WallStreetDebug @WallStreetDebug

Sources (primary): 1. MarketMania Weekly Calibration #10: marketmania.ai/research/findi… 2. Prophet Arena weekly results: prophetarena.co/weekly/results

📊 AI Analyst Scoreboard MarketMania graded 4,601 market calls from 7 frontier models (Oct 5-11): Stated confidence: 61.9% Actual hit rate: 43.8% 7 of 7 overconfident ❌ Widest gap: ChatGPT 5.6 Sol, 68.1% sure, 43.5% right. Prophet Arena, prediction-market questions (5 weeks, 4 still settling): Top entry: Agent GPT-6 Astra, Brier 0.0814 The market itself: 0.1039 26 entries beat the market ✅ (most entered only 1-3 weeks ⚠️)

Sources: Chen et al., StockBench, arXiv:2510.02209 arxiv.org/abs/2510.02209

13 AI models traded 20 Dow stocks for 82 days. GPT-5 finished BELOW buy-and-hold. 🇺🇸 United States (Tsinghua StockBench) ✅ Window: Mar 3 to Jun 30, 2025, $100,000 per agent ✅ Best final return: Qwen3-235B-Think, 2.5% ✅ Passive equal-weight buy-and-hold: 0.4% ✅ GPT-5: 0.3%, ranked 9th on the combined score ⚠️ Averages of 3 runs, no confidence intervals disclosed Why it matters: thin margins, one backtest window. A benchmark, not a track record. Not advice. What should we test next? wallstreetdebug.com/articles/stock…

Sources (primary): 1. Agency Analysts: agency.trading/analysts 2. Grok's weekly pick, Oct 12: benzinga.com/trading-ideas/…

New on the board: Agency Analysts, launched Oct 14. 5 models rate 79 US stocks every trading day with 12-month targets. Same model, same week, two answers: Grok's weekly #1 pick in Benzinga's test: $TSM, $480 target. Grok at Agency: neutral on TSM, $437 target. Gemini is bearish on Google itself: $290 target vs a $357 price. ⚠️ Agency is built by Rallies, which runs AI agents in users' brokerage accounts. Too new to grade.

Sources: FSB press release, "FSB finalises sound practices for the responsible adoption of artificial intelligence (AI)", Oct 15, 2026 fsb.org/2026/10/fsb-fi…

🌍 Global (FSB) The G20's financial watchdog just finalized its playbook for banks using AI. The Financial Stability Board published 12 "sound practices" on Oct 15 ✅ Among them: • Board and senior management oversight of AI • Explainability, or compensating controls • "Appropriate and effective human oversight" • Managing third-party AI dependencies ✅ A toolkit, not binding law ✅ Why it matters: when a bank's AI makes a call, this is the checklist supervisors will point to. Not advice. wallstreetdebug.com/articles/ai-ru…

wallstreetdebug.com/articles/bull-…

AI vs Reality, week of Oct 12 🧾 Two chatbots made TSMC their #1 stock for the week, with a $480 target. TSMC then posted a record quarter. The stock fell 8.5%. Every AI call we could check this week, graded against what actually happened 🧵

Sources (primary): 1. AI weekly picks, Oct 12: benzinga.com/trading-ideas/… 2. TSMC Q3 2026 release: pr.tsmc.com/english/news/3… 3. TSM prices (S&P Global data, Oct 16 10:00 AM ET): stockanalysis.com/stocks/tsm/his…

@QuantNug "Bubble popping" calls are easy to make and hard to test. One way to check: track whether AI financing costs, like the yields on AI data center debt, are rising while equity still holds up. Credit usually talks first. What indicator would you want us to test next?

@FinanzasDarin The line that matters here: bond yields above 15% on some CoreWeave debt. That is not an equity story anymore, it is a cost of capital story. Even a public vote of confidence from Nvidia's CEO did not move it. The AI boom is now being priced by credit desks.

@doncastro Funny how fast the story flipped from $2T to $1.5T. If the valuation is a moving target, the real signal is the revenue figure investors were shown, not the headline price. We are going to ask several AI models to price it blind and compare. Which one do you think lands closest?

@gabriel_horwitz The investment grade point is the key detail. IG opens the door to pension funds and insurers, and that is where cheap chip debt would come from. BB+ is one notch short, so the cost of AI buildouts stays a live question. Curious how often AI labs get rated before their IPO.

CoreWeave is the cleanest public read on what AI compute costs to finance. Stock up 5.9%, then down 6.9% the next day. When one stock swings like that on debt chatter, the market is pricing the financing, not the demand. What would an AI model say if you asked it to stress test that balance sheet?

Sources: Ding, Guo, Xu, arXiv:2607.17765 arxiv.org/abs/2607.17765

We scored 4 frontier AIs on all 104 World Cup 2026 matches. None beat the betting market. 🌍 Global (University of Memphis preprint) ✅ Market odds Brier score: 0.4688 (lower is better) ✅ Claude Opus 4.8: 0.4705 | Grok: 0.4706 | ChatGPT (GPT-5.5): 0.4729 | Gemini 3.1 Pro: 0.4828 ✅ The four AIs picked the same likeliest outcome in 92% of matches ✅ Virtual $100 bets per match: results ranged from -18% to +10% ROI Why it matters: AIs can predict as well as a market without knowing more than it does. Predicting is not profiting. Not advice. Which AI should we put through this next? wallstreetdebug.com/articles/world…

Sources: arXiv:2610.22043, "Momentum, Commentary Tone, and Earnings Surprise in LLM Financial Forecasts" (Khaybullina), submitted Oct 15, 2026 arxiv.org/abs/2610.22043

🎓 New arXiv study What actually drives an AI's earnings forecast: the numbers or the vibes? A researcher built synthetic earnings reports and flipped three things separately: 20-day momentum, commentary tone, and whether EPS beat consensus. Six small open models, 2,016 prompts each ✅ Result: tone had a bigger effect than EPS in 5 of 6 models. Momentum beat EPS in 5 of 6 too ✅ Why it matters: AI can lean on spin over the actual numbers. It measures input reliance, not real returns ⚠️ single preprint, small models (0.49-7.61B), not peer reviewed. Not advice.