Earnings Mesh - Intellidimension

Goals & Setup

For this case study, we built a catalog of 145 liquid public tickers and ran meshes on their 2025 earnings events to answer two questions:

Qualitative: Does a multi-agent mesh surface deeper or different qualitative insights than you'd get from the usual "web consensus" (news, broker notes, blogs, etc.)?
Quantitative: Do mesh-derived scores carry real, testable signal about short-term returns—and do they beat a baseline model that scores the same brief without the mesh deliberation?

Off-Platform: Event Monitoring & Context Assembly

We wrote an orchestration layer that sits outside the Mesh Platform to:

Monitor earnings calendars and detect new events,
Gather the necessary context—transcripts, filings, background briefs, prior quarter data—from external sources, and
Submit fully-assembled case versions to the platform via the API.

This orchestration is not part of the core platform—it's case-study-specific scaffolding we built to automate the workflow. In a production setting, clients would build their own event triggers and context pipelines tailored to their data sources and research processes.

On-Platform: Mesh Execution & Output Generation

Once a case version is submitted, the platform:

Treats each ticker × quarter as a separate case version (e.g., AAPL FY2025 Q2),
Runs a mesh that ingests the background brief, earnings call transcript, and any additional filings (10-Q/10-K, ESG, policy docs, etc.), and
Produces a set of artifacts (transcript, analysis, FAQ) plus scores that make meshes comparable across the catalog.

What You Get for Each Earnings Event

For every case version (e.g., AAPL FY2025 Q2), the platform produces:

Investor FAQ (for consumption & sharing): 30-second TL;DR, "What changed vs last quarter?", near-term risks, long-term opportunities, geography lens, bull/bear setups, and what to watch next 1–2 periods. View example →
Mesh Analysis Report (for analysts): Executive snapshot, emergent themes with evidence tags, cross-theme interactions, faction analysis, systematic blind spots, and data gaps. View example →
Mesh Transcript (for audit & research): Full, role-labeled transcript of the multi-agent debate with judge scores, Oracle data inserts, and amplification patterns—lets you see why the mesh landed where it did and where experts disagree. View example →
Scores (for models & comparison): Pointwise metrics (pillar scores, narrative quality, uncertainty) and pairwise ELO rankings that let you compare events across the catalog and use mesh outputs as features in predictive models.

Because all four outputs are generated from the same mesh, your human readers and your models are working off a single, coherent view.

Using Mesh Scores as Features in a 5-Day Excess Return Model

The mesh isn't just for reading—it produces structured scores that can be used as features in predictive models.

To validate that those scores actually contain signal, we built two models that predict 5-day excess return after the print:

Mesh model: uses features derived from mesh outputs—pillar metrics, narrative/complexity/uncertainty scores, ELO rankings.
Context-only baseline: uses features scored directly from the input brief (same rubric, no mesh deliberation).

Model Setup

Target: 5-day excess return vs a benchmark after the earnings date.
Regime: train on history before a cutoff date, test on future events only, with group-aware splits by ticker to avoid leakage.
Comparison: both models use the same training data, test window, and evaluation framework—only the feature source differs.

Results: Mesh Model vs Context-Only Baseline

The mesh model significantly outperformed the context-only baseline:

Higher AUC: better discrimination between positive and negative excess returns.
Lower Brier score: more calibrated probability estimates.
Higher trading profits: in simulated backtesting with a 3-bucket rule (short when strongly negative, no trade in the middle, long when strongly positive), the mesh model produced meaningfully better risk-adjusted returns in both long and short buckets.

What This Tells You

The reference model is not the product; it's a validation that the mesh deliberation process adds real information beyond what you'd get from scoring the input brief alone.
These results are backtested, over a limited period, and are intended to demonstrate signal, not promise performance.

Case Study: Public Earnings Analysis