How everything is calculated

Every number on this site is derived from public data using transparent formulas. If you spot an error or want to reproduce a calculation, you have what you need.

Primer: the semiconductor industry in 90 seconds

Semiconductor companies don't all do the same thing. Knowing where a company sits in the value chain matters because each segment has different margins, cyclicality, and exposure to AI vs broader demand.

The value chain

Design software (EDA) — Synopsys (SNPS), Cadence (CDNS). Sells the tools every chip designer uses. Recurring revenue, very high margins, least cyclical part of the chain.

Chip design (fabless) — NVIDIA (NVDA), AMD, Broadcom (AVGO), Qualcomm (QCOM), Marvell (MRVL). Designs chips but outsources manufacturing. High margins, moderate-to-high cyclicality.

Manufacturing (foundry) — Taiwan Semiconductor (TSM), GlobalFoundries (GFS), United Microelectronics (UMC). Makes chips for everyone. Capital-heavy, huge fab investments, exposed to geopolitics.

Memory — Micron (MU), Samsung (005930.KS), SK Hynix (000660.KS), Western Digital (WDC), Kioxia (285A.T). Most cyclical part of the industry. DRAM/HBM tied to AI; NAND tied to PCs and phones.

Equipment (WFE) — ASML, Applied Materials (AMAT), Lam Research (LRCX), KLA (KLAC), Tokyo Electron (8035.T). Sells the machines that make chips. Leads the cycle by 6–12 months — when fabs order more equipment, more chips are coming.

Integrated device manufacturers (IDMs) — Texas Instruments (TXN), Analog Devices (ADI), NXP (NXPI), Infineon (IFX.DE). Design and manufacture their own chips. Most exposed to industrial, automotive, and broad-based demand — the opposite of the AI trade.

Why semis are cyclical

Building a fab takes 2–4 years and tens of billions of dollars. Demand can shift in months. So supply and demand rarely match: when chips are scarce, prices and earnings surge, capacity gets added; by the time that capacity comes online, demand has often cooled. Cycles tend to last 2–4 years peak-to-peak.

The implication: peak earnings often coincide with cycle tops, not bottoms. Stocks trading at the lowest P/E multiples can sometimes be the most expensive in absolute terms, because the "E" is unsustainable. This is the peak earnings trap, and it's why cycle-aware analysis matters.

What's different about today (mid-2026)

The current cycle is unusual. AI infrastructure demand has decoupled top-tier names (NVDA, AVGO, TSM, ASML, the HBM trio) from the broader cycle. Memory pricing, historically the most cyclical input, is now bifurcated — HBM is sold out into 2027 while NAND tracks a more conventional cycle. Inventory normalization has been uneven. Geopolitical risk (US–China export controls, tariffs) is structural, not transient.

That's the case for splitting sentiment into two parallel gauges. The old "single semi index" approach hides what is actually a two-track market.

Sentiment composite (v2)

Two parallel gauges score sentiment from 0 to 100 for two ticker groups. Same nine components, different baskets, so the gauges can diverge when the AI economy and the broad-based semi cycle move differently.

Ticker groups

Composite formula

The composite is 60% technical (price-driven, fast-moving) and 40% fundamental (slower-moving, regime-aware).

Why percentile rank for technicals

Distance-from-MA, RS-vs-SPY, and vol use percentile-rank rescaling against each gauge's trailing 252-day distribution. This makes the gauge self-calibrating: today's reading reflects "where this is vs. where it has been recently," which keeps the gauge useful in both quiet markets and supercycles. The pure-linear technicals (RSI, breadth) are already 0–100 by construction.

Why fundamentals at 40%

Pure-technical sentiment leads cycle peaks by zero days — by the time price momentum breaks, the move has happened. Fundamentals lead by months. Memory pricing, capex revisions, and EPS revisions all flag regime shifts before they show in the price tape.

The cost of including fundamentals: the composite moves more slowly. A pure-technical gauge can swing 10 points in a week; this one rarely moves more than 5. That's by design. We aren't trying to call short-term tops and bottoms.

Score interpretation

Cycle position (v3)

Estimates where the industry sits in its multi-year cycle. Independent from short-term sentiment. Four signals — two coincident, two forward-looking — weighted toward the forward ones, because price and revenue confirm where we already are while inventory and capex indicate where things are heading.

Signals

SOX carries the lowest weight for two reasons. It is the most lagging of the four, and semiconductor equity prices and industry revenue largely measure the same thing — at equal weights the two together would have been half the composite while capturing one underlying signal twice.

How forward signals work

Inventory days = (inventory dollars / quarterly revenue) × 91. We compute this for memory + IDM filers (MU, WDC, TXN, ADI, NXPI, ON, MCHP, INTC), then average across the universe. The cycle signal is the change in average inventory days over the trailing 4 quarters. Rising = oversupply building (bearish for prices, late-cycle). Falling = supply tightening.

WFE capex YoY = sum of TTM capex across AMAT, LRCX, KLAC, TER, ONTO vs. prior TTM. Equipment makers spend capex to expand their own production capacity based on order book visibility, which leads chip company orders by 1–2 quarters. Surging WFE capex = they see more orders coming. Falling WFE capex = order pipeline thinning.

Reading inventory against the demand regime

Falling inventory days normally indicate an early or mid-cycle phase: supply has been worked down and demand has not yet caught up. That reading holds in a supply-driven cycle. It inverts in a demand shock, where inventory days fall precisely because demand is outrunning what the industry can produce — a late-cycle symptom wearing an early-cycle appearance.

To separate the two, the inventory score carries a floor tied to revenue growth. Below +20% YoY billings growth the floor is inactive and the signal behaves exactly as described above. Above that it rises, reaching 70 at +60% growth, on the reasoning that the industry cannot be emerging from a trough while revenue is compounding at those rates.

This is a floor rather than a reversal. At extreme demand, both falling inventory (supply cannot keep pace) and rising inventory (the build finally catching up as a boom crests) are late-cycle indications, so simply inverting the mapping would discard the inventory-build warning that matters most near a turn. The underlying curve is left intact; the floor only prevents readings the demand data rules out.

Scale ranges

SOX YoY is scored across −50% to +150% and SIA billings YoY across −30% to +120%. Both ranges were widened from earlier versions, which capped at +100% and +60% respectively — levels the current cycle has exceeded, leaving those gauges pinned at their maxima and unable to register further change. Readings beyond the top of each range still score at the ceiling.

Zone interpretation

Memory pricing

Memory is the most cyclical sub-segment of semis and the most reliable leading indicator for the broader cycle. We track three categories:

• DRAM contract — high-volume server/PC memory. Direction reflects supply-demand balance for traditional computing.
• NAND flash — storage memory. Tied to PCs, smartphones, and enterprise SSD demand.
• HBM (high-bandwidth memory) — what AI accelerators use. Independent cycle from DRAM/NAND because supply is constrained by advanced packaging capacity.

Source: TrendForce press releases. We scrape the latest commentary and classify direction (rising / falling / flat) based on headline language. Falls back to a manually-curated file when the scrape fails.

Taiwan monthly revenue

Taiwan is where most leading-edge semiconductor manufacturing physically happens, and Taiwanese regulation requires every listed company to publish its monthly revenue by the 10th of the following month. US-listed chip companies report quarterly. That difference is the whole reason this is here: it is the same industry, read weeks earlier, because of a disclosure rule rather than any cleverness on our part.

August 2026 figures were published on 15 September — roughly a fortnight after month end, against about six weeks for SIA billings.

What the number is

The aggregate year-over-year change across every company the Taiwan Stock Exchange classifies as 半導體業 (semiconductor industry) — 96 companies as of August 2026, spanning foundry, IC design, packaging and test, and memory. TSMC is carried separately because it alone is a meaningful read on leading-edge demand.

We sum revenue across the sector and then take the year-over-year change. We do not average the individual companies' percentages. The distinction matters more than it sounds: in August 2026 the aggregate was +61.3%, while the mean of the per-company percentages was far higher because Nanya Technology posted +560% on a small base. An average lets the smallest company move the sector reading.

Only companies reporting both a current and a prior-year figure are included in the ratio, so the basket stays consistent month to month rather than shifting as companies list or delist.

How to read it against SIA billings

These two numbers will often disagree, and neither is wrong. SIA measures global semiconductor sales and therefore carries memory price inflation; Taiwan's listed names are weighted toward logic and foundry. In August 2026 SIA read +135.1% while Taiwan read +61.3%.

The gap between them is itself informative — it widens when memory pricing is driving the cycle and narrows when logic demand leads. That is why this sits alongside SIA rather than replacing it.

What it is not

• Not global. Taiwan-listed companies only. Samsung and SK Hynix, which dominate memory, are not in it.
• Not part of the cycle composite. It is displayed and recorded, but does not currently feed the cycle position score. Adding a signal to that model requires enough history to judge whether it improves the reading, and we are collecting that history now rather than guessing at a weight.
• Not revenue in USD. Figures are New Taiwan dollars, so currency moves affect the year-over-year comparison.

Source: Taiwan Stock Exchange open data, dataset t187ap05_L. Free, no key. Dates in that feed use the ROC calendar — 11508 means August 2026.

Korea chip exports

Korea is where most of the world's memory physically ships from. Samsung and SK Hynix dominate DRAM and NAND, and neither appears in Taiwan's listed basket or as a separable line inside SIA billings. Korea Customs publishes export values monthly by HS code, free and without registration, which makes this the cleanest public read on memory shipments that exists.

What the number is

HS 8542 — 전자집적회로, integrated circuits — is Korea's single largest export line. In August 2026 it was $38.2bn, roughly 39% of everything Korea exported that month. We also carry HS 8523 (solid-state storage: SSDs and flash modules), which moves with the NAND cycle.

Year-on-year is computed by matching the same HS code between the current month and the same month a year earlier. Where a code is absent from either month's top twenty we show no change rather than a computed one, because absence means unranked, not zero.

How to read it against SIA and Taiwan

These three disagree by design, and the ordering is the information. SIA is global and carries memory pricing but cannot separate it; Taiwan is weighted toward logic and foundry; Korea is the most memory-exposed of the three. When memory pricing leads a cycle, they rank Korea first, SIA second, Taiwan third. In August 2026 they read +203%, +135% and +61% respectively.

If that ordering inverts — Taiwan running ahead of Korea — the cycle is being driven by logic and foundry demand rather than memory. That is the signal this panel exists to make visible.

The headline hides the shape

Korea's total exports rose 68.7% year-on-year in August 2026, which averages together very different things. Integrated circuits were up 203%, solid-state storage 450%, and computer components 289%, while passenger vehicles fell 30% and shipbuilding 46%. The boom is narrow and almost entirely semiconductors. An aggregate trade number cannot show that; an HS-coded one can.

What it is not

• Not company revenue. These are customs clearance values in US dollars. Currency moves affect the comparison, and clearing customs is not the same as revenue recognised.
• Not memory alone. HS 8542 is all integrated circuits, so logic and analogue ship inside it too. It is memory-heavy because Korea is, not because the category excludes anything.
• Not part of the cycle composite. Displayed and recorded, but it does not feed the cycle position score. Adding a signal to that model needs enough history to judge whether it improves the reading.
• Not final. Monthly customs figures are revised.

Source: Korea Customs Service trade statistics portal, HS4 item statistics. Collected by the KOSPI property and read from there, so one pipeline serves both sites.

Earnings revisions

Tracks net analyst rating changes (upgrades minus downgrades) over the trailing 30 and 90 days, per ticker. Aggregate breadth — what % of the universe has net up-revisions vs net down-revisions — is a stronger predictor of forward earnings momentum than absolute consensus levels.

Source: Finnhub upgrade/downgrade endpoint, daily refresh.

Movers and leaders

Four short lists from the 38-ticker universe:

• Top gainers / laggards — biggest one-day percent moves.
• Most extended — trading furthest above their 50-day MA. Momentum leaders with mean-reversion risk.
• Least extended — trading closest to or below their 50-day MA. Relatively cool names in an otherwise hot sector.

None are buy or sell signals. They surface where the action and the gaps are.

Data sources and refresh cadence

Known limitations

• SOX year-over-year reading is computed as today's 21-day median price vs. the 21-day median centered on the same date one calendar year ago. Other publishers may report different "1-year" figures using rolling, annualized, or total-return methodologies. Our number is honest about what it measures: literal price ratio, smoothed against single-day outliers. Not directly comparable to "1Y annualized return" elsewhere.
• EDGAR data has a structural lag: 10-Q filings drop ~45 days after quarter-end. So inventory-days direction reflects the most recent quarter where filings exist, not real-time inventory.
• The AI gauge contains 4 of the 5 largest semi names by market cap. When NVDA, AVGO, TSM, or ASML have large moves, the composite reflects that concentration. This is a feature for tracking AI infrastructure exposure, not a bug — but worth knowing.
• Foreign filers (TSM, ASML, Samsung, SK Hynix, Infineon, Tokyo Electron) report on different cadences (20-F annually) and are not included in our EDGAR aggregates. This means inventory days and capex YoY are US-filer subsets, not global.
• Korean and Japanese tickers report on a different trading calendar than US markets. The composite handles missing days gracefully but means weekend US data won't include Friday Asia moves.
• Policy timeline is manually maintained, with a target of adding new BIS rules within 1–2 days of release.
• Yahoo Finance chart API and TrendForce scraping are unofficial; both can break without notice.

This is not investment advice

Everything on this site is publicly available information processed through transparent formulas. None of it constitutes investment advice. The semiconductor sector is highly cyclical and individual stocks can move significantly on company-specific news. Always do your own research and consult a qualified financial advisor before making investment decisions.