AI Supply Chain Research · Credibility Audit

How Reliable Is SemiAnalysis?

An independent event study of what happens after the most influential newsletter in AI hardware praises or pans a stock.  ← back to the series

Everything in this book is distilled from SemiAnalysis, so the reader deserves to know how much this source can be trusted, and for what. We ran an independent event study: 27 parallel agents extracted every stance-bearing company mention from 311 articles (664 single-stock events across 119 tickers; 384 bullish / 168 bearish), then measured geometric excess returns vs. the SMH semiconductor ETF over 1M/1Y/3Y from publication date. Methodology: equal-weight by name with ticker clustering (to kill the soft look-ahead of winners being covered repeatedly), pre-IPO look-ahead removed (ARM/CRWV etc.), delistings settled at last price, and three adversarial audit rounds fixing eight methodological defects. One-line verdict: the narratives are remarkably right, but the stock picks carry no alpha — which makes this book's use of the source (mechanisms and bottlenecks, not tickers) exactly the right one.

Key Findings

1.
The only significant leading signal lasts ~1 month: bullish names earn +2.0% excess at 1M (t=1.63), and the bull-minus-bear spread is +3.9% (Welch t=2.33) — the study's only significant positive result.
2.
Long horizons reverse: the typical bullish name lags SMH by -5.2% at 1Y and -27.2% at 3Y (t=-4.49, significantly negative). Buying the bull list loses to simply holding SMH.
3.
Aggregate returns are a beta illusion: the big absolute gains come from the semi bull market plus NVDA being covered 62 times; excluding the five most-covered names (NVDA/TSM/AMD/GOOGL/AVGO), the remaining bullish names' 1Y excess is -6.1%.
4.
Bearish calls beat bullish ones: bearish names lag SMH by -10.7% at 1Y (t=-2.90, significant). INTC was panned 30 times and Samsung 15 — both played out. Whom SA pans is closer to a tradeable signal than whom it praises.
5.
Right industry, wrong stock: at sector level only AI accelerators beat an equal-weight peer index (entirely carried by NVDA/AVGO), while packaging/OSAT and foundry are significantly negative at 3Y. The memory-shortage, CoWoS-bottleneck and datacenter-power narratives were all right — but the money went to NVDA/TSM, and same-theme second-tier names underperformed.

Post-Coverage Drift at a Glance

Bullish 1M
+2.0%
Bullish 1Y
-5.2%
Bullish 3Y
-27.2%
Bearish 1M
-1.9%
Bearish 1Y
-10.7%
Bearish 3Y
-17.4%

Bars show mean excess return vs SMH after coverage (equal-weight by name). For bearish coverage, negative = the call was right.

Bullish Scorecard

Bullish scorecard (names covered ≥5 times; equal-weight by name, geometric excess vs SMH)

TickerN1M1Y3Y
ARM5+0.7%+54.3%
NVDA62-0.1%+30.4%+168.4%
GFS6+16.3%+22.0%-61.3%
META5+1.1%+17.2%-38.5%
AVGO20+2.2%+15.9%+76.9%
000660.KS15+3.1%+10.6%+49.3%
MU11+8.7%+8.7%-30.9%
LRCX8+3.3%+8.0%-14.9%
005930.KS5-4.9%+2.4%-58.0%
QCOM6+3.7%+0.2%-31.1%
TSM32+0.3%-3.1%-4.3%
AMD27-0.2%-4.6%-22.5%
GOOGL23+0.9%-6.4%-15.5%
AMAT10-3.3%-8.0%-16.4%
2454.TW8+4.2%-10.8%-15.1%
ASML6-2.0%-12.0%-37.8%
MSFT9-2.7%-17.7%-53.0%
8035.T6+0.2%-18.8%-38.6%
MRVL7-3.5%-22.2%-39.8%
INTC12-1.9%-27.3%-73.8%
0981.HK6-4.7%-31.8%+11.1%
AMZN7-7.2%-33.6%-25.4%

Click a column header to sort. Excess returns vs SMH, equal-weight by name, geometric, from publication date.

Bearish Scorecard

Bearish scorecard (covered ≥3 times; negative excess = the bearish call was right)

TickerN1M1Y3Y
NVDA7+6.2%+63.5%+242.3%
MRVL3-11.9%+34.0%-30.8%
AVGO6-2.7%+20.5%+107.2%
AAPL3-0.5%+12.5%-19.7%
CRDO3-1.1%+10.4%+260.8%
LRCX3+9.0%+1.2%-3.1%
MU6-4.3%+0.6%+69.1%
2454.TW4+3.5%-4.7%-22.0%
GOOGL3+6.4%-8.2%-6.0%
005930.KS15-1.1%-8.8%-50.3%
TSM3-1.2%-11.5%+7.9%
QCOM5+0.5%-16.1%-42.8%
ASML4+0.6%-19.4%-35.0%
GFS3+12.0%-21.9%-76.8%
AMD8-4.0%-23.1%-36.7%
INTC30-0.5%-24.5%-61.7%
ARM6-7.1%-53.5%

Click a column header to sort. Excess returns vs SMH, equal-weight by name, geometric, from publication date.

How to Use the Source

How to use: treat SemiAnalysis as a microscope on mechanisms and bottlenecks, not a buy list. That is precisely what this book distills — chain structure, bottleneck propagation, cost economics — the parts of the source that verify best. For trading: only the ~1-month drift is usable (and thin), bearish calls deserve real weight, and at long horizons return to a portfolio/beta framework.

Methodology & Limits

Scope: events dated at publication; benchmark = SMH; corpus includes paid-article previews (79% of full text), stances extracted from available text. This is the author's independent analysis, unaffiliated with SemiAnalysis.

Independence & sourcing. This is independent analysis by Yicheng Yang, distilled from publicly accessible SemiAnalysis articles (free posts and free previews; no paywall circumvention) and verified against the underlying text. It is not affiliated with, endorsed by, or a substitute for SemiAnalysis — subscribe there for the full research. All referenced claims are sourced and linked per SemiAnalysis's attribution terms. No SemiAnalysis images are reproduced. Nothing here is investment advice.