Methodology
How Early Signal analyzes the AI ecosystem and constructs portfolios.
10-Layer AI Supply Chain Taxonomy
Every stock in the universe is assigned to one of ten layers that trace the AI value chain from raw materials to secondary beneficiaries. Signals propagate along upstream/downstream edges so that, for example, a foundry capacity squeeze in Layer 1 ripples into chip and infrastructure layers before reaching applications.
| Layer | Name | Example Tickers |
|---|---|---|
| 0 | Raw Materials & Components | MP, ENTG, MKSI |
| 1 | Semiconductor Manufacturing | TSM, ASML, AMAT, LRCX, KLAC |
| 2 | Chips & Accelerators | NVDA, AMD, AVGO, MRVL, MU, ARM |
| 3 | Infrastructure | ANET, VRT, EQIX, DLR, CRDO |
| 4 | Cloud & Compute Platforms | MSFT, AMZN, GOOGL, META, ORCL |
| 5 | AI Platforms & Tools | MDB, SNOW, DDOG, PLTR, ESTC |
| 6 | Foundation Models & AI Services | MSFT, GOOGL, META, AI |
| 7 | AI Applications | CRM, NOW, WDAY, CRWD, PANW |
| 8 | AI-Enabled Transformation | JPM, UNH, WMT, JNJ, CAT |
| 9 | Secondary & Tertiary Beneficiaries | VST, CEG, NRG, PWR, EMR |
Signal Architecture
Composite signals are built from three timeframe-specific components, each composed of weighted sub-signals.
Daily Signal
- Momentum
- 30%
- Options Flow
- 20%
- News Sentiment
- 20%
- Earnings Surprise
- 20%
- Insider Trading
- 10%
Weekly Signal
- Analyst Revisions
- 30%
- Sector Rotation
- 25%
- Supply Chain Leading Indicators
- 25%
- Technical Levels
- 20%
Quarterly Signal
- Fundamental Shifts
- 35% — revenue acceleration, margin trends
- Valuation Resets
- 25% — multiple expansion / compression
- CapEx Trend Analysis
- 20%
- Market Share Changes
- 20%
Multi-Timeframe Aggregation
The three timeframes are blended with layer-specific weights. Layers closer to raw materials lean quarterly; layers closer to applications lean daily.
| Layer | Daily | Weekly | Quarterly |
|---|---|---|---|
| 0 Raw Materials | 15% | 35% | 50% |
| 1 Semiconductor Mfg | 20% | 35% | 45% |
| 2 Chips & Accelerators | 30% | 35% | 35% |
| 3 Infrastructure | 15% | 40% | 45% |
| 4 Cloud & Compute | 25% | 35% | 40% |
| 5 AI Platforms & Tools | 25% | 35% | 40% |
| 6 Foundation Models | 30% | 35% | 35% |
| 7 AI Applications | 25% | 35% | 40% |
| 8 AI Transformation | 15% | 35% | 50% |
| 9 Secondary Beneficiaries | 10% | 35% | 55% |
Conviction Scoring
- HIGH
- Composite signal ≥ 0.75
- MEDIUM
- Composite signal ≥ 0.50
- LOW
- Composite signal ≥ 0.25
When timeframes disagree in direction, a 30% disagreement penalty dampens the composite. Weekly signals decay over 5 days; quarterly signals decay over 65 days.
Cross-Layer Propagation
Strong signals in one layer propagate to connected layers with a lag that models real-world supply chain transmission times.
- Attenuation per hop
- 60% retained (40% decay)
- Maximum hops
- 3
- Minimum source signal
- 0.20 to initiate propagation
- Floor after propagation
- 0.05 (signals below this are discarded)
- Memory window
- 90 days
Propagation lags range from 7 days (Foundation Models to downstream) up to 90 days (upstream commodity layers). For example, a strong signal in Layer 1 (Semiconductor Manufacturing) reaches Layer 2 (Chips) after ~45 days at 60% strength, then Layer 3 (Infrastructure) after another ~30 days at 36% strength.
Bottleneck Detection
The engine continuously monitors supply chain chokepoints using capacity utilization, lead time stress, and inventory signals. Each chokepoint is classified into a severity tier.
| Tier | Severity Score | Beneficiary Signal | Victim Signal |
|---|---|---|---|
| CRITICAL | ≥ 0.75 | up to +0.80 | down to −0.60 |
| MODERATE | ≥ 0.50 | up to +0.48 | down to −0.36 |
| MILD | ≥ 0.25 | up to +0.24 | down to −0.18 |
| NORMAL | < 0.25 | No signal generated | |
- Tight utilization threshold
- 90%
- Normal utilization threshold
- 75%
Tickers are classified as beneficiaries (alternative suppliers, substitutes) or affected (dependent on the bottlenecked chokepoint). Bottleneck signals feed into the composite before portfolio construction.
Portfolio Construction
Positions are allocated across three sleeves based on conviction and timeframe alignment.
Sleeve Architecture (Moderate Risk)
- Core (60%)
- High-conviction, quarterly-driven — up to 15 names, min signal 0.50
- Satellite (30%)
- Medium-conviction, weekly-signal — up to 25 names, min signal 0.25
- Tactical (10%)
- Daily-signal short-term trades — up to 20 names, min signal 0.15
Position Sizing
Each position is sized by blending four methods:
- Fractional Kelly
- 40% weight — quarter-Kelly (0.25) default
- Volatility-adjusted
- 25% weight
- Risk-parity
- 20% weight
- Signal-proportional
- 15% weight
Concentration Limits
- Max single-position weight
- 8% (mega-caps: 10%)
- Per-layer caps
- Defined by layer metadata (1.5× typical annualized vol)
Risk Management
A drawdown-aware regime system dynamically scales gross exposure as the portfolio moves through four states.
| Regime | Drawdown Trigger | Gross Exposure |
|---|---|---|
| NORMAL | < 5% | 100% |
| CAUTION | ≥ 5% | 80% |
| DEFENSIVE | ≥ 10% | 60% |
| CRISIS | ≥ 20% | 30% |
Regime upgrades (back toward NORMAL) require 5 consecutive days below the stricter threshold.
- Per-layer caps
- 1.5× layer typical annualized vol
- Max single-position weight
- 8% (mega-caps: 10%)
Walk-Forward Backtest
Historical performance is estimated using a strict walk-forward methodology that prevents look-ahead bias.
- Rebalance frequency
- Weekly
- Warmup period
- 26 weeks (6 months)
- Score lookback
- 12 weeks
- Transaction cost
- 5 bps flat per round-trip turnover
- Top holdings
- 10 names, long-only by default
At each rebalance, the engine re-scores the full universe, selects the top-ranked names by composite signal, and re-weights using the same construction logic applied in live runs.
Stress Testing
The engine applies predefined macroeconomic and geopolitical shocks at the layer level to estimate portfolio impact.
| Scenario | Description | Heaviest Layer Shocks |
|---|---|---|
| Rates +100 bps | Higher discount rates compress growth multiples | Foundation Models −12%, AI Platforms −10% |
| Taiwan Supply Shock | Major disruption in advanced foundry output | Semiconductor Mfg −18%, Chips −12% |
| Hyperscaler CapEx Cut | Large cloud operators cut AI infrastructure spend | Infrastructure −15%, Chips −10% |
| Valuation Compression | Broad de-rating of high-multiple AI exposures | Foundation Models −16%, AI Platforms −14% |
Each scenario applies layer-level shocks and optional ticker-specific overrides, then rolls up estimated P&L across the current portfolio.
Data Sources & Quality
- Price data
- Live market feeds with daily close updates
- Synthetic fallback
- When live data is unavailable, the engine uses synthetic simulations with clear labeling
- Update frequency
- Daily signals updated each trading day; weekly and quarterly on schedule
- Data provenance
- Each run reports whether prices were live, mixed, or synthetic
- Freshness monitoring
- Dashboard displays data age and mode badges