Methodology

How Our 5 AI Engines Work

Full transparency into the algorithms, models, and techniques that transform 260,000+ consumer reviews into institutional-grade trading signals. No black boxes.

Neural Core
Anomaly Detection & Embeddings Engine

Neural Core is the foundation of ReviewSignal's intelligence stack. It transforms raw review text into 384-dimensional semantic embeddings using MiniLM, then applies Isolation Forest anomaly detection to identify statistically significant deviations in rating, sentiment, and review volume across all tracked locations.

MiniLM 384-dim Embeddings — all-MiniLM-L6-v2 captures semantic meaning far beyond keyword matching. Each review is encoded into a dense vector space where similar opinions cluster together.
Isolation Forest Anomaly Detection — Trained on 8,700+ real samples, the model identifies locations exhibiting statistically unusual rating drops, sentiment shifts, or volume spikes.
Welford's Online Statistics — Incremental algorithm tracks running means, variances, and z-scores per entity without reprocessing historical data. Scales to millions of data points.
Zero External API Cost — The entire inference pipeline runs locally on our infrastructure. No OpenAI, no cloud NLP services. Every embedding computed in-house at zero marginal cost.
384
Vector Dimensions
<1s
Inference Latency
$0/mo
API Cost
8,700+
Training Samples
Echo Engine
Sentiment Propagation & Monte Carlo Simulation

Echo Engine models how consumer sentiment cascades across geographic and brand networks. Using sparse matrix propagation and Monte Carlo simulation with 1,000+ paths, it generates confidence-weighted BUY/HOLD/SELL trading signals by analyzing how sentiment at one location predicts movements at related locations and competing brands.

Sparse Matrix Propagation — Sentiment signals propagate through a weighted adjacency matrix connecting locations by geography, brand affiliation, and competitive relationships.
1,000+ Monte Carlo Paths — Each trading signal is validated through thousands of simulated scenarios, producing probability-weighted confidence scores rather than point estimates.
Distance-Decay Weighting — Propagation strength decreases with geographic distance following an exponential decay function, ensuring local signals carry appropriate weight.
BUY/HOLD/SELL Signal Generation — Final output is a discrete trading recommendation with confidence interval, direction magnitude, and temporal horizon.
1,000+
Monte Carlo Paths
57,725
Locations in Matrix
0.78
Avg Confidence
3
Signal Types
Singularity Engine
7-Level Causal Analysis & Semantic Resonance

Singularity goes beyond correlation to establish causation. Its 7-level causal analysis framework traces sentiment shifts back to their root causes through temporal manifold analysis, semantic resonance mapping, and causal archaeology. It classifies events as STRUCTURAL (permanent operational change) or EPISODIC (temporary fluctuation), giving investors critical context for position sizing.

7-Level Causal Depth — From surface symptoms to root causes: event detection, pattern matching, temporal alignment, semantic clustering, causal chain reconstruction, structural classification, and impact projection.
Temporal Manifold Analysis — Maps sentiment trajectories in multi-dimensional time space, separating seasonal patterns from genuine structural shifts in brand perception.
Semantic Resonance — Detects when review language patterns across locations converge on similar themes, indicating systemic rather than isolated issues.
STRUCTURAL vs EPISODIC Classification — Automatically determines whether a sentiment shift represents a permanent operational change or a temporary disruption, directly informing position duration.
7
Causal Levels
2
Event Classes
90-day
Lookback Window
Enterprise
Tier Required
Cortex AI
Institutional Narrative Generation

Cortex AI transforms quantitative signals into institutional-grade written analysis. Powered by Claude AI with 14 section-specific prompts calibrated to a Goldman Sachs institutional voice, it generates research reports that read like they came from a top-tier equity research desk. Every output passes through a quality gate before delivery.

Claude AI Foundation — Built on Anthropic's Claude for nuanced, accurate, and factual narrative generation that avoids hallucination and maintains analytical rigor.
14 Section-Specific Prompts — Each report section (executive summary, sentiment analysis, risk factors, trading recommendation, etc.) uses a dedicated prompt template tuned for that analytical context.
Goldman Sachs Institutional Voice — Writing style calibrated to match the tone, structure, and vocabulary expected by institutional investors and portfolio managers.
Quality Gate — Every generated section passes automated quality checks for factual consistency, numerical accuracy, and analytical coherence before inclusion in final reports.
14
Prompt Templates
14-day
Response Cache
45-55pp
Enterprise Report
100%
Quality Gate Rate
Beacon Intelligence
Employee-Customer Sentiment Correlation

Beacon Intelligence correlates employee sentiment from Glassdoor with customer sentiment from Google Maps to detect early warning signals. When employee morale crashes, customer experience typically follows within 60-90 days. Beacon detects this divergence before it appears in traditional financial metrics, providing a unique leading indicator for hiring intent and operational health.

Glassdoor-Google Maps Correlation — Cross-references employee review sentiment with customer review sentiment across 238 brands to identify divergence patterns.
Employee vs Customer Divergence — Detects when internal sentiment decouples from external perception — a leading indicator of operational problems not yet visible to the market.
Hiring Intent Detection — Analyzes Glassdoor review patterns and job posting sentiment to identify companies in expansion, contraction, or restructuring phases.
SEC 13F Integration — Cross-references detected signals with institutional filing data to identify smart money positioning aligned with or contrary to sentiment indicators.
238
Brands Tracked
60-90d
Lead Time
2
Data Sources
Pro+
Tier Required
Get started

See our engines in action with a
free 14-day trial

Access Neural Core, Echo Engine, and Cortex AI with your trial. No credit card required.