Research Report · Consolidated Backtest & Forward-Looking Portfolio · v2 (rationalized)

@jukan05 Consolidated Analysis: 16-Month Sector-Controlled Backtest, the Real Edge Is Theme Identification (Not Stock Picking), and a 12-Position Niche Portfolio for What Comes Next

Full consolidated analysis of @jukan05 (120K verified followers, supply-chain narrative analyst, “Tech otakus save the world”). Two parts in one report: a sector-controlled 16-month backtest that shows moderate but real signal-level alpha, plus a counterfactual mapping analysis proving the skill is theme identification (not specific stock picking) — and a 12-position forward-looking niche portfolio with explicit evidence labels separating direct cashtags from inference mappings, and a calibrated forward expectation that does not project a repeat of the historical outcome.
Published 2026-05-23 Backtest window: Jan 18, 2025 → May 22, 2026 (489 days, 16 months) 352 topic-classified signals Sector-control: vs SMH Buy-and-hold basket: 7 names Counterfactual: 5 alt mappings Forward portfolio: 12 positions
VERDICT: MODERATE SECTOR-CONTROLLED EDGE · SKILL = THEME ID · FORWARD EXPECTATION CALIBRATED
+5.82%
60-day alpha vs SMH (n=294, WR 57%)
+31.85%
180-day alpha vs SMH (n=175, WR 68%)
+26.44%
MEMORY_BULL 60d alpha (only strong-edge topic)
+294.86%
7-name basket 16m (concentrated in WDC+MU)
+120.33%
SMH semi ETF same window (fair benchmark)
~2.5×
Basket vs SMH (the honest excess)
1 of 11
DIRECT ×5 cashtag conviction (INTC, 17%)
12
Positions in forward niche portfolio
Prefer the print version? Download the full 13-page consolidated PDF (v2, rationalized): profile + methodology, signal inventory, horizon-based alpha, per-topic and per-ticker edge map, memory canonical topic, 16-month buy-and-hold honest decomposition, equity-curve chart with 5-name vs 7-name basket comparison, counterfactual alternative mappings, “why we don't repeat the memory trade”, fresh narratives table with EVIDENCE column, 12-position portfolio table with EVIDENCE column + theme breakdown, realistic forward expectations, full management rules, and the calibrated decision chain.
Download PDF ↓

Profile & What Makes This Account Different Not a stock-call account. A supply-chain analyst whose forecasts you map onto tickers yourself.

The personality. @jukan05 is a 20-month-old account (joined Sep 2024), 120,209 verified followers, 16,277 lifetime tweets (~27 per day). Bio: “Tech otakus save the world | Not Investment Advice | DYODD.” The content is a continuous feed of Korean and English posts about memory pricing, foundry capacity, AI compute procurement, and packaging bottlenecks. No ticker calls, no buy/sell language, no portfolio screenshots.

Why the backtest is non-trivial. Because the posts are topic-level, you cannot just track cashtags — you have to classify each tweet by topic and map each topic to its natural ticker basket. We did that with a regex-based classifier validated on a 50-tweet sample (~92% agreement) and built a signal stream of 352 distinct topic events over 16 months, averaging ~22 signals per month.

Part 1 — Backtest Evidence (16 months, sector-controlled) Does the supply-chain methodology actually produce alpha? Yes — and the magnitude is large.

Topic → Ticker Mapping Every signal in the dataset is one of these seven topics, mapped to its natural ticker basket.

TopicMapped TickersDirectionn (60d)Hypothesis
MEMORY_BULLMU, WDCLONG78Samsung HBM tight / DRAM allocated / SK Hynix HBM lead / NAND rally
MEMORY_BEARMU, WDCSHORT30Memory cycle topping / NAND oversupply
TSMC_TIGHTTSM, AMAT, LRCXLONG105Foundry sold out / capacity push / 2nm node / EUV bottleneck
NVDA_BULLNVDA, AVGOLONG68Jensen / Colossus / Blackwell demand
AVGO_BULLAVGOLONG11Custom silicon / hyperscaler ASIC
AMD_BULLAMDLONG1MI300 / MI325 demand (single signal)
APPLE_WEAKAAPLSHORT1iPhone / supply softness (single signal)

Entries: next RTH open after the signal tweet. Holds: 5 / 30 / 60 / 180 trading days. For each entry we record the ticker return, SPY, QQQ, SMH and the alpha against each. Because every mapped ticker sits inside the semiconductor sector, alpha vs SMH is the only honest benchmark — positive numbers there mean the topic call predicted outperformance against other semi names, not just against the broad market.

Sector-Controlled Alpha By Horizon Positive at every horizon and growing with the hold — the signature of structural edge, not noise.

HorizonnMean ReturnMean SMHAlpha vs SMHWR (α > 0)Alpha vs SPYAlpha vs QQQ
5-day352+2.61%+1.61%+1.00%55.7%+2.09%+2.03%
30-day339+8.13%+5.46%+2.67%57.5%+6.88%+6.54%
60-day294+11.92%+6.10%+5.82%56.5%+10.90%+10.71%
180-day175+62.47%+30.62%+31.85%68.0%+54.79%+53.27%

The signal needs time. Sector-control alpha grows with hold horizon.

5-day alpha is +1.00% vs SMH — right at the noise floor. The topic signal does not produce an immediate intraday-to-week pop. Move out to 30 days (+2.67%), 60 days (+5.82%), and 180 days (+31.85% with a 68% win rate). The supply-chain prediction takes weeks-to-months of earnings cycle, guidance updates, and sell-side estimate revisions to translate into a sector-relative repricing.

Per-Topic Asymmetry Not every topic carries alpha. MEMORY_BULL is the canonical edge; NVDA_BULL is fully priced in; MEMORY_BEAR is systematically inverted.

NO EDGE / INVERTED

NVDA_BULL + MEMORY_BEAR

Mapped: NVDA · or short MU/WDC
−1.65% · −16.60%
60-day alpha vs SMH (NVDA / MEMORY_BEAR)
  • NVDA_BULL (n=68): alpha −1.65%, WR 35% — over-mentioned, market priced in
  • MEMORY_BEAR (n=30): −16.60%, WR 33% — direction is inverted
  • Contrarian setup: when jukan05 turns BEAR on memory, the long pays +16.6% alpha
  • Even the BEAR call is an information signal — just flip the sign
  • Do not mirror these as posted

All Topics, 60-Day Alpha The full topic table. Sample size matters — single-signal topics (AMD_BULL, APPLE_WEAK) are anecdotal, not statistical.

TopicnAlpha vs SMHWRRead
MEMORY_BULL78+26.44%84.6%canonical edge
AMD_BULL1+26.30%100.0%n=1 anecdotal
TSMC_TIGHT105+2.44%56.2%marginal positive
AVGO_BULL11−1.61%54.5%noise — sample too small
NVDA_BULL68−1.65%35.3%no edge / priced in
APPLE_WEAK1−4.89%0.0%n=1 anecdotal
MEMORY_BEAR30−16.60%33.3%inverted — fade it

Per-Ticker 60-Day Alpha WDC and MU carry the basket. NVDA on jukan05 mentions actually underperforms the sector.

TickernAlpha vs SMHWRWhy
$AMD1+26.30%100.0%n=1 (single AMD_BULL signal)
$WDC54+16.23%70.4%Storage / NAND beneficiary of memory tightness
$MU54+12.73%70.4%Direct DRAM/HBM pure-play; jukan05's clearest call
$LRCX35+4.76%54.3%Memory etch / dep equipment; benefits from capex cycle
$AMAT35+2.77%62.9%Foundry / memory capex cycle
$TSM35−0.22%51.4%Coin flip — the TSMC topic is fully priced
$AVGO44−0.85%50.0%Mega-cap, no edge in mentions
$NVDA35−2.64%22.9%jukan05's NVDA mentions actually fade vs the sector
$AAPL1−4.89%0.0%n=1 (single APPLE_WEAK signal)

16-Month Buy-and-Hold — Honest Decomposition The basket beat NASDAQ by 8× over 16 months. The fair benchmark is SMH, not NASDAQ. The honest excess is closer to 2.5× sector.

The headline number sounds huge. Most of it is sector beta.

An equal-weight basket of the seven most-mentioned jukan05 tickers (AMAT, AVGO, LRCX, MU, NVDA, TSM, WDC) returned +294.86% over 16 months versus QQQ +37.59%. That is the eye-catching framing — but the broader semis sector (SMH) returned +120.33% in the same window. Comparing a semis basket to broad NASDAQ measures mostly the sector rally, not the signal. The fair benchmark is SMH, and the honest excess is roughly +175 percentage points over SMH (~2.5× sector).

Series16-Month Total ReturnNote
jukan05 basket (EQ-weight, 7 names)+294.86%AMAT, AVGO, LRCX, MU, NVDA, TSM, WDC
SMH (semiconductor ETF) — fair benchmark+120.33%Honest excess: ~+175 pts (~2.5× sector)
NASDAQ-100 (QQQ)+37.59%Includes sector beta — flattering comparison
NASDAQ Composite (^IXIC)+33.34%Broad tech benchmark
S&P 500 (SPY)+25.44%Broad market benchmark

Per-Ticker 16-Month Contribution The excess over SMH is concentrated in two names: WDC and MU. The remaining five basically track the sector.

Ticker16-Month ReturnDriving Theme
$WDC+854.28%Memory supercycle + spin-off rerate — largest driver
$MU+589.14%Canonical DRAM / HBM tightness thesis — second driver
$LRCX+278.55%Wafer-fab equipment (3D DRAM transition)
$AMAT+126.87%Memory + foundry capex
SMH (sector reference)+120.33%5 of 7 names cluster around this line
$TSM+87.99%Foundry capacity / EUV bottleneck
$AVGO+74.26%Custom ASIC for hyperscalers
$NVDA+52.95%AI compute beta — smallest contribution despite biggest narrative

The concentration fact, stated plainly. The two memory names (WDC + MU) carry the majority of the excess over SMH. Removing them, the remaining 5-name basket basically tracks SMH. This is not a “we got lucky on two names” story (the counterfactual section below shows why), but it is a “the canonical edge was the MEMORY theme, expressed through two memory-beneficiary tickers” story. Other topics (TSMC tight, NVDA bull, AVGO custom) rode the broader semi rally but added little on top.

Counterfactual: What If We Had Picked Different Memory Names? The skill jukan05 demonstrated was theme identification — not picking WDC vs MU vs SNDK. Any reasonable memory mapping beat the sector.

The question. jukan05 never said “buy WDC” or “buy MU.” We mapped his “Samsung HBM tight” narrative to WDC + MU because they were the cleanest US-listed memory beneficiaries. Is the result robust to that choice, or did we get lucky on the picks?

The method. Keep the other 5 names (AMAT, AVGO, LRCX, NVDA, TSM — fixed-leg mean +124.12%) constant. Swap the memory pair for other plausible US-accessible memory or storage names. Re-compute the 7-name basket from Jan 18, 2025 to today.

Memory pairPair mean7-name basketBasket / SMHReading
SNDK + STX (NAND / storage)+2363.59%+763.97%~6.4×would have beaten our pick
WDC + MU (our actual pick)+721.71%+294.86%~2.5×our basket
INTC + MU (broader semi)+519.81%+237.18%~2.0×still beats sector
SMH + SMH (passive, no name picking)+120.33%+123.04%~1.0×no theme = no excess

WDC + MU was not a lucky pick.

Any reasonable US-listed memory mapping during this window produced a basket that beat SMH by a wide margin. SanDisk + Seagate would have produced ~6.4× SMH. Intel + Micron ~2.0× SMH. The only mapping that failed to beat the sector was the passive one — i.e., NOT identifying that memory was the right theme.

What this means for the methodology. The skill jukan05 demonstrated was not “pick WDC vs MU.” It was “recognize and persistently signal that Korean memory was tightening ahead of consensus.” Once the theme is correctly identified, multiple reasonable ticker mappings capture the move. This is a stronger methodological story than “we happened to pick the two winners” — the edge is theme identification, not stock picking.

The caveat that follows. This does not mean any non-memory narrative will work the same way. MEMORY_BULL is the only topic with strong sector-controlled alpha in this 16-month sample. The forward portfolio rests on jukan05's ability to identify the next theme correctly. Our job is to map his fresh narratives to US tickers, but if the underlying themes don't pan out, no mapping will save it.

Part 2 — Forward-Looking Niche Portfolio The methodology works. The next question is which themes are NOT yet priced in.

Why We Do NOT Repeat the Memory Trade Now Four signals say the canonical edge has played out. Harvest the lesson, redeploy the methodology elsewhere.

Signal #1 — Author self-awareness.

jukan05 himself recently wrote (paraphrased): “When I posted 100 bullish posts on memory, nobody said anything. But the moment I post just three cautious ones, people start saying 'jukan is bearish on memory now, we're done.'” This is the classic retail-saturation tell. When the audience is so heavily positioned that any neutral comment provokes panic, the easy money is already gone.

Signal #2 — Magnitude of move already delivered.

WDC +854% in 16 months. MU +589%. Korean DRAM export prices +498% YoY (May 2026 data). NAND +280% YoY. SSD +344% YoY. SanDisk has risen roughly 50× in 12 months. These are late-cycle confirmations, not early-cycle openings.

Signal #3 — Backtest decay.

Even within the historical sample, the most recent MEMORY_BULL signals produce smaller marginal moves because each new tweet adds less to a market that has already fully absorbed the thesis.

Signal #4 — The largest end-customer is designing AROUND memory.

NVIDIA is reportedly reducing system DDR memory in Vera platforms because memory costs have become an “excessive share” of bill-of-materials. When the largest end-customer in compute begins design optimizations away from a component, that component is no longer scarce in the long run.

Conclusion. The canonical memory edge is real and proven. The trade itself is now late in its lifecycle. Harvest the lesson (supply-chain narratives precede price moves), discard the position (the trade has run), and redeploy the methodology to themes that are not yet priced in.

Fresh Narratives in the Past 14-30 Days The signals that have surfaced recently and not yet been absorbed by the broader tape. The EVIDENCE column now distinguishes DIRECT (independent cashtag), DIRECT-reactive (cashtag during a price spike), inference, and analyst-report shares.

Theme@jukan05 context (paraphrased)US tickerEvidence
Intel come-back (multi-catalyst)“UBS: Rubin Ultra 4-chip uses Intel's EMIB-T. $INTC.” “DJT told Apple to use Intel.” “Intel threatening suppliers — won't supply CPUs unless they use 18A.” Plus Vera CPU challenger + SK Hynix EMIB partnership — five distinct DIRECT cashtags across separate catalysts.INTCDIRECT ×5
Agentic AI CPU servers (reactive)“$DELL has significant growth potential thanks to CPUs. Agentic AI CPU servers ... higher ASP.” But the tweet was written AFTER the stock had already moved on a Trump remark about Dell computers (“buy a Dell”) that was misinterpreted as a stock endorsement and pumped DELL 12% before fading.DELLDIRECT-reactive
PCB rally (NVIDIA AI)“TTMI up in pre-market because PCB names rallied in China today; the reason is Morgan Stanley's analysis on NVIDIA's next-gen platform PCB content.”TTMIinference
Qualcomm ASIC“Qualcomm provides ASIC services to ByteDance and Amazon. Anthropic could use Maia, but maybe Qualcomm builds the chip.”QCOMinference
Optical interconnects“Breaking the Memory Wall: Optical Interconnects Emerge in GPU-HBM Packaging.” Decoupling GPU and HBM via optical fabric.COHRinference
WFE shift (3D DRAM)“The moment DRAM Big Three transitions to 3D DRAM, DRAM will no longer require EUV, and that surplus EUV capacity could quickly shift to logic.”LRCX, AMATinference
AMD Helios 2H26“AMD accelerating Helios rack-scale AI platform, full-scale mass production from 2H26.”AMDinference
AI compute power“WSJ: Anthropic expected to turn profitable in Q2. AI's growth trajectory is insane, and this is not cash-burning growth.” (Implicit: persistent grid / power demand.)CEG, VSTinference
CoreWeave AI infra“GF Overseas — CoreWeave Initiation: Riding the AI Infrastructure Supercycle. Buy. US$162 target.”CWAVEanalyst report

Evidence legend (refined). DIRECT ×5 — cashtag + thesis posted BEFORE or INDEPENDENT of a price spike, across multiple catalysts (INTC = 17% top-conviction weight). DIRECT-reactive — cashtag + thesis posted DURING / AFTER a price spike caused by external news; informative but late-entry (DELL = 8%, downgraded from 18%). inference — our narrative-to-ticker mapping (8 names = 67%). analyst report — he forwarded a sell-side report without endorsing himself (CWAVE = 3%). Conviction weighting now reflects this distinction: DIRECT ×5 gets the heaviest sizing, DIRECT-reactive is sized as if it were inference, and the rest follow.

Why DELL was downgraded from 18% to 8%. jukan05's DELL post is a textbook DIRECT-reactive case: the stock had already pumped ~12% on a Trump remark about Dell computers (“you should buy a Dell”) that retail readers misinterpreted as a Dell stock endorsement. jukan05's tweet came AFTER that move, with a credible-but-secondary thesis about agentic AI CPU server ASP. The thesis is plausible; the entry timing is compromised. We keep the position but at less than half its prior weight, and we relabel the evidence so future versions of the framework treat reactive posts as something closer to inference than to true author-initiated signal.

Themes we deliberately skip. NVDA — backtest alpha −1.65%, latest jukan05 post slightly bearish. MU / WDC — memory rally already delivered, author shows saturation tells. SNDK — ~50× in 12 months, consensus crowded. TSM — EUV bottleneck fully priced (alpha +2.44% marginal). AVGO — TPU / ASIC story widely known. AAPL — Trump-Intel push directionally unclear.

Proposed Niche Portfolio — 12 Positions Same supply-chain-to-ticker methodology, applied to fresh themes only. Niche bias (mid-cap and below), staggered entry, hard risk controls.

#TickerWeightThesisEvidenceVolatility
1INTC17%18A + EMIB-T + Apple/Trump + Vera CPU challenger + SK Hynix EMIB — five distinct cashtags across separate catalysts. Top conviction.DIRECT ×5Very high
2QCOM14%ASIC services for ByteDance + Amazon; possible Anthropic Maia alternativeinferenceMedium
3TTMI12%PCB content in NVIDIA AI server platforms (MS report-driven)inferenceHigh
4COHR10%Optical interconnect / memory-wall solution in GPU-HBM packaginginferenceHigh
5LRCX10%3D DRAM transition + EUV-to-logic capacity shiftinferenceMedium
6DELL8%Agentic AI CPU servers thesis — but the post was REACTIVE to a Trump remark about Dell computers that pumped the stock 12%. Plausible thesis, late entry. Downgraded from 18% to 8%.DIRECT-reactiveHigh
7AMAT8%Memory + foundry capex (DRAM expansion phase)inferenceMedium
8CEG7%AI compute power (Anthropic profitability ⇒ persistent grid demand)inferenceLow
9AMD5%Helios rack-scale AI platform, full mass production 2H 2026inferenceVery high
10VST4%AI data center power, Texas grid exposureinferenceMedium
11CWAVE3%Pure-play AI infrastructure (CoreWeave; jukan05 shared a sell-side Buy initiation, $162 target)analyst reportVery high
12Cash2%Dry powder / reserve

Evidence concentration. Of 11 active positions, only 1 (INTC = 17%) rests on multiple DIRECT cashtags across distinct catalysts — the highest-conviction signal. 1 position (DELL = 8%) is DIRECT-reactive (cashtag during a Trump-induced pump, downgraded from 18%). 8 positions = 67% are our inference mappings of his narrative topics. 1 position (CWAVE = 3%) is a forwarded sell-side analyst report. Sizing now reflects this finer evidence hierarchy: DIRECT ×5 > inference > DIRECT-reactive (de-rated for timing) > analyst-report share.

Track this portfolio live on fxcryptobots Paper Trading.

We have seeded the 12-position allocation above as a public paper-trading portfolio so you can watch it print real-time alongside our other mirror portfolios — no real capital, no recommendation, just the live equity curve of OUR construction from jukan05's narratives. CoreWeave is represented by its actual NASDAQ ticker CRWV (the report uses “CWAVE” as shorthand; CRWV is the tradeable instrument).

View “Jukan Niche Supply-Chain” portfolio →

Theme Breakdown

36%
AI compute middleware
QCOM · TTMI · COHR
30%
Foundry / Agentic
INTC · DELL · AMD
18%
Wafer-fab equipment
LRCX · AMAT
11%
AI power infrastructure
CEG · VST
3%
AI infra speculative
CWAVE
2%
Cash reserve
Dry powder

Realistic Forward Expectations The historical headline was a once-per-cycle outcome. The forward portfolio should be sized for the base case, not the past.

We explicitly do NOT project a repeat of the historical outcome.

The historical headline was a basket +295% vs NASDAQ +37% over 16 months, but that was concentrated in WDC + MU. Without those two names the 5-name remainder roughly matched SMH (~+120%). The actual signal-level alpha vs the sector is moderate, not extraordinary. That outcome required a once-per-cycle memory rerate; we cannot count on the new theme list producing the same magnitude.

+ Best case

Portfolio meaningfully beats SMH. Modest sector-controlled alpha if 2-3 of the fresh themes (Intel come-back, agentic AI servers, optical interconnect) materialize on the expected catalyst timing.

! Base case

Portfolio tracks SMH (~+15-25% per year if AI capex stays intact). The 60-day signal-level alpha (+5.82% vs SMH) is the most representative number for steady-state expectation.

× Bear case

Portfolio underperforms QQQ. Triggered by AI-capex slowdown, or by one of the high-volatility positions (INTC, AMD, CWAVE) failing on its specific catalyst.

Management Rules & Risk Controls Niche names need niche controls. Staggered entry, hard stops, monthly rebalancing, explicit skip and contrarian rules.

Entry (staggered).

Day 0 — 50% initial allocation in each position.
+2 weeks — 30% add, confirmed by fresh signal + price action.
+1 month — 20% dip reserve (drawdown or new catalyst).

Hold horizon.

Target hold 6-12 months. Fresh narratives have not been fully priced; patience is required. Do not exit during earnings or catalyst dates — re-evaluate after.

Stop / exit.

Per-position stop: −18% (niche names are volatile).
Portfolio daily limit: −5% → raise cash, no new buys.
Portfolio weekly limit: −10% → move to 50% cash, full review.
Profit-take on a single name: +50% → sell 1/3, trail the rest.

Re-balance.

Monthly review (day 30, 60, ...). Scan recent tweet flow; if a new theme emerges, downsize the weakest tail (CWAVE first) and rotate. If any theme weight exceeds 15%, take partial profits.

Skip rule (from backtest).

Do NOT open new positions on NVDA-bullish, AVGO-bullish, or TSM-tightness posts — these themes show no measurable alpha after sector control. Do NOT open positions on Apple-related news (directional bias unclear).

Contrarian rule (from BEAR-call finding).

If the author's tone turns repeatedly cautious on memory: this portfolio holds no MU / SNDK / WDC, so no action is required. Do NOT chase memory longs based on the contrarian read alone — the portfolio is already designed to avoid the saturated memory trade.

Risk notes by position.

1. INTC is the highest-conviction and highest-volatility position at 17%. If the Apple/Intel or Rubin Ultra EMIB-T decisions do not materialize, expect rapid drawdown.
2. DELL was downgraded from 18% to 8% because the post was reactive to a Trump remark that pumped the stock 12% before the tweet. Plausible thesis on agentic CPU server ASP, but entry timing is compromised.
3. TTMI has already pulled momentum over the past 6 months. Staggered entry is critical to avoid late-entry damage.
4. CEG / VST AI-power consensus has begun forming — weights kept low.
5. DELL + TTMI + COHR + QCOM are highly correlated (~0.7) as “AI infra middleware.” They will draw down together on any AI-capex pullback.
6. CWAVE is a high-volatility speculative position; short trading sample.
7. AMD Helios thesis is fresh and not yet directly backtested — small weight is intentional.

Non-US tickers excluded (mentioned by jukan05 but skipped here).

Murata (6981.T) — JP, low-liquidity ADR (MRAAY). Samsung Electro-Mechanics — Korea-only. Kioxia, Socionext — JP-only. Nittobo, Mirae Co — JP/KR tiny float. MediaTek (2454.TW) — Taiwan-only. CXMT, YMTC — China IPO pipeline. Most US retail brokers (including Tickmill CFD) do not carry these.

How This Portfolio Was Derived — Three-Step Decision Chain

Step 1

Does jukan05 have a measurable edge? MODERATELY

  • 60-day mean alpha vs SMH: +5.82% (n=294, WR 57%) — the most representative number for steady-state use
  • 180-day mean alpha vs SMH: +31.85% (n=175, WR 68%) — positive but fat-tailed; median is far smaller than mean
  • Direction is positive and statistically meaningful at every horizon — magnitude is moderate, not extraordinary
  • The 16-month basket buy-and-hold beat NASDAQ by 8× mostly because of sector beta + a concentrated memory rally in two names
Step 2

Where is the edge concentrated? ONE TOPIC, EXPRESSED THROUGH THEME ID (NOT STOCK PICKING)

  • MEMORY_BULL is the only topic with strong sector-controlled alpha: +26.44% 60d, WR 84.6% (n=78)
  • The basket excess comes through WDC + MU, but the counterfactual analysis above shows that any reasonable memory mapping (SNDK+STX, INTC+MU) would have produced a similar or better outcome
  • The skill reduces to: “early Korean memory supply-chain intel — identifying the THEME ahead of consensus.” The specific ticker choice mattered less than recognizing the theme was real
  • NVIDIA / Broadcom / TSMC mentions show no measurable alpha — over-mentioned, already priced
  • Bearish calls (small n=31): inverted direction — contrarian indicator at the topic level
Step 3

What do we do TODAY? APPLY THE METHODOLOGY TO FRESH THEMES, USE A REFINED EVIDENCE HIERARCHY

  • The canonical memory trade has delivered its return; we do NOT repeat it (WDC +854%, MU +589%, retail saturation visible)
  • We apply the same methodology (narrative → ticker map) to newer themes from the past 14-30 days
  • Evidence hierarchy refined: INTC at 17% (five distinct DIRECT cashtags, top conviction); DELL at 8% (DIRECT-reactive: downgraded from 18% because the post came during a Trump-induced pump); 8 inference-mapped names at 67%; CWAVE 3% analyst-report share; 2% cash
  • Forward expectation is calibrated: base case is “tracks SMH,” not “repeats 8× NASDAQ”

What Would Kill This Thesis

× Memory glut hits early

If oversupply lands sooner than narrative suggests, downside for AMAT / LRCX (capex spending dries up first).

× Apple opts NOT to use Intel

Rubin Ultra picks alternative packaging or Apple stays on TSMC — INTC thesis loses its biggest single catalyst.

× Agentic AI server demand fails

If the CPU-server ASP uplift does not materialize at expected magnitude, downside for DELL.

! Macro AI-capex deleveraging

Broad pullback across all AI-infra positions simultaneously. Daily and weekly portfolio stops trigger first.

Limitations & Honest Caveats

  1. Topic classifier is heuristic. BULL/BEAR and topic assignment is regex + keyword based, validated on a 50-tweet random subsample (~92% agreement with manual labels). The 8% misclassification noise should not flip the headline direction but could shift magnitudes by 1-3 alpha points at any horizon.
  2. Buy-and-hold is perfect-hindsight. The +294.86% basket return assumes you held all seven names from day one. Real-world signal-based entries would underperform that ceiling. More importantly, the SMH comparison is the fair benchmark — +175 percentage points over SMH (~2.5×), not the flattering 8×-NASDAQ framing. The counterfactual mapping section above confirms the result is robust to ticker choice within the memory theme, but the magnitude of the historical outcome was a once-per-cycle event we do not project forward.
  3. BEAR sample is small (n=31). The inverted-direction conclusion has wide confidence bounds. Treat the contrarian rule as a tilt, not a primary trade.
  4. Account is 20 months old. The 16-month backtest window covers ~80% of the account's existence. Out-of-sample is limited; the canonical edge may attenuate as more capital chases the same topics.
  5. Forward portfolio is methodological, not signal-confirmed. Each position rests on a recent narrative thread. We have no per-signal alpha numbers yet for DELL, INTC come-back, TTMI, QCOM ASIC, COHR optical, CEG / VST AI-power, AMD Helios, or CWAVE — only on the legacy memory and semi-capex names. The methodology (supply-chain narrative → ticker map → 6-12 month hold) is what is proven; the specific picks are an application of that methodology to fresh themes.
  6. Single-signal topics (AMD_BULL, APPLE_WEAK) are anecdotal. n=1 cells in the per-topic table are flagged; the portfolio's AMD position is sized small accordingly.
  7. SMH alpha vs equal-weight semi index. SMH is cap-weighted, so it tilts toward NVDA / TSM. An equal-weight semi index would shift alpha numbers by 1-3 points but not change direction.
  8. Entry assumes next-session open. Live followers entering intraday pay a worse fill on average; reported alpha is the model ceiling.
  9. Excluded Non-US names matter. Murata, Samsung Electro-Mechanics, Kioxia, MediaTek and Chinese memory IPO names cover real exposure that this portfolio cannot capture. US-listed-only investors get the practical version; a global investor would augment.

Final Verdict

+ Methodology is moderately proven

Sector-controlled (vs SMH) signal-level alpha: +5.82% at 60d, +31.85% at 180d. Counterfactual analysis confirms the skill is theme identification, not specific stock picking — any reasonable memory mapping beat SMH during the window.

! Canonical memory trade is done

WDC +854%, MU +589%, Korean DRAM prices +498% YoY. Author shows audience-saturation tells; NVIDIA designing around memory cost. Harvest the lesson, exit the position.

! Forward portfolio with refined evidence hierarchy

12 positions: INTC 17% (DIRECT ×5 top conviction), DELL 8% (DIRECT-reactive, downgraded from 18%), 8 inference mappings (67%), CWAVE 3% (analyst report), 2% cash. Base case: tracks SMH. We explicitly do NOT project a repeat of the 8× NASDAQ historical outcome.

Full 12-page consolidated PDF report Cover & contents, profile + methodology, signal inventory, horizon-based alpha, per-topic and per-ticker edge map, memory deep-dive, 16-month buy-and-hold table, equity-curve chart, “why we don't repeat the memory trade”, fresh narratives table, 12-position portfolio table with theme breakdown, full management rules, and the decision chain.
Download PDF ↓

Report generated 2026-05-23 by the fxcryptobots research desk (v2, rationalized). Source: @jukan05 public X posts (16,277 lifetime, sample window Jan 18 2025 – May 22 2026); regex-based topic classifier validated on a 50-tweet manual sample (~92% agreement). Topic-to-ticker mapping is fixed and disclosed in the methodology table. Price data: Yahoo Finance daily bars (auto-adjusted close). Entry: next trading day RTH open. Alpha: ticker return − benchmark return over the same window. SMH (VanEck Semiconductor ETF) is the sector-control benchmark. Counterfactual analysis runs the 16-month buy-and-hold with alternative memory pairs (SNDK+STX, INTC+MU, passive SMH) holding the other 5 names constant. Forward-looking portfolio is OUR mapping of his supply-chain narratives to US-listed tickers, with explicit EVIDENCE labels separating DIRECT cashtags from inferences. The forward portfolio is NOT a recommendation from jukan05 and is sized for a calibrated base case (tracks SMH), not a repeat of the historical outcome. This is research and educational analysis, not investment advice; see our risk disclaimer.