Contextual Intelligence for Equity Research

The filing satisfies a legal obligation.
The language reveals the truth.

An AI-driven contextual intelligence platform that reads corporate disclosures as a connected system. Built on three independent studies: a 185-company diversified market study, a 37-company financial sector study, and a 34-company healthcare study. 222 unique companies, 2,761 company-years, 2015-2025. Validated out of sample across earnings prediction, balance sheet forecasting, and credit risk indicators.

32,983SEC Filings Analyzed
222Companies Tracked
11 Years2015-2025
500M+Words Analyzed by NLP
500K+Features Generated
3Independent Studies
The Platform

Read every filing as a connected system

ContextQuant ingests the full textual and financial output of publicly listed companies and measures every disclosure against what direct competitors are saying. The result is a peer-relative intelligence layer that surfaces what changed, where it diverges from the peer group, and what it historically predicts.

>_
NLP Filing Parser
Every 10-K, 10-Q, 8-K, and DEF 14A decomposed into individually labeled sections. Over 500 million words processed through the NLP pipeline. Risk factors extracted as discrete items. Sentiment scored using Loughran-McDonald financial dictionaries. TF-IDF specificity analysis separates company-tailored language from boilerplate.
Δ
Competitive Deltas
Every metric measured as a peer-relative deviation. Direct competitor mappings at the business-line level, not SIC codes. When a company's risk language becomes more specific while peers remain generic, that divergence is the signal.
Macro Regime Engine
The system knows which signal matters in which environment. Risk specificity strengthens during high-uncertainty periods. Sentiment drift strengthens in calm markets. Signal weights adjust dynamically to VIX, rate cycles, and policy regimes.
Multi-Source Ingest
SEC filings, earnings call transcripts, proxy statements, executive compensation, political contributions (FEC), 65 macroeconomic time series (FRED), XBRL financial ratios, and daily price data. 500,000+ features across three databases from a single integrated pipeline.
Out-of-Sample Validation
Every signal tested using a seven-test validation battery: walk-forward IC, Benjamini-Hochberg correction, panel fixed effects, year-by-year stability, VIX regime conditioning, Fama-French five-factor alpha, and economic magnitude. Replicated across three independent universes. No lookahead. No in-sample optimization.
Scalable Architecture
Built for expansion from 222 companies to full market coverage. North American markets first (EDGAR + SEDAR+), then European and Asia-Pacific filing repositories. The methodology is jurisdiction-agnostic. What changes is the filing format, not the insight.

Three studies. 222 companies. 135 signal tests. 35 significant out of sample.

We tested every signal on data the model had never seen, train before year t, test on year t, repeat ten times from 2016 to 2025. That span includes a global pandemic, a historic rate hiking cycle, a regional banking crisis, and a bull market. The original study found three confirmed earnings signals. The combination study found that filing language also predicts balance sheet changes, surviving all macro controls. All results replicated across three independent universes.

Risk Factor Specificity
H1: Specific risk language predicts underperformance
Avg IC (180d)-0.065
Avg Spread-5.9%
Consistency8 / 10 years
Best Year2023: t = -3.48
Confirmed · 10-Year Sample
Outlook Divergence
H9: Peer-relative forward tone predicts returns
Avg IC (90d)-0.134
Avg Spread-3.85%
p-value0.045
EdgePeer delta essential
Confirmed · Strongest IC
Transcript–Filing Gap
H11: Tone inconsistency between calls and filings
Avg IC (180d)-0.086
Avg Spread-6.59%
p-value< 0.05
SignalCredit + equity
Confirmed · Novel
Macro Regime Conditioning
H10: Signals are regime-complementary
Risk in HIGH4.8× stronger
Sentiment in LOW2.1× stronger
Composite IR1.41 · zero sign flips
Confirmed
8-K Events / Compensation
H5 & H6: Event clustering & pay structure
8-K EventsDirectional only
CompensationN=15 insufficient
StatusNeeds more data
Exploratory
Political Spending
H7: Direction depends on absolute vs. size-relative measurement
Total Spend IC+0.159
Intensity IC3/3 negative
Sample134 tickers / 406K records
Size-Dependent
AI Confidence Predicts ROA
Haiku confidence score predicts balance sheet health 1Q ahead
Financials OOS IC+0.398
Healthcare OOS IC+0.289
Diversified OOS IC+0.100
Macro controlt = +4.77, survives all
Confirmed · Strongest ROA Signal
Cross-Sector Replication
LM sentiment replicates IC = -0.090 to -0.103 across three universes
FinancialsIC = -0.103 (N = 1,965)
HealthcareIC = -0.100 (N = 2,009)
DiversifiedIC = -0.090 (N = 4,650)
FF5 alphat = -2.81, p = 0.005
Confirmed · Three Studies
Forward-Looking Ratio (Contrarian)
More future talk = present weakening
Financials OOS IC-0.282
Healthcare OOS IC-0.367
MechanismDeflecting from results
Contrarian · ROA Early Warning
Accruals Interaction
NLP 3x stronger when earnings quality is poor
High-accruals IC-0.128 OOS
Low-accruals IC-0.043 OOS
Walk-forward6 / 8 years confirmed
Confirmed · Walk-Forward
Regulation FD Amplifier
Voluntary disclosures amplify earnings signal 10x
With Reg FDIC = -0.138 (N = 1,411)
WithoutIC = -0.013 (N = 3,218)
Amplification~10x
Confirmed · Event-Driven
Credibility Selectivity
Signal works best where management is least reliable
Slight miss (ROA)IC = +0.376
Unreliable (ROA)IC = +0.334
Vol beater (EPS)IC = +0.090
Confirmed · Signal Routing
YearNICSpreadt-stat
2016106-0.056+0.3%+0.08
2017111-0.221-6.1%-2.29 *
2018111-0.227-9.9%-2.03 *
2019109+0.028+1.7%+0.42
2020113-0.163-16.1%-2.39 *
2021112-0.082-1.2%-0.20
2022110-0.110-7.7%-2.07 *
2023108-0.275-18.5%-3.48 **
2024110-0.061-4.8%-0.83
202574-0.034+6.9%+1.09
AVG (10yr)106-0.121-6.1%
* significant at 5%    ** significant at 1% (exceeds Harvey-Liu-Zhu t>3.0 threshold)
High Uncertainty Regime
Risk specificity IC-0.066
Composite 90-180d IC-0.137
Walk-forward hit rate6 / 6 test years
Low Uncertainty Regime
Sentiment IC-0.059
Sentiment 0-30d IC-0.030
Walk-forward hit rate4 / 4 calm years
Regime: VIX > 20 or Fed Funds Δ > 50bps / 6 months  ·  2016–2025 sample
Who It's For

One platform. Multiple applications.

Institutional

Contextual intelligence for equity research, credit risk, and trading desks. Four products: Filing Signal Feed, High-Conviction Alert, Management Credibility Index, Balance Sheet Health Signal.

  • Earnings prediction (IC = -0.103, FF5 orthogonal)
  • ROA prediction (IC = +0.398, survives macro)
  • Credit early warning (60-90 day lead time)
  • Reg FD event amplification (10x signal)
Explore Business Applications →
Academic

A peer-relative textual analysis framework validated across three independent universes with a seven-test battery, grounded in established literature.

  • Extends Campbell (2014), Jiang (2019), Cohen (2020)
  • Competitive delta framework (novel)
  • Haiku confidence as ROA predictor (novel)
  • Credibility-based signal selectivity (novel)
Explore Research →
Wealth Advisory

Plain-language quality metrics backed by quantitative validation. When your client asks "is this company getting better or worse," the confidence score answers.

  • Management credibility classification
  • AI confidence score (0-1, predicts ROA)
  • Forward-looking ratio as early warning
  • Client-ready narrative, not raw data
Explore Wealth Applications →

From 222 companies to global coverage

The methodology works. We proved it on three independent universes. The architecture is built to scale across jurisdictions, languages, and filing formats. The insight that peer-relative disclosure behavior predicts returns and balance sheet changes is not limited to US markets.

EDGAR (US)
SEDAR+ (Canada)
Companies House (UK)
AMF (France)
BaFin (Germany)
EDINET (Japan)
HKExnews (HK)
SGXNet (SG)
Get in Touch

Let's talk about what this means for you

Whether you are an institution exploring new signal sources or a researcher interested in collaboration, we would welcome the conversation.

info@contextquant.com

We typically respond within 24 hours.