How an LLM Discovers Stocks — The 9-Stage Pipeline
PIPELINE v1.0
Click any stage to explore the details
IN: $AXTI 10x move
OUT: 9-factor checklist
In-context reasoning
Strip ticker-specific details
Extract 9 generalizable factors
Build scoring checklist
No tools used — pure reasoning
IN: 9-factor checklist
OUT: Target layers 2-4
Domain knowledge (training data)
Layer 1: Compute (GPU/HBM) — discovered, mega-cap
Layer 2: Packaging (CoWoS/glass) — active bottleneck
Layer 3: Optical (InP/lasers) — AXTI lives here
Layer 4: Raw materials & equipment — undiscovered
Layer 5: Assembly & test — competitive, no moat
Layer 6: DC infrastructure — emerging
IN: Target layers + exchange
OUT: Expected company types per exchange
Prior knowledge pre-filters search
Frankfurt → Layers 2-4 (equipment + materials)
LSE AIM → Layer 3 (epi wafers — IQE)
HKSE → Layers 1,3 (chips + optics IPOs)
SGX → Layers 4-5 (precision eng, no chokepoints)
Bursa → Layer 5 (OSAT, service companies)
IN: Exchange + target layers
OUT: ~10 results per query with snippets
web_search tool (3-5 word queries)
Pass 1: Broad — '{exchange} small cap semiconductor AI 2026'
Pass 2: Specific — 'MOCVD equipment InP epitaxial Germany'
Pass 3: Gap-fill — 'compound semiconductor Hong Kong listed'
Key: domain knowledge shapes query precision
BAD: 'best stocks to buy' → noise
GOOD: 'pBN crucible manufacturer listed' → exact hit
IN: Search result snippets
OUT: Enriched entity profiles
Entity extraction + web_fetch
Step A: Extract names, tickers, market caps from snippets
Step B: Hard filter — too large (>$10B), too small (<$10M)
Step C: Source quality ranking (SEC filing > news > forum)
Step D: web_fetch full articles for 'investigate' entities
Extract: business model, customers, competitors, margins
IN: Enriched entity profiles
OUT: Chokepoint score (0 to 1.0)
Single-question filter — most candidates fail here
How many exact competitors? (0-1 = monopoly)
Customer qualification required? (12-24mo = locked)
Product commodity or differentiated?
What % of customer COGS? (<5% = price insensitive)
YES = not chokepoint (score 0)
NO = genuine chokepoint (score 1.0)
IN: Chokepoint scores + entity data
OUT: Total score out of 9.0
≥6.0 = strong candidate | 4-5.9 = trade | <4 = compounder
F1: Micro-cap (<$500M = 1.0)
F2: Chokepoint (from Stage 5)
F3: Real revenue (>$50M + named customers = 1.0)
F4: Revenue inflection (backlog, beats, acceleration)
F5: Loss-to-profit transition (highest re-rating power)
F6: Narrative scarcity (only listed pure-play = 1.0)
F7: Capacity-constrained (sold out + pricing power)
F8: Under-followed (<3 analysts = 1.0)
F9: Float dynamics (tight float + high SI)
IN: Scored candidates
OUT: Validated + warning-flagged candidates
Reverse DCF + scenario stress test
Check 1: Already moved? (>300% from low = late entry)
Check 2: TAM ceiling? (10x implies what revenue?)
Check 3: Kill switch? (single event that zeros the stock)
Example: LPKF kill switch = Intel cancels glass program
Example: IQE kill switch = no buyer, balance sheet breaks
IN: Validated candidates
OUT: Final ranked watchlist
Adjusted score = raw - timing - risk - access
Timing discount: catalyst >24mo = -0.5
Risk adjustment: kill switch >40% = -0.5
Accessibility: thin liquidity or local broker required = -0.25
Sort by adjusted score descending
Output: Rank | Name | Exchange | Score | Upside | Timing
💡 KEY INSIGHT
The LLM's edge is not in searching — it's in knowing what to search for. A traditional screener filters 50,000 stocks by P/E in milliseconds. But it cannot express: "find the company that makes the furnaces used to grow the crystals that become the substrates that go into the lasers that power AI optical interconnects." That query requires understanding supply chain structure — domain knowledge, not data.
⚠️ KNOWN LIMITATIONS (7)
Language Gate
30-40% of Asian micro-cap alpha is in Mandarin/Japanese/Malay broker reports
No Quant Screener
Cannot run SQL across 50,000 stocks — reasons from structure, not data
Survivorship Bias
Checklist built from 1 winner — next 10x may have different profile
Temporal Blind Spot
Recent IPOs and strategy pivots may not be in training data
Private Companies
Best candidate might not be public yet (e.g., AMF acquired by GFS)
SEO Bias
Search ranks by SEO quality, not investment relevance
Confirmation Bias
Pipeline searches for supply chain fits — misses non-supply-chain catalysts
No redactions applied.
Published via wiis. Originals are encrypted at rest; this page reflects the redacted version. report abuse.
Discussion
No comments yet.