Technical Methodology

NEXUS Alpha Signal Generation

Version 1.0 Published 2026-05-21 Coverage: 85 US Equities Validated: 2022–2025

1. Overview

NEXUS Alpha generates long-only investment signals (BUY, OVERWEIGHT, NEUTRAL) by applying FinBERT-based sentiment analysis to SEC EDGAR filings — 10-K annual reports, 10-Q quarterlies, and 8-K earnings releases — supplemented by large language model enrichment. Signals are produced within 10 minutes of a new filing appearing on EDGAR.

    Core thesis: Management language in regulatory filings contains predictive information about near-term stock performance. Positive sentiment in MD&A sections and earnings releases correlates with 60-day forward returns above the S&P 500. This edge persists after controlling for sector and market regime.
  

2. Data Sources

SEC EDGAR — Public filings database. 10-K, 10-Q, 8-K forms. Access via EDGAR full-text search API and Atom RSS feeds. No licensing cost; governed by SEC fair-use policy.
Management Discussion & Analysis (MD&A) — Primary text input for 10-K and 10-Q. Section 7 of the filing. 2,000–15,000 words per filing.
Earnings Press Releases — Item 2.02 and Item 7.01 of 8-K filings. 500–3,000 words. Filed 3–6 weeks before the full 10-Q, providing earlier signal.
Universe — 85 US-domiciled equities: S&P 500 top 25 + NASDAQ 100 additions. China ADRs excluded from backtest due to 20-F filing language differences.

3. Signal Pipeline

SEC EDGAR RSS

→

Filing Download

→

Text Extraction

→

FinBERT Scoring

→

LLM Enrichment

→

Signal Engine

→

DB + API + Alert

3.1 Text Extraction

Raw HTML from EDGAR is parsed using beautifulsoup4. For 10-K/10-Q, the MD&A section is isolated using regex pattern matching on Item 7 headers. For 8-K, Item 2.02 and Item 7.01 are extracted. Maximum input to NLP is 15,000 characters to stay within FinBERT's effective context window.

3.2 FinBERT Sentiment Scoring

ProsusAI/finbert — a BERT model fine-tuned on 10,000 financial sentences from analyst reports and earnings calls. Outputs three class probabilities: Positive, Negative, Neutral. Net sentiment is computed as Positive − Negative, ranging from −1.0 to +1.0.

Long text is split into 512-token chunks, scored independently, and averaged by token count. This preserves sentiment signal from both the beginning and the end of the MD&A, which contain different types of information (historical vs forward-looking).

3.3 LLM Enrichment (Ollama / Claude)

Optional second-stage enrichment generates a structured investment thesis. The prompt asks for: signal direction, confidence, bull case, bear case, key risks, revenue tone (BEAT/MISS/IN-LINE for 8-K), and guidance tone (RAISED/LOWERED/MAINTAINED).

Production uses qwen2.5:14b via Ollama (local, zero marginal cost). The enrichment model override is suppressed if LLM confidence is below 0.60 threshold.

3.4 Signal Engine

Net Sentiment	Trend Adjustment	Raw Signal
`> 0.30`	—	BUY
`0.10 – 0.30`	—	OVERWEIGHT
`−0.10 – 0.10`	—	NEUTRAL
`< −0.10`	—	NEUTRAL (long-only mode)

Trend adjustment: The current net sentiment is compared against the rolling average of the prior 4 filings for the same ticker. A positive trend (improving management tone QoQ) upgrades NEUTRAL → OVERWEIGHT. Negative trend is suppressed in long-only mode.

Confidence cap at 0.75: Information coefficient (IC) analysis shows negative correlation between stated confidence and actual 60-day return above 0.75. Signals are capped at 0.75 to avoid over-promising on less reliable high-sentiment readings.

4. Backtest Validation Results

    Methodology: Walk-forward validation. Training window: 2 years. Test window: 1 year. Windows rolled forward by 1 year. Signals evaluated at 20, 40, and 60-day forward returns. Benchmark: SPY total return. Universe: 25 US large-cap equities (2022–2025). Long-only BUY and OVERWEIGHT signals only.
  

1.20

Sharpe (60d)

70.6%

Win Rate vs SPY

+2.45%

Avg Excess vs SPY

+4.79%

Avg Signal Return

−20%

Max Drawdown

Total Signals

4.1 Alpha Decay by Hold Period

Hold Period	Avg Return	vs SPY	Win Rate	Sharpe	Max DD
20 days	+0.61%	−0.11%	58.8%	0.35	−27.5%
40 days	+3.63%	+1.34%	58.8%	1.13	−15.4%
60 days	+4.79%	+2.45%	70.6%	1.20	−20.0%

Key finding: The signal edge concentrates at longer holding periods. 20-day returns do not significantly beat SPY. The optimal holding period is 60 days, aligning with the quarterly reporting cycle.

4.2 Win Rate by Confidence Band (60-day hold)

Confidence Band	Signal Count	Avg Return	Win Rate
0.65 – 0.70	10	+12.3%	90%
0.70 – 0.80	20	+1.6%	60%
> 0.80 (pre-cap historical)	4	+2.1%	75%

Note: The confidence cap at 0.75 was applied after this backtest. The >0.80 bucket reflects historical signals generated before the cap; no live signals will exceed 0.75.

    Critical finding: The 0.65–0.70 confidence band dramatically outperforms higher confidence signals. IC (information coefficient) between stated confidence and 60-day return is −0.36 — a strong negative correlation. Higher confidence correlates with more extreme FinBERT readings, which occur on filings with boilerplate or sector-specific language that inflates sentiment artificially. The signal engine applies a hard cap at 0.75 and clients are advised to weight the 0.65–0.70 band most heavily.
  

4.3 By Signal Type (60-day hold)

Signal	Count	Avg Return	Win Rate
BUY	4	+2.1%	75%
OVERWEIGHT	30	+5.1%	70%

OVERWEIGHT signals — generated at moderate positive sentiment (net 0.10–0.30) with positive trend — comprise the majority of actionable signals and carry the strongest risk-adjusted return.

5. Long-Only Mode — Why SELL Signals Are Suppressed

The backtest evaluated long-short mode (including SELL and UNDERWEIGHT signals). Long-short Sharpe was −0.22 vs long-only Sharpe of +1.20. SELL signals on large-cap US equities systematically underperform due to:

Boilerplate risk language: SEC filings require disclosure of all material risks, creating negative language that FinBERT scores as bearish even in healthy companies.
Management hedging: Executives use cautious forward-looking language as legal protection, not as a signal of actual deterioration.
Short squeeze risk: Large-cap equities with high short interest can spike violently against short positions on any positive surprise.

SELL and UNDERWEIGHT signals are generated internally but not exposed via the API in long-only mode (default for all tiers). Long-short mode is available on the Institutional tier upon request with a separate risk disclosure agreement.

6. Live Signal Delivery

Signals are delivered via three channels simultaneously upon generation:

REST API — GET /signals/top, /signals/history, /signals/watchlist
WebSocket stream — wss://api.nexusalpha.io/ws/signals?api_key=... — pushed in real time
Email alert — HTML email to all active subscribers within seconds of signal generation

6.1 Latency SLA

Target: < 10 minutes from EDGAR filing acceptance to signal delivery. This is measured and published at /status. EDGAR updates RSS feeds every 10 minutes. Our watcher polls every 5 minutes, triggering the full NLP pipeline on first match.

7. Known Limitations & Risks

HIGH

Small sample size: 34 signals across 25 tickers over 3 years. Statistical significance is limited. The walk-forward methodology mitigates overfitting but does not eliminate it. More signals accumulate as universe expands to 85 tickers.

HIGH

Regime dependency: The backtest covers 2022–2025 — a period of rate rises, tech correction, and recovery. Performance in different macro regimes (deflationary, crisis, low-vol) is untested.

MED

Language model bias: FinBERT was trained on analyst reports, not MD&A sections specifically. Sector-specific boilerplate (pharma risk disclosures, energy regulatory language) may systematically skew scores.

MED

Filing timing: Some 10-Q filings lag the end of the quarter by 40–45 days. The signal may already be priced in by the time the filing is public. 8-K earnings releases (filed within 4 days of earnings) are the primary real-time signal source.

MED

Execution cost: Backtest returns do not include transaction costs, slippage, or market impact. For large allocations, these costs may materially reduce net returns.

HIGH

Signal overlay, not a portfolio: NEXUS Alpha generates ~40 signals/year on an 85-ticker universe. Capital is not continuously deployed. The per-signal excess return (+2.45% vs SPY at 60d) reflects information edge on each individual trade. A fully-compounded equal-weight backsimulation shows +9.5% total return vs SPY +26.2% over 2022–2025 — portfolio-level returns lag SPY because (a) only a subset of tickers generate signals in any quarter, and (b) the 2022 bear market hit concentrated filing-season entries. Institutional clients use these signals as overlays on existing book positions, not as a standalone fund.

8. Data Lineage & Audit Trail

Layer	Source	Immutability
Raw filings	SEC EDGAR HTTPS	Cached on ingestion; hash-verified
NLP scores	FinBERT inference	Written once, never overwritten
Signals	Signal engine	Unique constraint on (ticker, signal_date); upserts only
Paper track	Live pipeline	Append-only; `generated_at` is immutable; no backdating
API key usage	Every API call	Append-only audit log in `api_usage` table

9. Key API Endpoints

Endpoint	Tier	Description
`GET /signals/top`	1+	Highest-confidence BUY/OVERWEIGHT signals
`GET /signals/history`	1+	Full filterable signal time series
`GET /signals/watchlist`	1+	NEUTRAL signals — monitoring list
`GET /analytics/performance`	1+	Backtest stats + confidence bucket breakdown
`GET /filings/{ticker}`	2+	Raw sentiment history per ticker
`GET /data/export`	2+	Bulk CSV download — all signals
`GET /universe`	3	Full universe snapshot — latest signal per ticker
`GET /paper-track`	1+	Live paper track record with SPY comparison
`WS /ws/signals`	1+	Real-time signal push via WebSocket

10. Legal Disclaimer

NEXUS Alpha signals are generated by automated NLP models and are provided for informational and research purposes only. They do not constitute investment advice, a recommendation to buy or sell any security, or an offer of any investment product.

NEXUS Alpha is not a registered investment adviser under the Investment Advisers Act of 1940, nor is it registered with any equivalent regulatory body in any other jurisdiction. Past signal performance does not guarantee future results.

Users of NEXUS Alpha signals are solely responsible for their own investment decisions. NEXUS Alpha and its operators accept no liability for any losses arising from the use of these signals. By accessing the API, you confirm you have read and agree to the Terms of Service.