Methodology
This is a developer's methodology log, not an analyst's manifesto. Where an established method exists, I name it. Where I built something by intuition, I say so. The honesty grade on each item is the point — not the polish.
Honesty Grades
Tier 1 — Clear Academic Basis
Items that directly implement or are inspired by established academic papers or standard techniques.
9 binary checks on financial statement quality (0-9).
Academic Basis
Joseph Piotroski, "Value Investing: The Use of Historical Financial Statement Information to Separate Winners from Losers", Journal of Accounting Research, 2000.
Grade Note
F7 (share dilution) is approximated via EPS/NI direction comparison — we lack historical shares_outstanding data.
Key Formula
Profitability (4): NI > 0, OCF > 0, ROA improving YoY, OCF > NI. Leverage (3): Debt ratio down, Current ratio up, No dilution. Efficiency (2): Gross margin up, Asset turnover up.
Limitations
- F7 dilution check is an approximation (EPS vs NI direction), not true share count comparison.
- Uses TTM quarterly sums, not annual filing periods as in the original paper.
- Japan companies use annual data only (yfinance), which may lag by up to 12 months.
8-point checklist for growth stocks with fundamental backing (0-8).
Academic Basis
Partha Mohanram, "Separating Winners from Losers among Low Book-to-Market Stocks using Financial Statement Analysis", Review of Accounting Studies, 2005.
Grade Note
Stability checks include a direction-aware pass (consistent uptrend overrides variance) — this extension is not in the original paper.
Key Formula
Profitability (3): ROA, OCF/Assets, OCF > NI vs sector median. Stability (2): ROA variance, Revenue growth variance. Investment (3): R&D/Assets, CapEx/Assets, SG&A/Assets vs sector median.
Limitations
- Sector medians are computed from ~110 companies in our database, not the full market universe.
- Direction-aware pass is a custom extension; it may overcount for cyclical stocks with secular uptrends.
7 factors (C·A·N·S·L·I·M) each scored 0-100, heuristically adapted from O'Neil.
Academic Basis
William O'Neil, "How to Make Money in Stocks", 1988. Originally a screening tool, not a scoring system.
Grade Note
Scoring 0-100 per factor is our interpretation. I (Institutional) uses KR/TW institutional flow data, which differs from US mutual fund ownership in the original.
Limitations
- Converting a qualitative screening framework to quantitative scores is inherently subjective.
- I-factor data availability differs by market (KR: daily institutional flow, US: quarterly 13F only).
- M-factor (market direction) applies a single global market read to all tickers.
14-period Relative Strength Index.
Academic Basis
J. Welles Wilder Jr., "New Concepts in Technical Trading Systems", 1978.
Limitations
- Standard implementation with no modifications.
On-Balance Volume — cumulative volume tracking price direction.
Academic Basis
Joseph Granville, "Granville's New Key to Stock Market Profits", 1963.
Limitations
- Standard implementation with no modifications.
Relative Strength percentile rank within sector.
Academic Basis
O'Neil RS rating methodology. Formula: (count_below / (n-1)) × 99.
Limitations
- Sector assignments are manually tagged in our database; some companies may be mis-categorized.
Standard Pearson correlation used for cross-market lead-lag analysis.
Academic Basis
Karl Pearson (1896). Standard econometric lead-lag technique.
Grade Note
The JP→TW→KR ripple relationship discovered through this analysis may reflect a shared semiconductor cycle rather than true causation.
Limitations
- Correlation detected in the JP→TW→KR chain may be a shared-cycle artifact, not true causation.
- korea_exports table has multiple rows per month — always filtered by category with sufficient limit.
Additive smoothing for small-sample sentiment ratios.
Academic Basis
Laplace additive smoothing. Formula: (positive + 2.5) / (total + 5).
Limitations
- Prior strength (alpha=2.5) was set by intuition, not calibrated to data.
Logistic function for smooth 0-100 normalization without hard clamping.
Academic Basis
Logistic function (Verhulst, 1838). Center and steepness parameters are domain-tuned.
Grade Note
The function shape is standard; the center/steepness parameters are chosen by intuition, not calibrated.
Limitations
- Center and steepness parameters are set by intuition, not formal calibration.
Fact-axis composite using weighted geometric mean of GPU rental price and DC construction pipeline scores. CapEx news sentiment is separated into a narrative axis.
Academic Basis
OECD, "Handbook on Constructing Composite Indicators", 2008 — compensability problem: weighted arithmetic mean assumes full compensability, so a deficit in one axis is fully offset by a surplus in another. Geometric mean is less compensatory and penalizes low values, preventing one axis from masking another (the approach adopted by the UNDP Human Development Index in 2010).
Grade Note
Migrated 2026-06 from weighted arithmetic mean (gpu·0.4 + dc·0.3 + capex·0.3), which let CapEx news sentiment (a narrative input) dominate ~67% of the composite and mask an extreme DC pipeline reading. Separating sentiment follows OECD guidance that "highly different dimensions" should not be aggregated compensatorily. Internal precedent: the Narrative-vs-Fact Divergence card already separates sentiment from fact.
Key Formula
overall = (gpu^0.4 × dc^0.3)^(1/0.7), each axis clamped to [1, 99] to avoid log(0). CapEx sentiment returned separately; gap = fact − narrative.
Limitations
- Axis weights (GPU 0.4, DC 0.3) are still domain-intuition, not PCA-calibrated.
- DC pipeline MW is 94.9% note-extracted (inferred), not confirmed disclosure.
- Geometric mean reduces but does not eliminate compensability; not fully non-compensatory.
- Label thresholds (0-20-40-60-80) inherited from arithmetic era — distribution fit under geometric aggregation pending review as data accumulates.
28-day moving-average baseline per company×keyword; current value flagged new/stable/surge/drop by deviation from baseline.
Academic Basis
Standard moving-average baseline / deviation detection. Threshold-based anomaly flagging.
Grade Note
Evergreen keywords (e.g. CUDA) are tracked, not hidden — baseline deviation is the signal (surge = reinforcement, drop = pullback). Both directions matter.
Key Formula
baseline = mean(last 28 observed days); deviation% = (current − baseline) / baseline × 100. States: |dev| ≤ THRESHOLD → stable; > +THRESHOLD → surge; < −THRESHOLD → drop; <28 days → collecting.
Limitations
- Job posting APIs return only currently-open roles; no historical backfill (series starts at observation date 2026-06-18).
- THRESHOLD_PCT and 28-day window are intuition-set, not calibrated.
- Signal validity (does hiring lead the market?) requires accumulation to verify.
Tier 2 — Built by Intuition, Post-hoc Match
Items built by intuition, where similarities to known methodology were discovered after the fact.
Percentile-based min-max scaling with winsorizing at P5/P95.
Academic Basis
Winsorized normalization (Winsor; Ruppert, "Statistics and Data Analysis for Financial Engineering", 2004).
Grade Note
Built by intuition (clip extremes, scale to 0-100), later found to match standard winsorized normalization.
Key Formula
(return - P5) / (P95 - P5) × 100, clamped to [0, 100].
Limitations
- P5/P95 boundaries are hardcoded constants — they do not adapt as the data distribution shifts over time.
Excludes broad market selloff days from signal computation.
Academic Basis
Conceptually similar to detoning / market-common-component removal (Lopez de Prado, "Machine Learning for Asset Managers", 2020, Ch. 2).
Grade Note
A heuristic threshold (5+ tickers with ≥1% drop = market day, exclude) that later matched the concept of common-component removal.
Limitations
- Threshold of 5 is a heuristic magic number, not derived from eigenvalue decomposition.
- Does not distinguish sector-specific selloffs from true market-wide events.
Short-term fear-greed gauge (0-100) combining 3 weighted layers.
Academic Basis
Inspired by CNN Fear & Greed Index (7-indicator composite sentiment gauge).
Grade Note
CNN uses equal weights; we use differentiated weights (Market 40%, Supply 35%, Smart Money 25%) with fixed-range min-max instead of z-score.
Key Formula
Layer weights: Market Temperature 40%, Supply Chain Temperature 35%, Smart Money 25%.
Limitations
- Data accumulation period is short; extreme-value correction has not been validated.
- Fixed-range normalization may clip valid extremes in volatile markets.
Weighted composite indicator construction used across CPI, Tension, Divergence, HBM Tightness.
Academic Basis
OECD, "Handbook on Constructing Composite Indicators: Methodology and User Guide", 2008.
Grade Note
We follow the general pattern (normalize → weight → aggregate) but weights are set by domain intuition, not PCA or expert panel.
Limitations
- AI Infrastructure Tension migrated from weighted arithmetic to geometric mean (see dedicated entry in Tier 1); other composites (CPI, HBM Tightness) still use weighted arithmetic mean.
- Weights are intuition-based, not calibrated via PCA or sensitivity analysis.
- Data latency mixing: real-time prices coexist with quarterly filings (up to 45-day delay).
- GPU rental price sourced from Vast.ai only (single source risk).
- TSMC revenue used as CoWoS packaging proxy — not a direct measurement.
Product/tech keywords extracted from job titles and surfaced when newly appearing or surging, then cross-checked against the news corpus.
Academic Basis
Inspired by internal topic-surge detection (tag/n-gram surge). News cross-check labels keywords leading (hiring yes / news no) vs coincident (hiring yes / news yes).
Grade Note
News cross-check is a LABEL, not a filter — a keyword appearing in hiring but not yet in news is kept and flagged "leading" (highest value), not discarded.
Limitations
- Title parsing is INFERRED (stopword-filtered), distinct from facet-based MEASURED categories.
- "Leading" label needs new/surge precondition to exclude evergreen terms (refinement pending).
Tier 3 — Intuition, No Academic Source
Items built on intuition where no matching academic method has been found. This is honesty, not weakness.
Bitcoin hashrate used as a heuristic signal for GPU demand cycles.
Academic Basis
No academic source found.
Limitations
- The hashrate → GPU demand linkage is heuristic; ASIC mining dominance means the relationship is weakening.
- No backtesting performed on the causal claim.
Maps DRAM price direction × inventory direction to 8 fixed risk scores.
Academic Basis
No academic source found.
Grade Note
Score thresholds (±7.5, ±15, ±5) were set by intuition without backtesting. Flat-band boundary behavior is not validated.
Limitations
- All thresholds are intuition-based magic numbers with no backtest.
- Flat-band boundaries (DSI or DRAM near zero change) are not validated.
The "5 tickers" cutoff for identifying market-wide selloff days.
Academic Basis
No academic source found.
Grade Note
"5" is an undocumented magic number — a hard cut that ignores the underlying distribution.
Limitations
- Magic number 5 has no distributional basis; does not scale with universe size changes.
AI datacenter MW (construction/proposed, build-stage weighted) propagated equally to 4 supply-chain nodes (memory, power_infra, networking, packaging).
Academic Basis
No academic source found. Reuses internal hyperscaler-as-demand-source rule (big-tech capex sets big-tech ticker neutral, propagates demand to supply chain).
Grade Note
Equal 1/N split across 4 nodes is a placeholder — real part intensity differs (memory/power likely larger). Node coefficients not yet applied. Status weights (construction 1.0, proposed 0.3, operational/unknown excluded) are intuition-set.
Limitations
- Equal split is a placeholder, not a measured bill-of-materials.
- 94.9% of propagated MW is note-extracted (inferred), 5.1% confirmed.
- Hyperscaler operators excluded from demand (demand source, not signal).
- Signal validity unverified; accumulation required.
Rule-based extraction of MW and construction/completion dates from free-text note field.
Academic Basis
No academic source found. Regex/rule-based: GW×1000, total>phase>single, "potential/expansion" excluded as future capacity.
Grade Note
Raised MW total 3,519 → 39,673 (11×) by surfacing large projects previously unstructured (e.g. Fermi 11GW). Existing confirmed MW never overwritten; extracted values flagged inferred.
Limitations
- Free-text parsing; multi-MW notes without a "total" are queued, not force-resolved.
- Date precision is coarse (year-only → year start; flagged date_inferred).
- 94.9% inferred — supplemental to confirmed disclosure, not equivalent.
Appendix — Composite Score Detail
Weighted sum of 4 components (0-100). Overall investment attractiveness score per stock.
Formula
raw = (CAN SLIM × 0.40) + (Trend × 0.25) + (Quality × 0.25) − (Risk × 0.10) + 5
Components
| CAN SLIM | 40% | Growth, supply/demand, leadership |
| Trend | 25% | EMA structure, RSI, ADX, 3M momentum |
| Quality | 25% | ROE, operating margin, D/E, revenue growth |
| Risk | -10% | ATR, 52w drawdown, D/E, net margin, BB%B |
Grade Thresholds
≥75: STRONG · ≥60: WATCH · ≥45: NEUTRAL · ≥30: CAUTION · <30: AVOID
Limitations
- Component weights (40/25/25/10) are intuition-based, not statistically optimized.
- Risk is subtracted, so high-risk stocks can still score well if other components are strong.
- +5 baseline offset adjusts distribution center; no theoretical basis.
CapEx sustainability of 4 AI hyperscalers (Alphabet, Microsoft, Amazon, Meta). -50 to +50.
Structure
| Base Pressure | 0 ~ +25 | CAPEX/OCF percentile vs 9yr history |
| GPU Signal | -10 ~ +10 | 30d rental price trend |
| DRAM/DSI Signal | -15 ~ +15 | Inventory-price interaction |
Limitations
- NVIDIA is NOT included — it is a supplier, not a hyperscaler.
- CAPEX/OCF data depends on quarterly filings (up to 45-day delay).
Weighted 3-axis composite (0-100). Higher = tighter HBM supply.
Axes
| Hyperscaler CAPEX | ~36% | Demand pressure |
| HBM Supply Allocation | ~36% | SK Hynix, Samsung, Micron |
| TSMC Packaging Load | ~28% | CoWoS bottleneck proxy |
HBM ASP Direction is currently excluded — HBM contract pricing is not publicly disclosed. DDR5 spot is accumulating as a proxy candidate.
This document is a living methodology log, updated with each new feature. Last updated: 2026-06-18.
All data and analysis on this site are for informational purposes only and do not constitute investment advice. Investment decisions should be made based on your own judgment and responsibility.