Research Report · Twitter Alpha Backtest · Broad AI / Semis

@ParadisLabs 70-Day Backtest: BULL Signals Returned +55.58% Mean at 30 Days, Win Rate 97%, +28.21% Alpha vs XLK

Quantitative backtest of @ParadisLabs (Paradis Labs, 45,631 followers, bio: “AI / Semiconductor Analyst & Trader. Thematic investment research.”). 353 original tweets across the 70-day window, 231 ticker-mention tweets, sentiment mix BULL 103 / BEAR 27 / NEUTRAL 101. Universe: AI / semis thematic basket with high-conviction supply-chain mapping.
Published 2026-05-25 Backtest Window: Mar 15 -> May 24, 2026 (70 days) Benchmarks: SPY (broad) / XLK (tech sector) Entry: T+1 open · Exits: 5d / 14d / 30d / 90d close
VERDICT: STRONG ALPHA · FOLLOW (broad AI / semis, regime-bounded)
+18.66%
BULL mean / 14d (n=122)
+55.58%
BULL mean / 30d (n=37)
97.30%
BULL win rate / 30d
+28.21%
BULL alpha vs XLK / 30d
353
Original tweets sampled (70d)
231
Tickered tweets · 103 BULL
45,631
Followers · 1,923 lifetime tweets
70d
Account is <3 months old (no persistence data)
Prefer the print version? Download the 3-page PDF: profile dossier, full sentiment-by-horizon table, most-mentioned tickers, best/worst 30-day BULL calls with tweet previews, methodology, limitations.
Download PDF ↓

Profile & Persona Who Paradis Labs is and what kind of signal the timeline produces.

The author. 45,631 followers, account opened 2026-03-15 (less than 3 months old), 1,923 lifetime tweets, bio: “AI / Semiconductor Analyst & Trader. Thematic investment research.” The output style is high-conviction supply-chain mapping — long-form theses connecting upstream component plays (substrates, lasers, optical interconnects) to downstream AI-capex demand. Top names across the 70-day window: $NVDA (36 mentions), $TSM, $AAOI, $IQE, $SIVE, $SNDK, $LITE, $NBIS, $AXTI, $MU, $MRVL, $META, $AEHR, $QCOM, $COHR, $INTC, $AVGO, $GOOGL, $AMZN, $MSFT.

Signal shape. Mix of mega-cap thesis-anchor names ($NVDA, $TSM, $META, $QCOM) and mid-cap supply-chain plays ($AAOI, $SIVE, $IQE, $AEHR). The differentiator vs broad-tech commentary is the explicit mapping from end-demand to component supplier.

Forward Returns by Sentiment & Horizon Entry: next trading day open after tweet date. Exit: close at +5/+14/+30/+90 trading days. Alpha = signal return minus same-window benchmark return.

SentimentHorizonnMean RetMedianWin RateBestWorstAlpha SPYAlpha XLK
BULL5d160+6.88%+4.47%67.50%+50.51%-25.30%+4.86%+2.37%
BULL14d122+18.66%+13.93%78.69%+183.08%-38.45%+13.59%+6.19%
BULL30d37+55.58%+29.87%97.30%+270.77%-0.16%+44.11%+28.21%
BEAR5d43+9.11%+7.67%72.09%+53.08%-21.18%+6.40%+4.28%
BEAR14d32+28.38%+28.73%90.62%+73.70%-36.81%+21.35%+13.53%
BEAR30d14+67.22%+46.67%92.86%+157.37%-3.60%+52.43%+33.37%
NEUTRAL5d173+5.88%+3.33%63.01%+58.19%-19.93%+4.03%+1.77%
NEUTRAL14d110+17.95%+12.05%76.36%+111.72%-19.83%+12.46%+5.16%
NEUTRAL30d43+41.44%+31.98%88.37%+164.36%-35.42%+29.19%+13.75%

The headline read. The 14-day horizon (n=122 BULL signals) and 30-day horizon (n=37) are both sized large enough to read seriously inside this window. Mean returns +18.66% / 14d and +55.58% / 30d; win rates 78.7% and 97.3%. Alpha vs XLK is +6.19% / 14d and +28.21% / 30d — meaningful even against the same-period tech-sector ETF (the most demanding benchmark).

The BEAR bucket also produced positive returns in this regime: most BEAR-tagged tweets were ticker-mention posts where the directional language was secondary to the underlying name being inside the AI-rally tape. Read the BULL numbers as the primary edge.

Most-Mentioned Tickers (70-Day Sample) Where the conviction concentrated. Top 20 by mention count.

#TickerMentions#TickerMentions
1$NVDA3611$MRVL13
2$TSM2012$META13
3$AAOI2013$AEHR13
4$IQE1914$QCOM11
5$SIVE1915$COHR11
6$SNDK1716$INTC11
7$LITE1617$AVGO10
8$NBIS1618$GOOGL10
9$AXTI1519$AMZN9
10$MU1420$MSFT8

Best BULL Calls (30-Day Forward Return) Top 10 resolved 30-day signals.

DateTickerRet 30dTweet preview
2026-03-29$SIVEF+270.77%“$SIVE / $SIVEF Among small-cap photonics names, $SIVE offers arguably the most skewed R:R...”
2026-03-30$SNDK+146.12%“$SNDK Gross margin: 29.9% -> 51.1% in a single quarter. Guide for next quarter: 65-67%.”
2026-03-26$MU+121.01%“$MU buying back $5.4B in senior notes due 2031-35, settling April 3. That signals management...”
2026-03-15$AEHR+116.92%“The most asymmetric AI supply chain setup atm is $AEHR. A $1B company with a literal monopoly...”
2026-04-05$AXTI+114.46%“Finishing up my research on a couple of companies in crystal growth & substrate manufacturing...”
2026-04-05$INTC+112.60%“$AMD are executing strong. Their Zen 6 leak points to a huge architectural leap that fundamentally...”
2026-03-29$MU+111.38%“People think we're at peak memory demand? Lol. SK Hynix buying 2 yrs of competitive moat...”
2026-04-09$AMD+95.61%“I'm very bullish on Taiwan. Following on from my Browave thesis...”
2026-04-05$AMD+91.99%“$AMD are executing strong. Their Zen 6 leak points to a huge architectural leap...”
2026-04-04$SNDK+85.05%“$NVDA's Neural Texture Compression cuts VRAM usage by up to 90%. I bet markets will fear...”

Worst BULL Calls (30-Day Forward Return) Bottom 10. Note: in a strong tech tape every resolved 30-day BULL signal was still positive or near-flat.

DateTickerRet 30dTweet preview
2026-04-09$GOOG+19.22%“I'm very bullish on Taiwan. Following on from my Browave thesis...”
2026-04-09$AAPL+18.79%“I'm very bullish on Taiwan. Following on from my Browave thesis...”
2026-04-07$AVGO+18.77%“$AVGO up >5% off the back of the news that they've signed multiyear agreements with $GOOGL...”
2026-04-09$NVDA+16.83%“I'm very bullish on Taiwan...”
2026-04-09$AVGO+14.74%“I'm very bullish on Taiwan...”
2026-04-09$SPIR+12.25%“$SPIR down 10% on the news of a $70M ATM raise. Quite insignificant imo - ~$700M MC...”
2026-04-09$TSM+7.70%“I'm very bullish on Taiwan...”
2026-04-07$SPCX+2.74%“Matt Tuttle, you have my respect. His ETF ticker changed from $SPCX to $SPCK today...”
2026-04-07$SPCK+2.74%“Matt Tuttle, you have my respect. His ETF ticker changed from $SPCX to $SPCK today...”
2026-04-07$META-0.16%“Matt Tuttle, you have my respect. His ETF ticker changed from $SPCX to $SPCK today...”

Five Questions, Five Answers

Q1. Is there a real edge inside the sample window?

Yes. BULL alpha vs XLK is +6.19% / 14d (n=122) and +28.21% / 30d (n=37). Win rate 97% on the 30-day bucket is unusually high — every single resolved 30-day BULL signal except one was positive (or zero).

Q2. Will the edge survive a regime change?

Unknown. The entire 70-day sample falls inside the March-May 2026 AI rally. The account itself was only created 2026-03-15 (<3 months old), so we have no out-of-sample data at all — not in this lookback, not in any earlier regime. The +97% WR / 30d is almost certainly upper-bounded by the bull-tape regime.

Q3. What is the differentiator vs broad-tech commentary?

Supply-chain mapping. The thesis posts explicitly trace upstream component plays ($AAOI optical interconnect, $AEHR substrate test, $IQE / $AXTI substrate, $SNDK / $MU memory) to downstream AI-capex demand. This is harder to replicate from a generic mega-cap feed.

Q4. Where is the biggest measurement risk?

Sample size on the 30-day bucket (n=37) and the entire account history being <3 months. The biggest single winner ($SIVEF +270%) is a Stockholm listing whose cashtag does not always resolve in US pricing — included here but flagged as outlier-driver. Drop that one signal and 30d mean drops noticeably.

Q5. How should I use this feed?

As a high-quality thematic input. Use the supply-chain mapping to discover mid-cap component plays you would otherwise miss, then size moderate and time-bounded (≤30d). Wait for one corrective tape before sizing aggressively — we have no bear-regime data to anchor on yet.

Methodology

Data source. User-timeline endpoint, original tweets only (replies and pure retweets excluded). 353 tweets across 70 days — the maximum the timeline endpoint surfaces for this account at its posting cadence.

Ticker extraction. $TICKER cashtag regex with a curated non-ticker stop-list (e.g. $AI, $US, $YTD removed).

Sentiment classifier. Regex keyword lexicon (BULL: bullish, buy, long, added, breakout, ATH, target, runner; BEAR: bearish, short, sell, sold, dump, dilution, scam). BULL/BEAR if count strictly dominates; otherwise NEUTRAL.

Backtest mechanics. Entry next trading day open (T+1). Exits at +5d / +14d / +30d / +90d close. No transaction costs / slippage. Multiple mentions of the same ticker on the same day produce separate signals (no deduplication) — this reflects a literal “follow every tweet” baseline.

Benchmarks. SPY (broad market) and XLK (tech sector). XLK is the more demanding benchmark since virtually all signals are in AI / semis.

Limitations & Honest Caveats

  1. Account is <3 months old. Created 2026-03-15. The 70-day backtest is the entire account lifetime. We cannot evaluate persistence beyond this run.
  2. Regime bias. The entire sample falls inside the March-May 2026 AI rally (SPY +6%, XLK +12%). Stock-picking edge is inseparable from beta on a tape this strong.
  3. 30d sample is moderate (n=37). The +97% win rate is unusually high; expect mean-reversion toward the 14d bucket (n=122, WR 79%) as more signals resolve.
  4. One foreign-listing outlier. The biggest 30-day winner ($SIVEF +270%) is a Stockholm-listed laser play; pricing resolution varies by data provider. The published 30d mean has notable single-name sensitivity.
  5. BEAR / NEUTRAL buckets read positive. Inside an AI rally, most ticker-mention tweets resolve up regardless of directional tag. The BULL row is the cleanest read on directional skill.
  6. No risk-adjustment. Sharpe / Sortino not computed. Position-sizing, drawdowns, and correlation across signals are not modeled.

Final Verdict

+ Tier: STRONG ALPHA · FOLLOW

BULL signals returned +18.66% mean / 14d (n=122, WR 79%) and +55.58% / 30d (n=37, WR 97%) with +6.19% and +28.21% alpha vs XLK respectively. Supply-chain mapping is a real differentiator.

! Caveat: no out-of-sample data

The account is <3 months old. The 70-day window is the entire history. The +97% / 30d WR is regime-bounded — expect normalization. No bear-tape data to anchor sizing against.

+ How to use

High-quality thematic input on AI / semis with explicit supply-chain mapping. Discover mid-cap component plays through the feed; size moderate and time-bounded (≤30d). Wait for one corrective tape before sizing aggressively.

Full 3-page PDF report Profile dossier, full sentiment-by-horizon table, top-20 most-mentioned tickers, best / worst 30-day BULL calls with tweet previews, methodology and limitations.
Download PDF ↓

Report generated 2026-05-25 by the fxcryptobots research desk. Source: @ParadisLabs public X timeline (user-timeline endpoint, original tweets only); price data from Yahoo Finance daily bars. Sentiment classifier is regex-based. Per-name returns are total return including dividends via auto-adjusted close. This is research and educational analysis, not investment advice; see our risk disclaimer.