Market timing — Red Team review

Market timing — Red Team review

Background

• Spun out of Market-timing-refined-analysis (round 6, 2026-09-25) at Talbot’s request. It was called “Devil’s Advocate” until then; renamed to Red Team (Talbot, round 5).
• A second, independent model attacks the research before any of it leads a public report or the Market Drop Wins flagship. It must be a different provider (Gemini) so its blind spots differ from Claude’s.
• Project + data: /home/ta/projects/cape-timing/README.md · \\wsl$\Ubuntu-24.04\home\ta\projects\cape-timing\README.md
• Reports under test: /home/ta/artifacts/cape-timing-report.html and /home/ta/artifacts/valuation-trend-rule.html.

Tasks

1. Talbot: open https://aistudio.google.com, choose the strongest Gemini model, paste everything between the two BRIEF markers below, and paste Gemini’s full answer under ## Talbot Response.
2. Claude (/task-continue on this file): triage each objection as fatal, material, minor or wrong (check every factual claim at its source; a confident answer is not a correct one), re-run any test an objection breaks, and patch both reports and the parent’s findings.
3. Record the verdict (what survived, what changed) in the cape-timing README and the parent task.

Success Criteria

• Gemini’s answer is pasted in full.
• Every objection has a verdict with evidence.
• The reports and the brief’s findings are updated for every fatal or material objection, or the objection is refuted with a source.

The brief (paste this)

The brief is regenerated from results by uv run python scripts/build_brief.py in ~/projects/cape-timing. Don’t edit it by hand.

Red Team brief: equity allocation at record valuations

You are the red team: a sceptical quantitative finance reviewer. Your job is to disprove or weaken the claims and findings below. Rank your objections by how much each would change the conclusions if true. For every factual claim you make, cite a checkable source; say “unverified” when you can’t. Do not restate the brief.

Question

For a long-horizon investor (Canadian, financially independent, can tolerate a 50–80% drawdown), does an extreme starting Shiller CAPE justify holding less in stocks than the usual “invest the lump sum now” rule implies?

Headline claims under test (these will lead a public research report)

A. At extreme valuations (US CAPE ≈ 40 today, 99th percentile), almost every equity investor should reassess their equity allocation: the 10-year reward-to-risk is sharply lower, not just the expected return. B. A simple pre-committed rule, “halve the stock allocation while the 10-month trend is broken”, costs almost nothing over the long run and cuts the worst decline by a quarter to a third. C. The value of acting is large: from the Dec 1999 peak, doing nothing left a typical investor tens of thousands of dollars per $100,000 behind the rule (findings 12–13). D. Few practitioners act on this, although the research exists (Campbell & Shiller, Asness 2017, PWL 2024, Estrada 2026, Palazzo 2026).

Data (all total return, monthly)

• US: Shiller ie_data (S&P Composite, monthly-average prices, 1871–Sep 2026). Ken French 1-month T-bill + CRSP market, month-end (1926–Jul 2026). Gate: Shiller vs CRSP annual correlation 0.97. Findings 1–2 use Shiller; 3 and 5–6 use CRSP, so the same bucket can differ by a few points between them.
• Canada: StatCan TSX price + dividend yield (1956–2016), extended with XIC.TO rebuilt from close + dividends; StatCan 3-month T-bill; CPI; FRED CAD/USD. Gate vs XIC 2001–2016: correlation 0.998.
• Percentiles are real-time (expanding window from 1881, minimum 20 years). Strategies trade the month after the signal, with CAPE lagged one extra month and a 0.1% cost per switch.

Findings to attack

1. Base rate (Shiller): US stocks trailed T-bills over 10 years in 14% of 1926–2016 starts; from top-5% real-time CAPE starts, 55%. n_eff ≈ 1.1 independent decades (1929, 1998–2000).
2. Bayes: P(top-10% CAPE | lost to cash) = 69% vs P(top-10% | beat cash) = 21%; likelihood ratio 3.3.
3. Risk (CRSP): from top-5% starts, P(10-yr real > 0) fell from 88% to 52%. Median in-window Sortino fell from 0.76 to 0.13; median volatility 15.5% vs 16.2%.
4. Forecast: at CAPE 40.6, the log-CAPE regression (1881–2016 starts) gives +0.4%/yr real and the CAPE-yield model +2.8%. OOS R²: 0.05 (1921–2016), -0.47 (1990–2016). OOS calibration: realized ≈ 0.040 + 0.68 × forecast.
5. Merton share (σ 18.3%, risk aversion 4): 62% on history (excess over historical T-bills); 33% calibrated and 14% raw forecast (both excess over today’s 1.45% real 10-yr yield); 12% on top-5% history.
6. Strategy: “cash only when CAPE ≥ 90th pct AND price < 10-month SMA” earned 11.0% vs 10.3% lump sum (1926–2026), but 11.9% vs 12.0% over 1976–2026. Six rules were tested.
7. Canada, after tax (2026 Ontario rates, $200k bracket, gains deferred to year 10): the TSX beat taxable CAD T-bills in 98% of 10-year windows (unconditional) vs 83% pre-tax. US stocks held in CAD from top-5% US CAPE starts beat taxable CAD cash 35% of the time (33% pre-tax).
8. Halving rule (cut stocks by 50% when the 10-month trend breaks, no valuation filter), US 1926–2026: 10.2% vs 10.3% a year, worst decline -62% vs -84%. For 60/40 (30/70 when fired, bonds = Shiller 10-yr Treasury returns): 8.5% vs 8.5%, worst decline -40% vs -63%.
9. Cash sensitivity (cash = T-bill −1.0/−0.5/0/+0.5 pt, to represent savings accounts and GICs): P(stocks beat cash | top-5% CAPE) = 53% / 47% / 45% / 37%.
10. Real-time forecast check: at Dec 1999 (CAPE 44.2) the recursive log-CAPE forecast was -4.0%/yr real; realized -3.2%; historical-mean assumption +6.7%.
11. Japan: dividend yield below 1% from 1985-01; the index then rose 225% to the 1989-12 peak, fell -81%, and recovered in price by 2024-02. A 10-month trend rule exited 1990-02, -11% from the peak.
12. “If you had used it”, per $100,000 from Dec 1999 (10 years; US; 100/0 cash-based and 60/40 bond-based): do nothing $96,264 / $132,165; halve on trend break $144,121 / $182,297; all out on trend break $204,652 / $245,580; safe asset all decade $131,068 / $190,923. Worst decline: do nothing -50% / -28%; halving -30% / -10%.
13. Spread, not just the peak: over all 134 top-5% real-time CAPE start months with a full 10 years (start years 1928–2004, clustered in 3 episodes), the halving rule finished ahead of doing nothing in 100% (100/0) and 97% (60/40) of starts; median gain $29,016 / $29,401 per $100,000 (10th–90th pct $17,772 to $48,344 for 100/0). Holding the safe asset all decade finished ahead in only 53% / 63%.
14. Recent cost: 2017–2026 the all-out trend rule earned 9.7% vs 14.9% buy-and-hold (worst decline -18% vs -25%).

Known weaknesses (already acknowledged; go beyond these)

• Top-valuation evidence is 1–3 episodes; overlapping windows.
• Starts after 2016-09 are censored, and all are strongly positive so far.
• CAPE drift and accounting changes (Palazzo 2026 CAPE-H) are not corrected for.
• Merton is one-period and parametric; there is no horizon-dependent mean reversion.
• Today’s tax rates are applied to history.
• Japan returns are approximate (FRED Nikkei price, JST dividend yield and bill rate).
• The halving rule’s per-era results are not yet computed.
• Findings 12–13 start in hand-picked expensive months; every such 10-year window that ended by now started before 2017, so the expensive 2017–2021 starts (strongly positive so far) are excluded by censoring.
• The trend rule is Faber’s 10-month SMA (published 2007); its best episodes (1929, 2000–02, 2008) are in-sample.

Deliver

1. The 5 strongest objections, each with its likely effect on claims A–D and findings 1–14.
2. Any computational error you can infer from the numbers above (internal inconsistencies).
3. One better test we should run instead, fully specified.

Talbot Response

---