Regime layer · weekly snapshot · 2026-W37
Weekly Regime Outlook
Model-conditional probabilities that each asset is in a given market regime 21 trading days from now — from a preregistered, out-of-sample-validated covariate model, refreshed weekly. Data through 2026-09-04, generated 2026-09-07.
These are probabilities of membership in operationally defined regime classes (trailing 21-day volatility/drift quantile labels) — not a forecast of returns or market direction, and not investment advice. The number in parentheses is the unconditional baseline: what you would assign knowing nothing about current conditions. Where the model and the baseline differ is exactly what current observable state adds.
SPY
now: BULL21-day membership probability (unconditional baseline)
Stress classes (BEAR or CRISIS): 5.8% vs 14.7% unconditional (-9.0 pp)
QQQ
now: BULL21-day membership probability (unconditional baseline)
Stress classes (BEAR or CRISIS): 4.1% vs 8.8% unconditional (-4.7 pp)
GLD
now: CRISIS21-day membership probability (unconditional baseline)
Stress classes (BEAR or CRISIS): 57.4% vs 14.8% unconditional (+42.5 pp)
TLT
now: SIDEWAYS21-day membership probability (unconditional baseline)
Stress classes (BEAR or CRISIS): 12.1% vs 27.2% unconditional (-15.0 pp)
Cross-category market map
One compressed view across 18 category proxies — US large-cap, the nine SPDR sectors, international equity, bonds & credit, commodities, Bitcoin — on the same operational regime classes and the same 21-day horizon. No direction is forecast: each row carries regime-membership and stress probabilities only, and each (asset, horizon) cell ships exactly the model tier that passed a preregistered out-of-sample rule (covariate / persistence / unconditional).
| Category | Proxy | Regime (days) | P(stress, 21d) | unconditional | Validated tier |
|---|---|---|---|---|---|
| US large-cap equity | SPY | BULL (2d) | 5.8% | 14.7% | covariate |
| US tech / growth equity | QQQ | BULL (2d) | 4.1% | 8.8% | covariate |
| Technology sector | XLK | SIDEWAYS (24d) | 12.3% | 12.6% | persistence |
| Financials sector | XLF | BULL (2d) | 3.2% | 15.3% | covariate |
| Energy sector | XLE | BULL (2d) | 9.5% | 21.1% | covariate |
| Health-care sector | XLV | SIDEWAYS (22d) | 12.0% | 12.7% | persistence |
| Industrials sector | XLI | SIDEWAYS (6d) | 13.2% | 15.8% | persistence |
| Consumer-discretionary sector | XLY | SIDEWAYS (4d) | 10.1% | 14.5% | covariate |
| Consumer-staples sector | XLP | SIDEWAYS (6d) | 11.0% | 11.0% | unconditional |
| Utilities sector | XLU | SIDEWAYS (33d) | 12.8% | 14.3% | persistence |
| Materials sector | XLB | BULL (1d) | 9.7% | 13.9% | persistence |
| Developed ex-US equity | EFA | BULL (2d) | 12.5% | 17.3% | persistence |
| Emerging-markets equity | EEM | BULL (3d) | 8.3% | 16.9% | covariate |
| Long US Treasuries | TLT | SIDEWAYS (4d) | 12.1% | 27.2% | covariate |
| High-yield credit | HYG | SIDEWAYS (4d) | 9.9% | 9.8% | persistence |
| Gold | GLD | CRISIS (6d) | 44.0% • | 14.8% | persistence |
| Crude oil | USO | SIDEWAYS (7d) | 21.8% | 25.2% | persistence |
| Bitcoin | BTC | SIDEWAYS (2d) | 12.3% | 13.9% | persistence |
• = stress probability materially above the row’s own unconditional baseline (≥ max(15%, 1.5×)). US sector rows largely re-express one common equity factor — the map is fewer independent signals than rows. Full per-category detail (probability vectors, baselines, conditional forward-return quantiles as historical description, equity-factor commonality, per-cell DM p-values) is in /outlook.json under market_map, daily-refreshed via the market_regime_map MCP tool. Descriptive, not advisory; no ranking is implied.
How to read this
The four classes are operational labels computed from each asset’s own history — BULL (positive trailing drift, calm volatility), SIDEWAYS (the residual class — no strong drift/vol signature), BEAR (negative trailing drift, elevated volatility), CRISIS (volatility above its expanding 90th percentile). Thresholds are expanding-window quantiles, so a label never uses information from after its own date.
Validation: Out-of-sample 2013-2025, expanding-window yearly refits: covariate logit beat the persistence-Markov baseline (DM p<0.05) on SPY/QQQ/TLT at h=5 and h=21; GLD h=21 direction-only. Persistence baseline beat the unconditional distribution in 8/8 cells. Annual Fourier seasonality terms were tested and FALSIFIED out-of-sample (0/8 cells improved; significantly worse on GLD and SPY h=21) — excluded from this model.
Model tier: covariate_logit_v1, refit on the full available history (6,186 training pairs) — the preregistered expanding-window protocol, applied at load time. Probabilities are shrunk toward the unconditional distribution; stress-class estimates carry the widest uncertainty and tend to be slightly conservative.
Query this live
This page is a weekly snapshot. The regime_outlook tool on the open MCP server serves the same model with daily-refreshed data (free tier, no key) — plus the market regime map, portfolio stress, IPS gate, and 12 more diagnostics.
https://mcp.crashtestyourstrategy.ai/mcp
claude mcp add --transport http ctys https://mcp.crashtestyourstrategy.ai/mcp
Capability declaration →Data & archive
This snapshot as JSON: /outlook.json — dated snapshots accumulate at /outlook/archive/ (e.g. 2026-W37.json), so a citation of “the 2026-W37 outlook” stays checkable after the page moves on. Every payload carries data_through and per-asset data_staleness_days — verify freshness yourself.
Methodological limitations
- Probabilities are model-conditional statements about membership in operationally defined regime classes (trailing 21-day volatility/drift quantile labels) — not predictions of returns or market direction, and not a claim about future market behavior.
- Validated out-of-sample under a preregistered protocol (expanding-window, 2013-2025): the covariate model beat the persistence baseline at p<0.05 on SPY/QQQ/TLT; on GLD at h=21 the improvement was directional but not significant. Only horizons of 5 and 21 trading days and the four listed assets are validated — other inputs are rejected rather than extrapolated.
- The model mildly underpredicts the two stress classes (e.g. realized CRISIS share 5.6% vs predicted 4.4%, SPY h=21 test window); treat stress probabilities as slightly conservative.
- BEAR/CRISIS days are rare (~10% of history each); conditional estimates for stress regimes carry the widest uncertainty. Probabilities are shrunk toward the unconditional distribution.
- Annual seasonality terms were tested and falsified out-of-sample; calendar patterns are deliberately not part of this model.
- The underlying price history is refreshed daily with end-of-day bars (no intraday feed): the outlook is computed at the end of the stored series (see data_through / data_staleness_days), which typically lags the present by one trading day.
- Descriptive, not advisory. No suitability, timing, or ranking claim is made or implied.
Model-conditional probabilities of membership in operationally defined regime classes — descriptive, not a market prediction, not investment advice. Fields follow the ctys-agent schema family (snake_case); the live source of truth is the regime_outlook tool on the open MCP server. Operated under German jurisdiction (BaFin / WpHG framework): model-based scenario simulation — descriptive, not advisory; not a forecast; not investment advice.