Robotics & Physical AI
An editorial framework — AI value creation extends from software into the physical world; the exposure is built as a basket satellite, not a winner-pick.
Own the robotics build-out as a diversified satellite on a world-market core — because the theme’s worst enemy is not the crash, it is the rate regime.
Headline finding
The theme’s malus is regime-specific: in the COVID crash the robotics basket fell WITH the world market (not deeper, +1.2pp) — but in the rate-driven regimes it lost an extra 20–26 points (sleeve −46.7% vs VT −26.1% in the 2022 core). A growth theme carries duration character: it dies by the discount rate, not by panic. Concentration doubles the damage: all-theme −46% vs balanced −31% (2022), a paired synthetic gap of 14–19 points.
The framework
The economic thesis the portfolio is built on — stated as falsifiable claims with evidence grades.
The next phase of AI value creation extends from software into the physical world: robotics, industrial automation, autonomous systems. Demographics tighten labour supply while sensor, actuator and compute cost curves keep falling, and foundation-model progress lowers the marginal cost of automating physical tasks. The framework deliberately does NOT claim to know which company wins or when humanoids commercialise — which is exactly why the portfolio expression is a basket satellite with staged risk tiers, not a concentrated bet. The stress evidence adds the framework’s most important correction: the dominant risk channel of the theme is not the industrial capex cycle but the RATE regime — a growth theme carries duration character.
Pillars
Falsifiers
- F1 — If global industrial-robot installations shrink in real terms for ≥ 5 consecutive years outside a recession, the secular-adoption pillar fails.
- F2 — If automation capex stagnates through a full demographic tightening cycle (labour scarcity without an automation response), the demographics pillar fails.
- F3 — If in the NEXT rate-shock regime the theme sleeve does NOT underperform the world market materially (< 5pp extra drawdown), the duration-character finding is falsified and the sizing logic here is too conservative.
- F4 — If a commercially scaled general-purpose humanoid deployment emerges, P5 graduates from conjecture — and the basket, which holds the suppliers, captures it without having predicted the winner.
Read the full framework summary6 sections ▾
What this thesis is — and is not
This is the catalog’s first editorial thesis: no source paper, the framework is constructed here and stated as falsifiable claims. It is deliberately NOT a claim that robotics stocks are cheap, that a specific company wins, or that humanoids arrive on schedule. It is the narrower claim that physical-AI adoption is secular (P1/P4), that its winners are unknowable in advance (P5 is graded conjecture), and that the correct expression of that combination is a diversified basket, sized by risk tier, on a world-market core.
Why a basket, not a winner-pick
The all-theme benchmark (100% robotics basket) shows what concentration in the thesis costs: −46% in the 2022 core vs −31% for the balanced tier, −44% in the growth winter vs −27%, and a paired synthetic gap of 14–19 points of median drawdown at matched paths (seed-stable to <1pp). The pre-registered concentration thresholds were met on both the synthetic and the real path — it is the single most expensive mistake available inside this thesis.
The duration character of a growth theme
The central measured finding: the theme’s malus is regime-specific. In the fast, liquidity-driven COVID crash the basket fell alongside the world market (−32.9% vs −34.1% — NOT deeper). In the rate-driven 2022 regimes it lost an extra 20–26 points. Long-duration cash flows — growth stories whose value sits far in the future — are discount-rate assets: they die by the rate, not by the panic. This is the correction the stress test forces on the usual capex-cycle framing.
The two faces of the TLT insurance
The defensive sleeve holds long Treasuries deliberately, with the sharpest possible semantics: TLT insured the deflation crash (+14% in the COVID core, +2.1 to +2.9 portfolio points of cushion) and became a liability in the rate shock (−31% in the 2022 core, −4.3 to −6.0 portfolio points). 2022 was the regime where diversification AND the hedge failed simultaneously — and the tiers still staggered the damage monotonically (25% / 31% / 37% / 47% max drawdown). The insurance is real, and its failure mode is documented rather than hidden.
The honest counter-finding
One pre-registered implication was rejected by the data: the expectation that a theme basket underperforms the broad market (the well-documented thematic-ETF procyclicality pattern). In this window the sleeve BEAT the world market absolutely (+15.0% vs +12.7% p.a.) — bought with +8.8 points of extra volatility and a 52% vs 34% max drawdown, and the edge stems almost entirely from the 2023–2026 AI rally inside a window that starts at the youngest ETF’s inception. Risk-adjusted, the ranking reverses (~0.58 vs ~0.74 return/vol). We report the rejection rather than tuning the window until the expectation holds.
Limits
The window begins at BOTZ’s 2016 inception — no GFC test is possible for the theme, and the window contains the full AI rally, so the return comparison (B6) is window-sensitive. Tail quantiles are indicative (effective sample size ≈ 9). The factor model reproduces the calm correlation structure well (mean pair error 0.02) but understates the crisis deepening of the equity–TLT correlation, and its rate-shock correlation weakens toward zero rather than flipping positive as 2022 really did. All figures USD; no fees or taxes modelled; the TLT verdicts hold for the tested regimes only — hedge betas are non-stationary.
From framework to portfolio
If adoption is secular but the winners and the timing are unknowable (P5 is graded conjecture, not conviction), then the expression follows: a BASKET carries the theme (concentration is the most expensive error available — measured, not asserted), a world-market core carries everything the theme is not, and the defensive sleeve pairs cash with long Treasuries — an insurance whose payout (deflation crash) and whose failure mode (rate shock) are both documented in the evidence. The tiers express one decision only: how much of the portfolio may live at the theme’s duration risk.
The portfolio
ETF-only, no single names, no leverage. Three robotics ETFs (equal-weighted 1/N inside the sleeve) diversify index-construction risk: BOTZ (global robotics & AI, cap-weighted), ROBO (broad robotics & automation, modified equal-weight), ARKQ (active, autonomous-tech tilted). VT is the world-market core. TLT + BIL form the defensive floor — long-duration insurance plus optionality cash. Deliberately NO single-stock pure plays: the theme’s winners are the part the framework admits it cannot know.
| Building block | Role / regime defended | Conservative | Balanced | Offensive |
|---|---|---|---|---|
| Global X Robotics & AI (BOTZ) | Cap-weighted global robotics & AI basket — the theme’s large-cap expression | 5% | 10% | 18% |
| ROBO Global Robotics (ROBO) | Broad, modified equal-weight robotics & automation — widest supplier coverage | 5% | 10% | 18% |
| ARK Autonomous Tech (ARKQ) | Active autonomous-systems tilt — the aggressive edge of the theme | 5% | 10% | 18% |
| World equities (VT) | The non-theme core — everything the thesis is not | 45% | 40% | 26% |
| Long Treasuries (TLT) | Deflation-crash insurance with a documented rate-shock failure mode | 20% | 20% | 15% |
| T-bills / cash (BIL) | Optionality reservoir | 20% | 10% | 5% |
How the critical sizes were chosen
- The theme sleeve (15% / 30% / 54%) is the only lever the tiers move materially — it expresses how much of the portfolio may live at the theme’s duration risk, not a market view.
- Inside the sleeve, 1/N across the three ETFs: index-construction risk (cap-weight vs equal-weight vs active) is diversified instead of predicted, consistent with the catalog’s brute-force finding that 1/N is hard to beat ex-ante.
- TLT, not a bond aggregate: the sharpest insurance semantics available — it pays in the deflation crash (+14% COVID core) and its 2022 failure (−31%) is priced into the sizing rather than discovered later. The conservative tier caps the combined theme+duration exposure at 35%.
The stress evidence
Regime-switching SV factor model — 50 synthetic paths × 6 profiles, 252-day horizon, calibrated on 2016–2026 (2,471 trading days, bounded by BOTZ inception), all four allocations assessed on IDENTICAL factor paths (paired comparisons, seed-stable), plus 5 real historical episodes in both rebalancing conventions. All claims pre-registered before any run; one (B6) was rejected by the data and is reported as such. Tail quantiles indicative (ESS ≈ 9). Descriptive: assumptions stated, not a forecast.
Tiers — real backtest
| Tier | Return p.a. | Vol p.a. | Sharpe | Max DD | Worst regime |
|---|---|---|---|---|---|
| Conservative A 15% robotics satellite on a broad, bond-anchored core | +8.6% | 11.4% | 0.78 | −26.7% | Rate shock −23% · p95 −32% (moderate) |
| Balanced A 30% theme sleeve, still market-anchored | +9.9% | 14.2% | 0.74 | −33.1% | Rate shock −26% · p95 −37% (moderate) |
| Offensive Theme-led (54%) with a reduced defensive floor | +11.6% | 17.8% | 0.70 | −40.0% | Crisis shock −32% · p95 −48% (fragile) |
Real backtest over the common window (2014–2026). Worst regime = synthetic profile with the deepest median drawdown; p95 = 95th-percentile drawdown across the n=50 paths (the tail, not the typical case). The grade reflects the median — read it together with the p95.
Synthetic regimes — drawdown depth
Cells: median max-drawdown across n=50 synthetic paths. p95 = 95th-percentile drawdown (the tail — roughly the worst 1-in-20 path). p5 = 5th-percentile 1-year return. Colour ∝ median drawdown depth.
The hedge that breaks
Long Treasuries are the sharpest crisis insurance available in this universe — and the most honest about their failure mode. In the COVID deflation crash TLT returned +14% while the theme sleeve fell −33%, adding +2 to +3 portfolio points of cushion. In 2022 the same duration that pays in deflation turned against the portfolio: TLT −29% over the year, subtracting 4–6 portfolio points exactly when the growth theme was also bleeding. The real crisis correlation to world equity deepens from −0.04 (calm) to −0.27 (stress) — measured before any synthetic run, and reproduced qualitatively by the factor model.
Real historical episodes
| Episode | Window | Conservative | Balanced | Offensive | Single-thesis |
|---|---|---|---|---|---|
| COVID crash Fast deflation shock: the robotics basket fell WITH the market, not deeper (TLT insured, +14%). | Feb–Mar 2020 | −17.4% | −20.6% | −24.5% | −32.9% |
| 2022 rate shock The theme’s worst regime — duration character: sleeve −42% vs world −18%, and TLT broke (−29%). | full-year 2022 | −20.2% | −25.6% | −31.6% | −41.6% |
| 2018 Q4 selloff | Sep–Dec 2018 | −11.0% | −14.1% | −18.2% | −26.5% |
| Growth winter The concentration test: all-theme −44% vs balanced −27%. | Nov 2021–Nov 2022 | −20.9% | −26.7% | −33.3% | −44.4% |
| Current regime From the sleeve’s late-May all-time high — the theme sits ~12% below it; the tiers carry it mildly. | May–Jul 2026 | −2.2% | −4.0% | −6.7% | −11.7% |
Buy-and-hold end return per episode. Single-thesis = All-theme benchmark (100% robotics basket BOTZ/ROBO/ARKQ — the single-thesis concentration).
Document-claim verdicts
| # | Document claim | Verdict | Conf. | Note |
|---|---|---|---|---|
| B1 | The risk tiers stagger volatility monotonically | Supported | High | Full-window annualised vol 11.4% < 14.2% < 17.8% < 25.4% — strictly monotonic across tiers and benchmark. |
| B2 | The theme sleeve is NOT a crisis diversifier — it is high-beta equity | Supported | High | In every bear episode the sleeve returned at or below world equity (within the +2pp tolerance in COVID, −8.7 to −26.4pp in the rest). Nuance: the malus concentrates in rate/valuation regimes, not in the deflation panic. |
| B3 | Concentrating in the theme (all-theme vs balanced) costs double-digit drawdown | Supported | High | Paired synthetic gap +13.6/+17.9pp median drawdown (seed-11 check: +14.4/+18.6 — the DIFFERENCE is stable to <1pp); real gaps +16pp (2022) and +18pp (growth winter). Both pre-registered thresholds met. |
| B4 | TLT is insurance with a documented failure mode | Supported | High | COVID core +14.2% solo (sleeve +2.1…+2.9 portfolio pp); 2022 core −30.6% solo (sleeve −4.3…−6.0pp). Synthetic rate-shock reproduces the break (hedge sleeve −14%). Verdict holds for the tested regimes only — hedge betas are non-stationary. |
| B5 | The tiers stagger real crisis drawdowns and cap the conservative tier | Supported | High | Max drawdown strictly monotonic in 2022 (25/31/37/47%) and the growth winter (27/33/40/51%); conservative ≤ 0.6× all-theme in both (ratio 0.53). |
| B6 | Thematic-ETF procyclicality: the theme basket does not out-return the world market | Not supported↳ Overstated | Medium | Rejected by the data — the sleeve beat VT absolutely (+15.0% vs +12.7% p.a.) in this window, almost entirely via the 2023–2026 AI rally; risk-adjusted the ranking reverses (~0.58 vs ~0.74 return/vol). Window-sensitive (starts at BOTZ inception). Reported as rejected rather than re-windowed. |
Caveats
- Window starts at BOTZ inception (Sep 2016): no GFC test is possible for the theme, and the window contains the full 2023–2026 AI rally — the B6 return comparison is window-sensitive.
- Tail bands indicative: effective sample size ≈ 9 independent 252-day blocks; p95 drawdowns and p5 returns are directional, not calibrated quantiles.
- Monte-Carlo spread: the all-theme benchmark’s median drawdown varies ±5.8pp across seeds (core tiers ≤3.2pp); the paired tier-vs-benchmark DIFFERENCES are stable to <1pp and are the load-bearing numbers.
- The factor model matches calm pair correlations well (mean error 0.02) but understates the crisis deepening of the equity–TLT correlation and does not flip it positive in the rate shock as 2022 really did.
- TLT verdicts hold for the tested regimes only — hedge betas are non-stationary (a calm-window beta does not know flight-to-safety).
- Per-episode rebalancing convention matters mainly in the COVID recovery window (constant-mix up to +2.2pp better — it buys the dip); both conventions are in the artifacts.
- All figures in USD; no EUR overlay was run (an available engine option, deferred); no fees, taxes or tracking difference modelled.
- Calibration window 2016–2026 is shorter than thesis #1 (2014–2026) and thesis #2 (2008–2026) — regime tables are not directly comparable across theses.
Found this useful?
I run bespoke stress analyses and build portfolio frameworks like this one on request.
Commission your own →Descriptive stress-test case study — not investment advice, and not a suitability or recommendation statement. The framework is an editorial set of falsifiable hypotheses (no external source paper); the portfolio is an illustrative, rule-based expression for independent analysis. Stress results are model-dependent and not a forecast of future market behaviour.