Background
How we predict

The science behind anticipating earnings guidance revisions

Most market participants react to guidance changes after they are announced. We built a system that detects the linguistic fingerprints of an impending revision before the press release is drafted. First, we ingest thousands of corporate disclosures in real time. Next, we extract subtle shifts in language that human analysts routinely miss. Finally, we score the probability of a revision and deliver it to research teams with enough lead time to think, not just react.

Fully transparent

Every signal includes its linguistic triggers, confidence interval, and reasoning path — no black-box mystery.

Continuously validated

Models are backtested quarterly, retrained monthly, and audited against real outcomes, not simulated success.

Measurable lead time

Signals typically surface one to three weeks before official guidance revisions reach the market.

Three principles that keep our models honest

Fundamental intuition meets machine scale

Experienced analysts know that tone matters, that hedging intensifies before bad news, that precision often signals confidence. We encode this intuition into features that machines can measure at scale — across thousands of companies, every quarter, without fatigue or bias. The result is not a replacement for human judgment but a force multiplier for the patterns analysts already suspect.

Unstructured data deserves structured rigor

Earnings calls, SEC filings, and management presentations are messy. Language is ambiguous, context-dependent, and full of corporate euphemism. We treat this unstructured text with the same methodological discipline that quants apply to price data — defining features, establishing baselines, measuring drift, and validating signals against ground truth rather than convenient proxies.

Predictive power decays without maintenance

A model that worked last quarter may fail next quarter as language evolves, sectors rotate, and executives learn to manage narrative expectations. We do not deploy and forget. Every model is monitored for performance degradation, retrained on fresh data, and adjusted when real-world outcomes diverge from predictions. Maintenance is not an afterthought — it is the core of sustainable predictive accuracy.
Abstract technology pipeline visualization

From raw transcript to ranked signal

A five-stage pipeline that transforms corporate language into actionable foresight, with validation at every step and no stage left unexamined.
1

Ingest and normalize

We collect earnings call transcripts, SEC filings, press releases, and management presentations from multiple sources in real time. Text is cleaned, standardized, and tagged with metadata including company, sector, date, and document type so downstream models operate on consistent inputs.

Coverage spans thousands of companies across North America, Europe, and Asia-Pacific with daily refresh cycles.
2

Extract linguistic features

Natural language processing identifies hedging patterns, qualifier density, forward-looking statement degradation, sentiment trajectory, and lexical specificity. Each feature is measured relative to a company-specific baseline established over prior quarters, so signals reflect genuine change rather than cross-sectional differences.

Over four hundred engineered features feed into the scoring layer, with feature importance tracked and reported.
3

Score and validate

Ensemble models rank revision probability by direction, magnitude, and timing. Each prediction carries a confidence interval calibrated against historical accuracy. Predictions are backtested, stress-tested, and compared against actual outcomes to surface model weaknesses before they affect live signals.

Quarterly model audits publish accuracy metrics, coverage gaps, and acknowledged limitations for partner review.

Model evolution

Founders sketching early prototype
2023
Prototype built

A weekend experiment in parsing earnings call sentiment with off-the-shelf NLP tools produced surprisingly directional results on a small test set.

Analyst reviewing backtest results
2023
First backtest

We tested the prototype against five years of guidance revisions and found measurable predictive signal in linguistic features, but also significant false positive rates in cyclical sectors.

Developer working on feature engineering code
2024
Feature engineering

We moved beyond simple sentiment to engineered features measuring hedging intensity, qualifier density shifts, and forward-looking language degradation relative to company baselines.

Team launching the live platform
2024
Live deployment

The first version of the platform went live with coverage of two hundred companies, delivering signals directly to three research partner desks via API and email alerts.

Global expansion visualization
2025
Coverage expansion

We expanded to over three hundred forty companies across three continents, adding Asia-Pacific coverage and multilingual processing for non-English earnings commentary.

Formal audit documentation
2025
Audit framework

We introduced quarterly public model audits documenting accuracy, latency, coverage gaps, and acknowledged limitations — because trust requires verifiable transparency.

Advanced calibration interface
2026
Confidence refinements

A rebuilt confidence interval engine now accounts for sector-specific volatility patterns, reducing false positives in cyclical industries and improving signal prioritization.

The people building the prediction engine

Dr. Elena Vasquez

Dr. Elena Vasquez

Lead Research Scientist

Elena spent a decade in computational linguistics before applying her expertise to financial text. She designed the feature extraction pipeline that translates corporate language into measurable signals and insists that every model output include a human-readable reasoning path.

Frequently asked questions

Common questions about how our models work, what they can and cannot do, and how we validate the signals we deliver to research teams.

Every signal includes a reasoning path showing which linguistic features shifted, how they compared to the company's historical baseline, and what confidence interval the model assigned. We do not hide the logic behind black-box outputs.

We track directional accuracy, lead time, and confidence calibration against actual guidance revisions. Results are published in quarterly audits that include both successes and failures, with no selective reporting.

Transcripts and filings are typically ingested within hours of publication. The pipeline refreshes daily, and signals update as new disclosures arrive, so predictions reflect the most recent available language.

Coverage spans over three hundred forty companies across North America, Europe, and Asia-Pacific, with emphasis on sectors where management commentary is frequent and guidance revisions are material to valuation.

Models are retrained monthly on fresh data and audited quarterly against real outcomes. When performance degrades in specific sectors, targeted adjustments are made before drift affects live signal quality.

We document false positives, analyze whether the miss was due to model error, data quality, or genuine unpredictability, and feed those findings back into feature engineering and training protocols.