Reference

# Sources
brandos signals sources list
brandos signals sources add sec-edgar --tickers AAPL,GOOGL,MSFT
brandos signals sources add gdelt --query "brand name"
brandos signals sources test <name>

# Fetching
brandos signals fetch --brand acme
brandos signals fetch --brand acme --sources sec-edgar,gdelt

# Analysis
brandos signals stats --brand acme --period 7d
brandos signals anomalies --brand acme
brandos signals emerging --brand acme

# Agent integration
brandos agent run signal-digest --brand acme
brandos agent run threat-assessor --brand acme --signals latest

Summary

Layer Incumbents brandOS
L1: Raw data Free APIs Same (GDELT, SEC, etc.)
L2: Normalized Their moat We build this
L3: Interpreted Dashboards Agents do better
L4: Actionable Gap We own this
L5: Outcomes Gap We own this

The competitive landscape focuses on L1-L3. We differentiate at L4-L5.