Refitted each quarter, the S&P 500's widest band held 43 of 46 quarter-ends
How far could each market move over the next quarter and year, and what would each named path do to it?
In brief
The 's range for each market, centred on no change, the record of how often its bands have held, and what each named path does to the median- Refitted each quarter on the data to that date, the S&P 500's widest band held 43 of 46 quarter-ends since 2015.→ Exhibit 6
- Over the next quarter the S&P 500 lands between −11.4% and +12.9% in four cases of five, with no view on direction.→ Exhibit 1
- The 10-year moves between −58bp and +57bp over the next quarter in four cases of five.→ Exhibit 2
- WTI crude ranges from −23.1% to +29.3% over the next quarter in four cases of five.→ Exhibit 2
- Stocks and long Treasuries both fall 5% or more over the next quarter in 4.9% of the model's paths.→ Exhibit 12
| What would change this view | Trigger | Today | Decided by |
|---|---|---|---|
| The S&P 500 window from 2026 Q3 leaves its 90% band | S&P 500 change from the 30 Sep 2026 close outside −11.0% to +13.1% at session 65 | −1.1% after 32 of 65 sessions (13 Nov 2026)pending | the close of 4 Jan 2027 (session 65) |
| The 10-year window from 2026 Q3 leaves its 90% band | 10-year change from the 30 Sep 2026 close outside −69bp to +71bp at session 65 | −17bp after 31 of 65 sessions (16 Nov 2026)pending | the close of 6 Jan 2027 (session 65) |
| The bands are too narrow for the regime | more than one 90% miss for the S&P 500, or for the 10-year, in eight consecutive quarter-ends | 1 and 2 in the last eight (2024 Q3 – 2026 Q2)crossed | each window as it completes (next 4 Jan 2027) |
| The bands are too wide | eight consecutive quarter-ends inside the S&P 500's 50% band | 1 in a row to 2026 Q2not crossed | each window as it completes (next 4 Jan 2027) |
The reasoning behind each point
- CoverageRefitted each quarter on the data to that date, the S&P 500's widest band held 43 of 46 quarter-ends since 2015.
Out of sample: at each quarter-end from 2015 Q1 to 2026 Q2 the S&P 500's GARCH(1,1)-t is re-estimated on its returns since 2 Mar 2007 up to that date only, and the band drawn from it scores the next 65 sessions. The 90% band held 93% (Kupiec's test of the miss rate against 10%, p 0.40; Christoffersen's test that misses do not cluster, p 0.51); the 80% band 78% (p 0.77); the 50% band 33% (p 0.02) — the zero-drift centre sits under a market that rose. The 10-year, refitted the same way: 90% 85% (p 0.27), 50% 48% (p 0.77). Windows outside a refitted 90% band, by the quarter-end they start from: 2016 Q3; 2018 Q3; 2019 Q4; 2020 Q4; 2021 Q4; 2023 Q2; 2024 Q3; 2024 Q4; 2026 Q2. The cones' own record, with the volatility model's parameters fitted on the whole sample, held 43 of 46 for the S&P 500 and 85% for the 10-year: in-sample parameters, which have seen the shocks they are scored on. 46 windows are a short record: the tests' p-values are asymptotic (χ²) and only coarse at this size.
- EquitiesOver the next quarter the S&P 500 lands between −11.4% and +12.9% in four cases of five, with no view on direction.
The 10th to 90th percentile from the 13 Nov 2026 close, around a zero-drift median of ±0.0%; the 5–95 band is −15.8% to +19.0%. 26 weeks: −15.6% to +18.2%; 52 weeks: −21.4% to +27.4%. With the fitted mean instead of zero drift the 13-week median moves to +6.7%. Nasdaq-100 −18.2% to +22.1%, semiconductors −30.0% to +43.1%. The width is the model's current variance state (S&P 500 daily volatility 23.2% annualised at the start) pulled towards its long-run level over the horizon.
- RatesThe 10-year moves between −58bp and +57bp over the next quarter in four cases of five.
The yields model's 10th to 90th percentile; the 2-year −70bp to +70bp. 10-year over 26 weeks −79bp to +79bp; 52 weeks −110bp to +110bp. The yields model's DCC a = 0.0213, b = 0.9647; 10-year GARCH ν = 13.1. Long Treasuries (TLT, ten core markets) −7.8% to +8.5%.
- Oil, gold, dollarWTI crude ranges from −23.1% to +29.3% over the next quarter in four cases of five.
Gold −14.0% to +16.4%, the dollar (UUP) −4.1% to +4.2%; on the yields model USO −25.8% to +34.6%, gold −14.8% to +17.0%, UUP −4.1% to +4.3%. Two models on overlapping assets differ because each series' GARCH is fitted on its own set's sessions and the correlations are estimated jointly with different partners; the gap between them is a specification sensitivity in its own right. Bitcoin's cone (−32% to +48%) sits on a fit at the α + β = 1 boundary (α + β = 1.0000), so its expected variance grows with the horizon instead of settling at a long-run level.
- Joint eventsStocks and long Treasuries both fall 5% or more over the next quarter in 4.9% of the model's paths.
Against independence the bonds-fail event would be 5.6% — the difference is the conditional correlation (−0.04 S&P–TLT at the start) doing its work. P(10-year +50bp and S&P 500 −10%) = 1.3%; P(10-year −50bp and S&P 500 +5%) = 3.0%; P(USO +20% and S&P 500 −5%) = 5.2%; P(gold +10% and UUP +3%) = 0.9% (ten core markets). All are model-conditional: the share of paths the fitted model assigns under its own assumptions, not a judgement.
- QuarterOver the next quarter the factor model's S&P 500 median runs from −1.1% to −0.9% across the paths, all inside the range with no scenario.
The no-scenario 10th–90th band over 1 quarter is −11.6% to +15.8%. At a year the paths' medians lie within a fraction of a point of one another, because the model's stock–bond correlation reverts towards zero; a quarter is where they part.
- PathsThe rate paths put the 10-year between 4.07% and 4.17%; none of them carries odds.
The Fed paths — hold path −106bp, SEP path −100bp, above the dots −96bp — each set the 10-year at the fair value What is priced gives a policy path: the distance from today's yield, so a path asks what happens if the 10-year travels that distance, not when or why. The path above the dots takes the policy-rules grid's one path above the SEP's, +50bp then hold; its fair 10-year still sits below today's yield, so its move (−96bp) is the smallest fall of the three — the market prices more tightening than even that path. The 4 Hormuz paths are listed without odds and are not projected: they carry no oil level yet; the formula cannot price them either, because no elasticity of index earnings or of the premium to oil exists among the site's estimates. No path carries a probability: no source the site reads states one. Setting the 10-year conditions the draws rather than replacing them, so the other factors keep their tails.
- FormulaAt the hold path's 10-year the model's value is still −29.7% from the price; the formula's full pass-through, +29.6%, is arithmetic, not a target.
Valuation-model arithmetic: the Valuation page's two-stage model — price over trailing earnings = k [(1 + g_L) + H (g_S − g_L)] ÷ (r − g_L) — at its priced inputs (k 0.84, E0 ÷ P 3.985%, g_S 15.1% fading to g_L 4.0% over 6 years, r = 5.12% + 3.502pp), which return today's price. A path moves rf by its full distance and holds the premium and growth — full pass-through, the extreme case — so its move from today is (r0 − g_L) ÷ (r1 − g_L) − 1: +29.6% on the hold path, +27.8% on the SEP path, +26.0% above the dots. The response is convex — −22% per point of the discount rate at today's rate — because r − g_L is 4.62pp today. Oil, credit and the dollar are not modelled on the formula: the site estimates no elasticity of index earnings or of the premium to them.
- MarketThe market prices the 10-year at 5.12%, 106bp above the fair value of a hold.
From What is priced (16 Nov 2026): the hold path's fair value is 4.065%, the SEP path's 4.115% and the market's own, the ACM-fitted 10-year, 5.111%; the 1y1y forward sits +115bp over the policy rate; the SEP medians are 4.1 for 2026 and 4.1 for 2027, 3.2 in the longer run (16 Sep 2026).
- UnconditionalThe factor model's 1-year central band for the S&P 500 runs from −19.3% to +35.6%.
Over 1 year the S&P 500's 10–90 band is −19.3% to +35.6% with the median at +4.7% — close to the cash leg because the simulation carries no drift, not because the engine expects it.
- Projection coverageThe published band held 82.4% of outcomes over the 9 quarter-ends kept out of its calibration, from 30 Jun 2024.
Calibrated: the 10–90 band held 77.1% of 214 asset-quarters since 31 Mar 2022 (74.0–79.9%) against a nominal 80%. Calibrated, so the factor projections head this page beside the engine cones: the width scale estimated on the quarter-ends to 31 Mar 2024 (k = 1.07) held 82.4% of the held-out quarter-ends from 30 Jun 2024 (76.4–88.2), which contains 80%, and the published bands carry it. The 10–90 band held 79.4% (76.8–82.0) with fixed betas and a GARCH-t residual, 78.0% (75.0–80.9) with the betas drawn and 77.1% (74.0–79.9) with the residual bootstrapped; widened by k = 1.07 it would hold 81.3% (78.5–84.0) — applied. Drawing the betas moves the pooled coverage −1.4pp; resampling the residual instead of simulating it moves it −0.9pp. Before any scale the band held 76.4% of the first 9 quarter-ends (to 31 Mar 2024) and 77.8% of the 9 since. A scale estimated on the first half moves the second to 82.4%, which holds, so it is applied. The backtest published before 24 September left the residual out of the band it scored and mapped the four factor-proxy ETFs through betas, so it tested a narrower object than the page drew. The verdict sits on a knife-edge: the uncalibrated band's interval misses 80% by 0.1pp, and under 1 of the 4 alternative seed triples the spec lists it contains 80% instead — so the page's placement owes as much to the simulation's seed as to the bands.
- IndexOver three years each +10bp on the 10-year has moved the S&P 500 −1.08%, with credit and the dollar held.
Per 1% of crude ±0.00%; the 13-week band is about ±8%. Rates beta −1.08 (SE 0.32), oil beta −0.002 (SE 0.027), R² 0.26, residual 1.69% a week, credit and the dollar held; windows end on 13 Nov 2026. The band is almost all residual risk: √13 × 1.69% × 1.2816 = 7.8pp against a beta-uncertainty part below 2pp in every cell of the grid. Nothing in oil and rates alone justifies an index-level view over a quarter.
- Valuation routeOn the valuation formula a half-point rise in the 10-year moves the S&P 500 −9.8%; oil is not modelled.
A half-point fall lifts it +12.1%. The grid's valuation route is the Valuation page's two-stage model at the stated 10-year with everything else held (k 0.84, E0 ÷ P 3.985%, near-term growth 15.1% fading to 4.0% over 6 years, r = 5.12% + 3.502pp), which returns today's price at the zero column. It is one row, not a grid: the site estimates no elasticity of index earnings or of the premium to oil, credit or the dollar, so those are printed as not modelled rather than given a number. The response is convex (−22% per point of the discount rate at today's rate), which is why −50bp lifts the index by more than +50bp lowers it. Model-conditional: what the identity implies under stated inputs, not a forecast. The branch table on this page runs the same identity for every named scenario.
- PairThe energy − discretionary pair swings 15 points between the hollow-deal and escalation oil moves, against a band of about ±16.
−7.6 to +7.6: not larger than its uncertainty. Energy's oil beta +0.37 against discretionary's −0.06 per 1% (156 weeks); the band treats the two regressions' errors as independent (their residual correlation is +0.06, so the independent band is wider than a direct regression of the pair would give: 3.49 vs 3.38% a week).
- SectorsA rise in crude lifts oil & gas E&P most, +0.50% per 1%, and hurts homebuilders most.
Homebuilders −0.10% per 1%. Next: energy +0.37% and consumer discretionary −0.06% per 1% of crude. Per +10bp on the 10-year: homebuilders −2.38%, regional banks −2.21%; least affected consumer staples −0.54%. These are total effects (the market is not in the regression), so they include what comes through the index.
Projections · The claim is that the published bands hold what happens about as often as they say: widened by k = 1.07 on the first half, the 10–90 band held 82.4% of the 108 held-out asset-quarters from 30 Jun 2024 to 30 Jun 2026 (90% bootstrap interval 76.4–88.2%), which contains 80%, so the claim is not falsified on this sample.
Scenarios · The band, not the point, is the model's claim: it is contradicted if the S&P 500's realised 13-week return falls outside the 10th–90th band of the cell at the realised factor moves. To 13 Nov 2026 the moves were +14% oil, +47bp, credit −0.1%, dollar +3.1%: the model gives −7.0% (−14.9 to 1.0) against a realised −2.6% — inside the band.
Earnings · The read that the index's earnings level rests on an extended profit share fails if the after-tax share is back at its 2015–19 average (9.8%, the Profits page's threshold) for two consecutive quarters; the national-accounts read fails if the 0.2pp gap between index EPS growth and profits growth closes to under 0.1pp; the fan is contradicted if calendar-2027 EPS growth lands outside −9% to +22%, and its reading of consensus if the quoted calendar-2027 growth falls below +9.9%.
Next: Range record and stress tests — the range's record and worst weeks
How far each market could move over 3, 6 and 12 months, with no view on direction
The 's distribution over the next quarter, half-year or year, centred on no change so it measures width, not direction; the 's range for every asset beside itExhibit 1 · The selected market's recent path and its simulated range ahead
About this exhibit
Exhibit 2 · The selected market's return percentiles by horizon and drift setting
| Horizon | 5th | 10th | 50th | 90th | 95th |
|---|---|---|---|---|---|
| 13wzero drift | −15.8% | −11.4% | ±0.0% | +12.9% | +19.0% |
| 26wzero drift | −21.3% | −15.6% | +0.1% | +18.2% | +26.6% |
| 52wzero drift | −29.0% | −21.4% | −0.2% | +27.4% | +40.3% |
| 13wμ +0.099%/session | −10.2% | −5.5% | +6.7% | +20.4% | +26.9% |
| 26wμ +0.099%/session | −10.5% | −4.0% | +13.9% | +34.4% | +44.0% |
| 52wμ +0.099%/session | −8.1% | +1.7% | +29.2% | +64.9% | +81.6% |
About this exhibit
Exhibit 3 · Each asset's range on the factor model, by market group
The S&P 500 by two roads
Betas times factors, and earnings times a multiple — over the same year, from the same rates draw| Route | 5th | 10th | 25th | 50th | 75th | 90th | 95th |
|---|---|---|---|---|---|---|---|
| Factor routecash + Σ β·F̃ + ε | −22.4 | −19.3 | −7.2 | +4.7 | +18.1 | +35.6 | +40.9 |
| Valuation routeΔln P = Δln E + Δln PE, on today's multiple | −25.2 | −17.0 | −4.7 | +10.2 | +28.4 | +49.5 | +66.6 |
| — its earnings legthe Earnings page's default fan, NIPA bridged to EPS (−0.92 log points a year), by inverse CDF | −11.9 | −8.0 | −1.4 | +6.6 | +15.4 | +24.0 | +29.9 |
| — its multiple legthe two-stage identity at the drawn discount rate | −30.9 | −22.9 | −11.4 | +2.6 | +20.7 | +42.3 | +59.5 |
Why the two medians differ, and why this route's tails are long. The factor route's median (+4.7%) is the cash leg: it carries no drift. This route's (+10.2%) carries the Earnings page's bridged growth (median +6.6%) and a multiple leg whose median is +2.6%, because the premium's four-quarter changes, centred on their mean (−0.08pp removed), are skewed (median −0.25pp). Its tails are the identity's convexity: the multiple is close to k ÷ (r − g), with r − g = 4.62pp today, so the discount rate's 5th percentile (6.90%) lifts it by +59.5% while its 95th (10.69%) moves it by −30.9% — a straight line through today's slope would give +37.3% and −44.8%. The tails are drawn as they come, not trimmed.
The multiple moves from today's observed 18.66×, not from a fair one, so the route does not close the valuation gap — it moves what is there by the change in its inputs. r = rf + ERP, starting at 5.12% + 3.50pp, with rf moved by this page's own rates draw and the premium by the pool of 83 four-quarter changes the Valuation page draws its own distribution from, centred on their mean (5th −1.76pp, median −0.25pp, 95th +2.00pp). The three legs are tied by that page's Gaussian copula over 82 quarters. 1 of the draws (0.01%) put the discount rate within 0.5pp of the long-run growth rate, where the identity has no finite price, and are excluded and counted here rather than clipped.
About this exhibit
Four rate paths, no odds, beside the market's own path
What has to happen on each path, the 10-year it gives, the S&P 500 on the valuation formula and the , and four other assets over a quarterExhibit 4 · Each rate path's 10-year and its effect on stocks and other assets
| Scenario | What has to happen | 10-year, % | S&P 500, 1 quarterfactor model: median · 10th–90th | The model's value at that 10-year% from today's price | Formula's move% from today · premium and growth held | Goldmedian | WTI crude (USO)median | US dollar bullish (UUP)median | High yieldmedian | What we would need to see |
|---|---|---|---|---|---|---|---|---|---|---|
| What the market prices | No change: the policy path priced today, at today's term premium | 5.11%fair value on the market's path; par 5.12% on 16 Nov 2026 | +1.1%−11.6% to +15.8% | −41.5% | ±0.0%today's price, the formula's start | ±0.0% | ±0.0% | ±0.0% | ±0.0% | the reference: every row is read against it |
| Hold path | The Fed follows “Hold at IORB”; the 10-year falls to that path's fair value | 4.07%−106bp from today | −1.1%−13.4% to +13.3% | −29.7% | +29.6%full pass-through, no horizon | +8.2% | −10.4% | −2.2% | +2.1% | 10-year at 4.07%5.12% today, 106bp away |
| SEP path | The Fed follows “One further +25bp (the SEP path)”; the 10-year falls to that path's fair value | 4.12%−101bp from today | −1.0%−13.3% to +13.4% | −30.4% | +27.8%full pass-through, no horizon | +7.8% | −10.0% | −2.1% | +2.0% | 10-year at 4.12%5.12% today, 101bp away |
| Above the dots | The Fed follows “+50bp, then hold (above the SEP path)”; the 10-year falls to that path's fair value | 4.17%−96bp from today | −0.9%−13.2% to +13.6% | −31.0% | +26.0%full pass-through, no horizon | +7.4% | −9.5% | −2.0% | +1.9% | 10-year at 4.17%5.12% today, 96bp away |
No odds: the site reads no source that states one. At a year the factor model's medians for the paths lie within a fraction of a point of one another; a quarter is the horizon at which they part.
About this exhibit
Exhibit 5 · S&P 500 under each rate path against the no-scenario range, by horizon
About this exhibit
The bands have held about as often as they promise, since 2015
Coverage — how often past outcomes fell inside the band drawn at the time — at every quarter-end since 2015, refitted each quarter on the data to that date; the full record is on Range record and stress testsExhibit 6 · The selected market's outcome after each quarter-end against the band drawn then
About this exhibit
Exhibit 7 · How often each market missed its bands, scored both ways
| Market · record | 90% band held | Misses (expected) | Kupiec p | Christoffersen p | Both, p | 80% band held (Kupiec p) | 50% band held (Kupiec p) |
|---|---|---|---|---|---|---|---|
| S&P 500Refitted each quarter | 93% | 3 (4.6) | 0.40 | 0.51 | 0.57 | 78% (0.77) | 33% (0.02) |
| S&P 500The cones, full-sample parameters | 93% | 3 (4.6) | 0.40 | 0.51 | 0.57 | 74% (0.32) | 33% (0.02) |
| 10-year Treasury yieldRefitted each quarter | 85% | 7 (4.6) | 0.27 | 0.14 | 0.18 | 78% (0.77) | 48% (0.77) |
| 10-year Treasury yieldThe cones, full-sample parameters | 85% | 7 (4.6) | 0.27 | 0.14 | 0.18 | 76% (0.52) | 48% (0.77) |
| WTI crude (USO)Refitted each quarter | 89% | 5 (4.6) | 0.85 | 0.26 | 0.52 | 74% (0.32) | 50% (1.00) |
| WTI crude (USO)The cones, full-sample parameters | 87% | 6 (4.6) | 0.51 | 0.22 | 0.38 | 78% (0.77) | 52% (0.77) |
A p-value at or above 0.05 is read as consistent with the band's promise, not as proof the width is right. “–”: no miss was followed by another window, so clustering cannot be tested. The cones' own record has seen the shocks it is scored on; the refitted record has not.
About this exhibit
Have the projections' bands held?
Coverage of the 10-90 and 5-95 bands at every quarter-end, for every step of the calibration and both parameter variantsExhibit 8 · Factor model band coverage by calibration step and asset, against nominal
| Step | Coverage | 90% interval | Holds 80%? | Per asset, below |
|---|---|---|---|---|
| Fixed betas, residual by GARCH-tthe construction the page published to 23 September: each asset's betas at their point estimates and its residual simulated with its own GARCH(1,1)-t | 79.4% | 76.8–82.0 | yes | |
| (a) Betas drawn from their sampling distributioneach path draws its own betas from the standard errors of the regression that estimated them; the residual is still GARCH-t, which is the alternative the page keeps printing | 78.0% | 75.0–80.9 | yes | |
| (b) Residual by stationary block bootstrapthe residual resampled from the asset's own history of residuals, in blocks of 21 sessions on average, instead of simulated from a fitted GARCH | 77.1% | 74.0–79.9 | no | |
| (c) Calibrated width — publishedthe (b) band's 10–90 widened about its median by a scale k, and its 5–95 the widened 10–90 band plus a tail margin that cannot be negative. Both are fitted on the first half of the quarter-ends, and only the 10–90 band is tested on the second | 81.3% | 78.5–84.0 | yes |
The calibration, fitted on one half and tested on the other
k = 1.0721 for the 10–90 band and a tail factor of 0.4590 for the 5–95 — applied to every published band| Half | Quarter-ends | Before k | Interval | After k | Interval |
|---|---|---|---|---|---|
| First: k estimated here | 31 Mar 2022 – 31 Mar 2024 (9) | 76.4% | 70.7–83.8 | 80.2% | 76.8–85.0 |
| Second: held out, the test | 30 Jun 2024 – 30 Jun 2026 (9) | 77.8% | 71.5–84.7 | 82.4% | 76.4–88.2 |
calibratedthe (b) band's pooled 10–90 interval excludes 80% and the widened 10–90 band passes its held-out test, so the published bands carry k and the tail factor.
How much the verdict owes to the seed. Re-run under 4 alternative seed triples (factor draws, residual resample, beta draws), the uncalibrated band's pooled 10–90 interval contains 80% in 1 of them: 76.6% (73.9–79.4), 77.1% (74.3–80.1), 77.1% (74.5–79.9), 76.6% (73.8–79.4). The published bands use the spec's own seeds and are calibrated; 4 of the 4 alternatives would give calibrated.
Per asset: (c) Calibrated width.
Point in time throughout: the filtered state, the betas and their errors, the residual's variance and its bootstrap pool. The GARCH and DCC parameters are the full sample's, which contains the outcomes and therefore flatters these numbers. A factor-proxy ETF takes its raw factor draw, as on the page, so the betas and the residual change nothing for it; only the calibration does.
About this exhibit
Every asset under each path
The distribution of each asset's return over each horizon when a path's 10-year holds — the 's medians and bandsExhibit 9 · Each asset's return with and without the selected path
| Asset | Median, no pin | Median, pinned | Difference | 10th | 90th |
|---|---|---|---|---|---|
| MSCI ACWI | +4.7% | +4.8% | +0.1 | −18.6% | +33.9% |
| Dow Jones Industrials | +4.5% | +4.9% | +0.4 | −17.4% | +33.4% |
| Emerging markets | +4.6% | +6.3% | +1.7 | −25.8% | +50.7% |
| Developed ex-US (EAFE) | +4.5% | +7.7% | +3.2 | −16.3% | +38.9% |
| Japan | +4.6% | +4.4% | −0.2 | −22.2% | +39.3% |
| China large-cap | +4.3% | +8.2% | +3.9 | −24.6% | +54.2% |
| Russell 2000 | +4.6% | +6.7% | +2.1 | −23.5% | +48.7% |
| S&P MidCap 400 | +4.8% | +6.7% | +1.9 | −17.5% | +37.1% |
| Nasdaq-100 | +4.5% | +0.5% | −3.9 | −30.1% | +43.4% |
| S&P 500 equal weight | +4.7% | +7.8% | +3.1 | −12.7% | +32.6% |
| S&P 500 | +4.7% | +2.8% | −1.9 | −20.7% | +33.0% |
| Europe | +4.4% | +8.4% | +4.0 | −17.8% | +42.8% |
Why the pins barely move the index: a ten-year pin reaches the S&P 500 almost only through the market factor's co-movement with the ten-year in the engine's draws — +0.21% per 10bp over 1 month, +0.16% per 10bp over 1 quarter, +0.12% per 10bp over 6 months, +0.13% per 10bp over 1 year — today's correlation at short horizons, reverting to about zero, its average since 2018, by a year. Over 1 year the hold path's −106bp therefore shifts the index's mean return by −1.78pp, within noise in a distribution whose 10–90 band is 54pp wide: −1.38pp through the market, +0.02pp through credit, −0.08pp through the rest and −0.34pp through the index's own ten-year beta once the market is in the regression (+0.033% per 10bp, t = 3.44, beyond the site's |t| > 2).
What is priced puts the same path at +11.3% for the S&P 500 because it is a different model: a total-effect beta (−1.08% per +10bp from 156 weekly returns since 24 Nov 2023 on oil, the ten-year, HYG − LQD and the dollar, with no market factor, so the index's own co-movement with yields in that window is inside it) times the −106bp still to come, with oil, HYG − LQD and the dollar held fixed and no horizon; here nothing is held fixed and the market's response is the engine's correlation at each horizon.
About this exhibit
Earnings under nominal-GDP scenarios
Where the model takes EPS over eight quarters under stated nominal GDP growth - conditional arithmetic, not forecastsExhibit 10 · EPS growth under each nominal GDP growth rate, against consensus
About this exhibit
Exhibit 11 · Calendar-year EPS growth: the regression's range, the model and consensus
| Item | Value | Meaning |
|---|---|---|
| Low | +5.6% | 4% nominal GDP scenario, Q4 2027 |
| The regression's mid | +14.1% | 5% scenario (the regression's path at Q4 2027) named paths, no odds |
| High | +22.6% | 6% scenario |
| Out-of-sample RMSE | 21.8pp | headline model, Q1 2022 – Q2 2026 out-of-sample record |
| The model's expectation (the fan) | +4.7% | the model's fan, carried to EPS: median, 10th–90th −9.3 to +21.8% past misses |
| Consensus, calendar 2027 | +15.1% | FactSet, quoted 11 Sept 2026 |
| Gap, mid − consensus | −1.0pp | inside ± the out-of-sample error |
About this exhibit
What moves that hurt together are worth, against independence
What the model assigns to a few combinations of moves at once, and what independence alone would have given
Exhibit 12 · How often paired moves happen together by horizon, against independence
| Event | Engine | 13 weeks | 26 weeks | 52 weeks | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| P(both) | with mean | P(A) · P(B) | if independent | P(both) | with mean | P(A) · P(B) | if independent | P(both) | with mean | P(A) · P(B) | if independent | ||
| WTI crude (USO) +20% and S&P 500 −5% | spine | 5.2% | 2.9% | 18 · 27 | 4.8% | 7.6% | 3.3% | 25 · 33 | 8.3% | 10.8% | 2.8% | 32 · 38 | 12.2% |
| TLT −5% and S&P 500 −5% (bonds fail as a hedge) | spine | 4.9% | 1.6% | 21 · 27 | 5.6% | 8.2% | 1.8% | 28 · 33 | 9.4% | 11.7% | 1.4% | 35 · 38 | 13.3% |
| Gold +10% and dollar (UUP) +3% | spine | 0.9% | 3.0% | 20 · 18 | 3.7% | 2.5% | 10.4% | 26 · 26 | 6.9% | 4.7% | 24.0% | 31 · 33 | 10.3% |
| 10-year +50bp and S&P 500 −10% | yields | 1.3% | 0.6% | 13 · 14 | 1.8% | 3.1% | 1.0% | 20 · 21 | 4.1% | 6.2% | 1.2% | 27 · 28 | 7.8% |
| 10-year −50bp and S&P 500 +5% | yields | 3.0% | 5.6% | 13 · 30 | 3.9% | 5.4% | 12.1% | 20 · 35 | 7.1% | 8.6% | 18.9% | 27 · 39 | 10.7% |
| WTI crude (USO) +20% and S&P 500 −5% | yields | 6.3% | 3.6% | 20 · 29 | 5.8% | 8.6% | 3.9% | 26 · 34 | 8.9% | 11.4% | 3.1% | 32 · 39 | 12.4% |
| Gold +10% and dollar (UUP) +3% | yields | 0.9% | 3.0% | 21 · 18 | 3.9% | 2.5% | 10.0% | 27 · 26 | 7.1% | 4.6% | 23.5% | 32 · 33 | 10.6% |
About this exhibit
Factor paths
The simulated factors, with no scenario and under each path
Exhibit 13 · Simulated moves in the market factors, with and without a path
Yield factors, basis points
About this exhibit
The paths that are not projected
Every named path whose 10-year has no signed, dated source, and why the formula cannot price it either
Exhibit 14 · The paths that are not projected
- The ten-year up by the term premium's last three months
The expected path of short rates stays where it is and the ten-year rises by as much as the ACM term premium rose over the last three months — compensation for duration repricing again, not the Committee. The adverse branch for bonds: the only one in the set in which yields go up.
Not projected: no decision by the author.
- Hollow deal
Washington blesses the Iran–Oman corridor and keeps the blockade; a reopening by regional consent without a US–Iran agreement (Brent $80–95 in the grid).
Not projected: its pin would be read from the author's signed scenario file for the Hormuz thesis, which has not been written yet; and the valuation identity cannot price an oil branch either — no elasticity of index earnings or of the premium to oil exists among the site's estimates, so oil is not modelled. The pin would be the author's signed scenario file's “hollow” oil level.
- Attrition
Both sides hold to 3 November: transits 5–20 a day, shut-ins of 5.7 mb/d through 4Q (Brent $100–115).
Not projected: its pin would be read from the author's signed scenario file for the Hormuz thesis, which has not been written yet; and the valuation identity cannot price an oil branch either — no elasticity of index earnings or of the premium to oil exists among the site's estimates, so oil is not modelled. The pin would be the author's signed scenario file's “attrition” oil level.
- Escalation
A mistake path — a Pickaxe strike on IAEA-verified movement, Gulf energy infrastructure, or an Iranian missile that hits a US ship (Brent above $120).
Not projected: its pin would be read from the author's signed scenario file for the Hormuz thesis, which has not been written yet; and the valuation identity cannot price an oil branch either — no elasticity of index earnings or of the premium to oil exists among the site's estimates, so oil is not modelled. The pin would be the author's signed scenario file's “escalation” oil level.
- Genuine deal
A sequencing device is found — an escrow or custodian for the HEU that makes Iran's irreversible concession simultaneous with America's reversible one.
Not projected: the log states no Brent level for this branch, so it is listed and not projected.
No path carries odds: none of the 6 theme logs holds a dated row of judgement odds written by the author, so the set is printed as paths with their effects and no central path is claimed.
About this exhibit
The joint model's bands, held in a fold until their record passes
The conditional BVAR's 1- and 4-quarter bands on each path, and the coverage record that keeps them here
Exhibit 15 · The macro variables and the 10-year under each Fed and Hormuz path
| Path | Conditions at 4q | Nonfarm payrolls%, change from today | Unemployment ratepp | Core PCE prices%, change from today | CPI, all items%, change from today | Industrial production%, change from today | 10-year Treasurypp |
|---|---|---|---|---|---|---|---|
| Hold paththe policy steps of the path What is priced defines (no move); WTI held | funds rate 3.75%crude ±0% (log) | +3.11+0.88 to +5.34 simulated shocks | +2.96+1.38 to +4.55 simulated shocks | +3.18+2.35 to +4.00 simulated shocks | +4.11+2.90 to +5.33 simulated shocks | +3.54−0.31 to +7.40 simulated shocks | +4.00+3.18 to +4.82 simulated shocks |
| SEP paththe policy steps of the path What is priced defines (+25bp in quarter 1); WTI held | funds rate 4.00%crude ±0% (log) | +3.26+1.03 to +5.49 simulated shocks | +2.89+1.30 to +4.47 simulated shocks | +3.19+2.37 to +4.02 simulated shocks | +4.15+2.93 to +5.36 simulated shocks | +3.64−0.21 to +7.50 simulated shocks | +4.04+3.22 to +4.86 simulated shocks |
| Above the dotsthe policy steps of the path What is priced defines (+50bp in quarter 1); WTI held | funds rate 4.25%crude ±0% (log) | +3.42+1.19 to +5.65 simulated shocks | +2.81+1.22 to +4.40 simulated shocks | +3.21+2.39 to +4.04 simulated shocks | +4.18+2.97 to +5.40 simulated shocks | +3.75−0.11 to +7.60 simulated shocks | +4.08+3.26 to +4.90 simulated shocks |
| Hollow dealBrent at $87.5, the mid-point of range the log states (2026-09-13), from the September report's locked $104.61; the funds rate held | funds rate 3.75%crude −18% (log) | +3.18+0.95 to +5.41 simulated shocks | +2.87+1.28 to +4.45 simulated shocks | +2.96+2.14 to +3.79 simulated shocks | +3.44+2.22 to +4.65 simulated shocks | +3.50−0.36 to +7.35 simulated shocks | +3.86+3.04 to +4.68 simulated shocks |
| AttritionBrent at $107.5, the mid-point of range the log states (2026-09-13), from the September report's locked $104.61; the funds rate held | funds rate 3.75%crude +3% (log) | +3.10+0.87 to +5.33 simulated shocks | +2.98+1.39 to +4.57 simulated shocks | +3.21+2.38 to +4.04 simulated shocks | +4.21+3.00 to +5.43 simulated shocks | +3.55−0.31 to +7.40 simulated shocks | +4.02+3.20 to +4.84 simulated shocks |
| EscalationBrent at $120, the floor of the open range the log states (2026-09-13), from the September report's locked $104.61; the funds rate held | funds rate 3.75%crude +14% (log) | +3.06+0.83 to +5.29 simulated shocks | +3.04+1.45 to +4.63 simulated shocks | +3.34+2.51 to +4.17 simulated shocks | +4.63+3.41 to +5.85 simulated shocks | +3.58−0.28 to +7.43 simulated shocks | +4.10+3.28 to +4.92 simulated shocks |
The coverage record (thresholds: Kupiec p ≥ 0.05, independence p ≥ 0.05, at least 20 origins)
| Variable · horizon | Inside | Kupiec p | Independence p | Passes |
|---|---|---|---|---|
| Nonfarm payrolls · 1q | 81 of 103 (79%) | 0.00 | 0.00 (103) | no |
| Nonfarm payrolls · 4q | 65 of 103 (63%) | 0.00 | 0.04 (26) | no |
| Unemployment rate · 1q | 73 of 103 (71%) | 0.00 | 0.00 (103) | no |
| Unemployment rate · 4q | 62 of 103 (60%) | 0.00 | 0.51 (26) | no |
| Core PCE prices · 1q | 74 of 103 (72%) | 0.00 | 0.02 (103) | no |
| Core PCE prices · 4q | 58 of 103 (56%) | 0.00 | 0.16 (26) | no |
| CPI, all items · 1q | 67 of 103 (65%) | 0.00 | 0.02 (103) | no |
| CPI, all items · 4q | 52 of 103 (50%) | 0.00 | 0.62 (26) | no |
| Industrial production · 1q | 81 of 103 (79%) | 0.00 | 0.07 (103) | no |
| Industrial production · 4q | 69 of 103 (67%) | 0.00 | 0.56 (26) | no |
| 10-year Treasury · 1q | 80 of 103 (78%) | 0.00 | 0.56 (103) | no |
| 10-year Treasury · 4q | 58 of 103 (56%) | 0.00 | 0.19 (26) | no |
Model-conditional: a conditional forecast's bands are the plug-in Gaussian quantiles of the model around the stated paths; parameter uncertainty is not integrated out, and the record is what says whether that matters.
About this exhibit
What an oil move and a 10-year move do, with credit and the dollar held
What an oil move (rows) and a 10-year move (columns) do over the horizon, with credit and the dollar held unchanged: the point and its 10th–90th band; and the valuation formula's row
Exhibit 16 · The selected asset's change for each oil and 10-year move
| Oil \ 10-year | −75bp | −50bp | −25bp | ±0bp | +25bp | +50bpnow | +75bp | +100bp |
|---|---|---|---|---|---|---|---|---|
| −40% | +8.2%-0.2 to 16.5 | +5.5%-2.6 to 13.5 | +2.8%-5.2 to 10.7 | +0.1%-7.8 to 8.0 | −2.6%-10.7 to 5.5 | −5.3%-13.6 to 3.0 | −8.0%-16.7 to 0.7 | −10.7%-19.8 to -1.5 |
| −30% | +8.1%-0.2 to 16.5 | +5.5%-2.6 to 13.5 | +2.8%-5.1 to 10.7 | +0.1%-7.8 to 8.0 | −2.6%-10.6 to 5.4 | −5.3%-13.6 to 2.9 | −8.0%-16.6 to 0.6 | −10.7%-19.8 to -1.6 |
| −20%Hollow deal | +8.1%-0.2 to 16.5 | +5.4%-2.6 to 13.5 | +2.7%-5.1 to 10.6 | ±0.0%-7.8 to 7.9 | −2.6%-10.6 to 5.3 | −5.3%-13.5 to 2.8 | −8.0%-16.6 to 0.5 | −10.7%-19.7 to -1.7 |
| −10% | +8.1%-0.3 to 16.5 | +5.4%-2.7 to 13.5 | +2.7%-5.2 to 10.6 | ±0.0%-7.8 to 7.9 | −2.7%-10.6 to 5.2 | −5.4%-13.5 to 2.8 | −8.1%-16.5 to 0.4 | −10.7%-19.6 to -1.8 |
| ±0%Attrition | +8.1%-0.3 to 16.5 | +5.4%-2.7 to 13.5 | +2.7%-5.2 to 10.6 | ±0.0%-7.8 to 7.8 | −2.7%-10.6 to 5.2 | −5.4%-13.5 to 2.7 | −8.1%-16.5 to 0.3 | −10.8%-19.6 to -1.9 |
| +10%now +14%, +47bp | +8.1%-0.4 to 16.5 | +5.4%-2.8 to 13.5 | +2.7%-5.2 to 10.6 | ±0.0%-7.9 to 7.8 | −2.7%-10.6 to 5.2 | −5.4%-13.5 to 2.7 | −8.1%-16.5 to 0.3 | −10.8%-19.6 to -2.0 |
| +20%Escalation | +8.0%-0.5 to 16.6 | +5.3%-2.8 to 13.5 | +2.6%-5.3 to 10.6 | ±0.0%-7.9 to 7.8 | −2.7%-10.6 to 5.1 | −5.4%-13.5 to 2.6 | −8.1%-16.5 to 0.2 | −10.8%-19.6 to -2.1 |
| +30% | +8.0%-0.6 to 16.6 | +5.3%-2.9 to 13.6 | +2.6%-5.4 to 10.6 | −0.1%-8.0 to 7.8 | −2.8%-10.7 to 5.1 | −5.5%-13.5 to 2.6 | −8.1%-16.5 to 0.2 | −10.8%-19.5 to -2.1 |
| +40% | +8.0%-0.7 to 16.7 | +5.3%-3.0 to 13.6 | +2.6%-5.5 to 10.7 | −0.1%-8.0 to 7.8 | −2.8%-10.7 to 5.2 | −5.5%-13.5 to 2.6 | −8.2%-16.5 to 0.2 | −10.9%-19.6 to -2.2 |
| +60% | +7.9%-1.0 to 16.9 | +5.2%-3.3 to 13.8 | +2.5%-5.7 to 10.8 | −0.1%-8.2 to 7.9 | −2.8%-10.9 to 5.2 | −5.5%-13.7 to 2.6 | −8.2%-16.6 to 0.1 | −10.9%-19.6 to -2.2 |
Bold: |point| exceeds the half-band (the betas can tell the scenario from noise). Oil changes are log changes in USO; the 10-year change is in the Treasury par yield.
About this exhibit
Exhibit 17 · The selected asset's change by 10-year move, at the selected oil move
About this exhibit
Every asset at the cell
Every asset at the chosen cell of the surface, with its 10th–90th band and whether the point exceeds it
Exhibit 18 · Every asset's move at the selected oil and 10-year move
| Asset | Point | 10th | 90th | Half-band | Beta share of variance | Exceeds band |
|---|---|---|---|---|---|---|
| Index | ||||||
| S&P 500SPY | −5.4% | −13.5% | +2.6% | 8.0 | 6% | no |
| S&P 500 equal weightRSP | −6.4% | −13.4% | +0.6% | 7.0 | 6% | no |
| Nasdaq-100QQQ | −6.0% | −17.7% | +5.7% | 11.7 | 5% | no |
| Russell 2000IWM | −9.9% | −20.3% | +0.6% | 10.4 | 3% | no |
| Sectors | ||||||
| TechnologyXLK | −6.5% | −21.3% | +8.2% | 14.7 | 5% | no |
| Communication servicesXLC | −4.1% | −14.0% | +5.8% | 9.9 | 5% | no |
| Consumer discretionaryXLY | −8.2% | −19.3% | +2.8% | 11.0 | 4% | no |
| Consumer staplesXLP | −3.2% | −10.3% | +4.0% | 7.1 | 6% | no |
| Health careXLV | −3.6% | −13.1% | +5.9% | 9.5 | 5% | no |
| FinancialsXLF | −6.6% | −15.8% | +2.6% | 9.2 | 6% | no |
| IndustrialsXLI | −7.1% | −15.7% | +1.6% | 8.7 | 5% | no |
| EnergyXLE | +3.9% | −8.4% | +16.3% | 12.4 | 7% | no |
| MaterialsXLB | −6.4% | −15.9% | +3.1% | 9.5 | 6% | no |
| UtilitiesXLU | −4.5% | −13.7% | +4.8% | 9.3 | 5% | no |
| Real estateXLRE | −8.2% | −16.3% | −0.1% | 8.1 | 4% | yes |
| Industries | ||||||
| SemiconductorsSMH | −8.3% | −29.3% | +12.7% | 21.0 | 4% | no |
| Oil & gas E&PXOP | +6.8% | −7.4% | +21.0% | 14.2 | 6% | no |
| Regional banksKRE | −10.8% | −26.1% | +4.4% | 15.2 | 5% | no |
| HomebuildersXHB | −14.0% | −28.1% | +0.2% | 14.1 | 2% | no |
| TransportsIYT | −8.5% | −19.6% | +2.5% | 11.0 | 5% | no |
| RetailXRT | −8.3% | −20.2% | +3.7% | 11.9 | 4% | no |
| REITsVNQ | −8.4% | −16.3% | −0.5% | 7.9 | 5% | yes |
| Style spreads | ||||||
| Growth − valueGRW-VAL | −0.9% | −11.3% | +9.4% | 10.3 | 4% | no |
| Momentum − S&P 500MOM | −2.2% | −10.3% | +6.0% | 8.1 | 4% | no |
| Quality − S&P 500QUAL | +0.1% | −2.3% | +2.4% | 2.3 | 4% | no |
| Small − largeSMALL | −4.4% | −12.0% | +3.1% | 7.6 | 3% | no |
| Equal − cap weightBREADTH | −0.9% | −6.0% | +4.1% | 5.1 | 4% | no |
| Low vol − S&P 500LOWVOL | +1.6% | −4.4% | +7.5% | 6.0 | 3% | no |
| Cyclicals − defensivesCYC-DEF | −3.2% | −11.4% | +4.9% | 8.1 | 4% | no |
| AI complex − defensivesAI-DEF | −4.2% | −23.0% | +14.6% | 18.8 | 4% | no |
| Pair | ||||||
| Energy − Consumer discretionaryXLE − XLY | +12.2% | −4.4% | +28.8% | 16.6 | 5% | no |
The pair's variance adds both covariance matrices and both residual variances as if the two regressions' errors were independent (their residual correlation over 156 weeks is +0.06).
About this exhibit
What actually happened last quarter
Where the last quarter sits: what crude and the 10-year actually did over 13 weeks, and whether the index landed inside the band
Exhibit 19 · Rolling 13-week change in crude oil and the 10-year yield
About this exhibit
Appendix
How the models are specified and simulated, what the coverage backtest scores, and every source behind the page249 checks, all pass
The volatility model: GARCH(1,1) with Student-t shocks per market, a scalar DCC across them
| Symbol | Meaning | From | Assumed or estimated |
|---|---|---|---|
| daily log return × 100 (a yield's daily change in basis points) | Yahoo Finance adjusted closes; U.S. Treasury par yields | – | |
| the conditional variance of market j | filtered with the fitted parameters (arch's recursion, backcast start) | – | |
| , , , , | mean, variance intercept, shock and persistence weights, the t's degrees of freedom | maximum likelihood per series | estimated |
| , , | the correlation recursion's weights and its long-run matrix | quasi-maximum likelihood on the standardised residuals | estimated |
| independent Student-t draws scaled to unit variance | simulated | – |
Two engines: the core-markets fit (3,058 sessions from 18 Sep 2014; a = 0.0180, b = 0.9660), reused from the regime page without re-estimation, and the yields fit (3,213 sessions from 3 Jan 2014; a = 0.0213, b = 0.9647). The parameters are estimated once on the whole sample; the state (h, Q) is filtered to each date with them.
The cone, its record and the two coverage tests
| Symbol | Meaning | From | Assumed or estimated |
|---|---|---|---|
| the cumulative move over H = 65, 130 or 260 sessions on one simulated path (20,000 paths; the record 5,000) | simulated from the fitted model | – | |
| zero drift: the band measures width, not direction | assumed (the fitted mean is an option) | assumed | |
| the realised 65-session move from quarter-end q | Yahoo Finance; U.S. Treasury | – | |
| the market's GARCH(1,1)-t refitted on its data from 2 Mar 2007 to quarter-end q only — the out-of-sample record | maximum likelihood at each quarter-end | estimated | |
| , | the miss rate the band promises (1 − its coverage) and the rate observed, x misses in n windows | the record | – |
| , , | transitions from a hit (0) or a miss (1) to the next window's, and the miss rates after each | the record | – |
Model-conditional, not a forecast. The engine's own record uses the full-sample parameters (in-sample, printed beside); the refitted record uses θ̂ from data to each quarter-end only. For the S&P 500, the first series in the core-markets engine, the univariate refit is exactly the engine's marginal; for the 10-year and WTI the engine mixes t draws through the correlation factor, which keeps each variance recursion and changes only the shape. 46 windows: the χ² p-values are asymptotic.
Exhibit 20 · Specification
How the numbers are made, step by step
- Prices and yields. Source: Yahoo Finance daily closes; Treasury par yields (2- and 10-year). Transform: daily returns and yield changes on the sessions every series printed. Check: the backtest's outcomes equal the raw files (3 checks, all passing).
- The two engines. Source: the ten-core-markets fit the regime page cached (not refitted) and a yields fit here. Transform: GARCH(1,1)-t per series and a scalar DCC on the standardised residuals, filtered to today's state. Check: the filtered states reproduce the cached engine; stationarity of both recursions (5 checks, all passing).
- The cones. Source: step 2. Transform: core.cones: 20,000 paths from today's state at 65, 130 and 260 sessions, zero drift by default. Check: the simulator's variance identities; unit-variance t draws; quantiles monotone (7 checks, all passing).
- Joint events. Source: step 3. Transform: the share of paths in which both legs hold. Check: no joint probability exceeds either marginal (1 check, all passing).
- Coverage, two records. Source: step 2 at every quarter-end since 2015, and each asset refitted on the data to that date. Transform: the band drawn at the time against the outcome that followed; Kupiec's and Christoffersen's tests. Check: coverage nested across the 50, 80 and 90% bands; every complete quarter counted; the refitted record re-scores the same windows and fits nothing after a quarter-end (12 checks, all passing).
- SpecificationModel-conditional, zero drift, full-sample parameters for the cones.
The cones are model-conditional distributions, not forecasts: they are what the fitted engine implies if its parameters and its reading of today's variance and correlation state are right, with zero drift. Engine: GARCH(1,1) with Student-t innovations per series on daily returns ×100 (arch), DCC(1,1) on the standardised residuals by quasi-maximum likelihood — the ten-core-markets fit (the spine in earlier issues) is the cached 17 Sep run of the regime page (a = 0.0180, b = 0.9660, 3,058 sessions from 18 Sep 2014), reused without re-estimation (its coverage record is therefore in-sample in its parameters — the page's out-of-sample record refits each quarter); the yields engine is the same code on daily changes of the 2- and 10-year par yields (bp) and log returns of SPY, USO, GLD and UUP on 3,213 common sessions from 3 Jan 2014 (a = 0.0213, b = 0.9647).
- Simulation20,000 paths, 260 sessions, fixed seeds.
From the last filtered state — h_(T+1) = ω + α ε_T² + β h_T per series and Q_(T+1) = (1 − a − b) Q̄ + a z_T z_T′ + b Q_T, both from the fitted parameters run through the data — 20,000 paths of 260 sessions: each session draws independent Student-t innovations with the fitted ν, scaled to unit variance, correlates them with the Cholesky factor of R_t = diag(Q_t)^−½ Q_t diag(Q_t)^−½, scales by √h_t, then updates h and Q with the same recursions. Quantiles 5/10/25/50/75/90/95 of the cumulative simple return (yields: cumulative change in bp) at 65, 130 and 260 sessions (13, 26 and 52 weeks) and every fifth session for the chart. Seeds are fixed, so a rerun on the same inputs reproduces every number.
- DriftZero by default; the fitted mean as an option.
Zero drift is the default: the median path is flat and the cone is pure variance. The option adds the fitted constant mean μ of each GARCH to every session; because the variance recursion runs on the residual, the same paths serve both settings shifted by μ per session. The fitted means are sample averages over 2014–26 and carry no information about the coming quarter; they are shown so the size of the assumption is visible.
- Joint eventsShares of paths.
Joint probabilities are the share of the paths in which both legs hold at the horizon; the marginals and their product (independence) sit beside each so the part the correlation adds is visible. Events are computed in every engine that carries both assets. They are the fitted engine's probabilities under its own assumptions, not a judgement of the site's.
- CoverageQuarter-ends since 2015: refitted each quarter, and the engine's full-sample record beside it.
Two coverage records, one set of windows. At every calendar quarter-end from 2015 Q1 to the last quarter whose 65-session outcome is complete (2026 Q2; 46 quarter-ends) a band is drawn from the state at the last session on or before the quarter-end (5,000 paths of 65 sessions, zero drift) and the realised outcome (adjusted-close ratio; par-yield difference) is scored inside or outside the 50% (25–75), 80% (10–90) and 90% (5–95) bands. The out-of-sample record — the page's claim — re-estimates each asset's GARCH(1,1)-t at every quarter-end on its daily data from 2 Mar 2007 to that date only (an expanding window; the S&P 500, the 10-year and WTI) and draws the band from that fit. The engine's own record draws it from the full-sample parameters with the state filtered to the date: the parameters have seen the shocks they are scored on, so it is an in-sample-parameter record and is printed as that, never as out-of-sample. The same outcomes are also scored against the engine's bands shifted by the fitted mean (the drift option), to show how much of a central-band shortfall is the zero-drift centre rather than the width. The latest quarter-end's outcome is still running and is shown as pending, not scored.
- TestsKupiec and Christoffersen.
Kupiec's unconditional-coverage test compares the miss rate with the band's promise (10% outside a 90% band) by a likelihood ratio, χ² with one degree of freedom; Christoffersen's independence test asks whether a miss makes the next window's miss more likely (a first-order Markov chain against independence, χ²(1)), and the two add to the conditional-coverage test, χ²(2). With 46 windows the asymptotic p-values are coarse, and adjacent 65-session windows share a session or two, so they are close to but not exactly independent. A p-value at or above 0.05 is read as consistent with the nominal rate — not as proof that the width is right.
- SensitivityDrift, tails, correlation, engine.
Sensitivity of the 13-week edges: zero drift (default) vs the fitted mean; Student-t vs normal innovations at the same variances (a unit-variance t(ν) puts more mass near the centre and only exceeds the normal beyond about 2.5σ, so the 5–95 edges are slightly wider with normals while the far tails are thinner); the fitted DCC vs constant correlation Q̄ (a = b = 0; affects the joint events more than the marginal cones). The two engines' cones for the assets they share are a further specification sensitivity.
- CalendarTwo session calendars.
Core-market sessions are the days on which all ten core markets, bitcoin included, printed a close (the regime page's convention). Yields-engine sessions are the common days of the Treasury par curve and NYSE trading (Columbus Day and Veterans Day drop out, so the next change spans two days). A 65-session horizon is about 13 weeks on either calendar.
- LicenceNo level shown.
Prices are Yahoo Finance data, shown with their attribution; Treasury par yields are the Treasury's own. The engines, the simulation and the backtest are computed here.
- RuntimeNot printed.
Runtime is not printed: timings differ from run to run, and the page must print the same numbers whenever it is recomputed from the same files. The simulation's size is its paths and horizons above.
Exhibit 21 · How a path is priced
- What this is
Model-conditional. Every number on this page is what one named model says under stated inputs, and none is a forecast. The index-level answer to a branch is the Valuation page's two-stage identity with the pinned ten-year and everything else held; the distributions are the factor engine's — GARCH(1,1)-t per raw factor with a scalar DCC, simulated forward with zero drift from today's filtered state, mapped to each asset through betas drawn from their sampling distribution, with the asset's residual resampled from its own history — whose median is an artefact of the zero drift rather than a central case.
- The identity
The valuation identity, per branch (D7). Price over trailing earnings = k [(1 + g_L) + H (g_S − g_L)] ÷ (r − g_L), the Valuation page's two-stage identity, at its priced inputs — k 0.84, E0 ÷ P 3.985%, g_S 15.1% (the consensus) fading to g_L 4.0% over 6 years, r = rf 5.12% (16 Nov 2026) + ERP 3.502pp (the premium the price implies) — which return today's close (a check). A branch moves rf by its ten-year pin and holds the rest, so its move from today is (r0 − g_L) ÷ (r1 − g_L) − 1 (a check on every branch), printed as a percentage of today's price and never as an index level. The model's value under the branch is the same identity at the expected inputs — g_S 5.77% (the Earnings page's fan bridged to EPS) and ERP 5.170pp (the yield-gap premium) — with the same rf, printed as a percentage of today's price, the one hold-path gap every page copies; at a pin of zero it is that page's gap (−41.5%, a check). Held, not estimated: the premium's response to the ten-year, to spreads or to volatility, and earnings' response to growth or oil — no such elasticity exists among the site's estimates, so oil, credit and the dollar are printed as not modelled rather than given a number, and the two Hormuz branches with a Brent range cannot be priced through the identity until one is estimated. Every input is copied from the Valuation page by its published key and re-compared on each update.
- Market pricing
Market pricing beside the set is copied from What is priced (16 Nov 2026) by its published keys: the ten-year, each path's fair value and remaining move, the forwards against IORB, the SEP medians, that page's completion reading and its own equity amount on the SEP path. It is the observed counterpart the scenario set is read against; nothing here is recomputed from it.
- Conditioning
A branch pins a subset of the raw factors at the horizon. Each draw's unpinned factors become Q* = Q − B (P − x_P) with B estimated on those draws. For jointly Gaussian draws this is exact conditioning; for these fat-tailed draws it is a shift that keeps each draw's own residual, so the unpinned factors keep their tails instead of collapsing onto a mean. With nothing pinned the draws are unchanged, bit for bit, and a check says so.
- The scenario set
Every pin is transcribed from a dated source and re-read on each update, and printed in the source's own unit. A branch whose definition names a direction is projected only while its source still moves that way. A branch is projected only on a recorded decision — the author's decision of 3 Oct 2026 (D-2026-10-03-07); the Fed paths stand on the author's decision of 26 Sep 2026 (D-2026-09-26-PATHS). A branch with none is listed with its source and not projected — today that is the ten-year up by the term premium's last three months, hollow deal, attrition, escalation, genuine deal. The Hormuz branches' keys and names are the Hormuz log's own, read from it on every update. No branch carries a probability: the set is paths only, as decided on 26 Sep 2026 (the Hormuz page keeps its dated log), and no joint probability is computed anywhere.
- The mapping
R = cash + Σ β·F̃ + ε, with β drawn afresh on every path. The second regression's constant is not carried: it is an in-sample average of what the factors did not explain, and adding it would assert that the same average continues. An asset that *is* a factor proxy takes its own raw factor draw and no cash leg. An asset attribution excludes from a factor because it is a leg of it gets that factor back, raw, here — and its beta is applied to the raw draw, the variable it was estimated on.
- Parameter draws
Parameter uncertainty: each path draws its own betas as β + se·z, z standard normal, from the Newey–West standard errors the Return attribution page publishes for its second regression (the projection regression's own, by the same estimator, for an asset with a factor restored). Only standard errors are published, not the coefficients' covariance, so the draws are independent across factors, and a Dimson exposure's error is its two lags' errors in quadrature. A path keeps its draw at every horizon.
- The residuals
Each asset's own second-regression residual, resampled by the stationary block bootstrap (Politis–Romano): blocks of 21 sessions on average, of geometric length, drawn from the asset's whole residual history centred on zero, since the constant is not carried. Assets are resampled independently, which changes no number printed here — every band is one asset's. The alternative printed beside it simulates the residual with its own GARCH(1,1)-t, the 11 sector residuals drawn together with today's residual correlation from the Correlations & regime page and every other asset's independently.
- Calibration
A width scale about the median is estimated on the first 9 quarter-ends of the backtest, to 31 Mar 2024. It is the smallest scale that would have held 80% of outcomes there: k = 1.0721 for the 10–90 band. It is then tested, unchanged, on the 9 quarter-ends after. Under the author's decision of 24 Oct 2026 (D-2026-10-24-PJ2), the 5–95 band is the widened 10–90 band plus a tail margin that cannot be negative. The margin is the distance the simulation puts between the 10th and 5th percentiles, and between the 90th and 95th, times a tail factor of 0.4590. That factor is the smallest that would have held 90% of the same outcomes in the 5–95 band. Because it cannot be negative, the 5th and 95th can never cross the 10th and 90th. The 5th and 95th percentiles changed method on 24 Oct 2026: before, each was scaled about the median by a factor of its own. The scale is needed only when the uncalibrated band's pooled 10–90 interval excludes 80%. It is applied only when the widened 10–90 band then passes its held-out test. A scale that fails its own test is printed as attempted and not drawn. When applied, every published 10th and 90th percentile is p50 + k·(q − p50). Every 5th and 95th is that value plus k_tail·(q05 − q10) or k_tail·(q95 − q90), and the median and quartiles are as simulated. Today: needed, and the widened 10–90 band passes its held-out test, so it is applied. The bands head Distributions. The alternative (GARCH-t residual) is shown before any calibration.
- Licence
Prices are Yahoo Finance data and Treasury yields are free to use; every other input is another page's published number, with that page's own sources behind it.
Exhibit 22 · Definitions
How the numbers are made, step by step
- Prices and factors. Source: Yahoo Finance ETF closes; Treasury par yields; the Board's H.10 dollar. Transform: weekly log returns; the four factors (oil, the 10-year per 10bp, credit, the dollar); the factors page's weekly data. Check: the weekly data is the factors page's, to its last week (3 checks, all passing).
- The regressions. Source: step 1. Transform: each asset on the four factors over 156 and 52 weeks, Newey–West covariance (4 lags). Check: betas and covariances equal statsmodels and the factors page's estimator (3 checks, all passing).
- The grid and its bands. Source: step 2 and the stated moves (oil rows, 10-year columns, horizons). Transform: point = Σ β f; variance = fᵀ Σ_β f + H σ²; band = point ± 1.2816 √variance, at every cell, asset, window and horizon the page offers. Check: a Monte Carlo of the betas and the residual reproduces the band; the pair is energy minus discretionary; the published grid reproduces the example cells and the report's daily construction (8 checks, all passing).
- The valuation route. Source: the Valuation page's two-stage identity and its inputs (copies). Transform: the identity at each stated 10-year, everything else held. Check: at the priced inputs the identity returns today's price; each column re-run through the shared identity (7 checks, all passing).
- The September issue, reproduced. Source: the September review's quoted constants and printed cells. Transform: the September review's daily betas, TLT mapping and base row recomputed from the prices. Check: the betas, the mapping and the base-row levels reproduced (4 checks, all passing).
- Copies from other pages. Source: the other pages' published data, through the site's manifest. Transform: each value read from the published version of its page. Check: each source published with its checks passing; every copy equal to its source (17 checks, all passing).
- ModelFour factors, weekly, no market.
Route A: weekly log returns (%) of each asset regressed over 156 and 52 weeks on four factors — crude (USO), the change in the Treasury 10-year par yield per 10bp, high yield over investment grade (HYG − LQD) and the Board's broad dollar — without the market, so each beta is a total effect (it includes what reaches the asset through the index). Newey–West covariance (4 lags, Bartlett, no small-sample correction). Windows end on the last complete week, 13 Nov 2026; the same weekly data as the Factors page (a check).
- GridPoint and band.
A cell is a pair of factor moves over the horizon H: oil in per cent, the 10-year in basis points (÷10 to the beta's unit); credit and the dollar are held at zero (the grid's stated moves are oil and the 10-year). Point = Σ β_i f_i; variance = fᵀ Σ_β f + H σ² where Σ_β is the betas' covariance and σ the weekly residual standard deviation; the 10th–90th band is point ± 1.2816 √variance. The band assumes the factor moves are given (a scenario, not a forecast) and normal residuals; it is model-conditional. A cell 'exceeds its band' when |point| is larger than the half-band. Every cell the page prints — each asset and the pair, the 156- and 52-week windows and the report's daily one, each horizon, each oil row and 10-year column — is computed here with its band, its half-band, the beta and residual parts and the beta share of the variance; checks hold the published grid to the example cells and to the report's daily construction. A Monte Carlo checks the band's arithmetic: 200,000 draws of the betas from their covariance plus a normal residual over the horizon (seed 20260917) put the 10th and 90th percentiles of 3 example cells within 0.2pp of the band's edges.
- PairEnergy − discretionary.
The energy − discretionary pair is energy's cell minus discretionary's; its variance adds the two covariance matrices and the two residual variances as if the regressions' errors were independent. Their residuals are correlated (+0.06 over 156 weeks), so the printed band is wider than a direct regression of the pair would give.
- Route BThe valuation route.
The valuation route (the Valuation tab) runs the 10-year column through the Valuation page's two-stage identity — price over trailing earnings = k [(1 + g_L) + H (g_S − g_L)] ÷ (r − g_L) — at that page's priced inputs (k 0.84, E0 ÷ P 3.985%, g_S 15.1% fading to g_L 4.0% over 6 years, r = 5.12% + 3.502pp), moving rf by the column and holding the rest, so a column's move from today is (r0 − g_L) ÷ (r1 − g_L) − 1, printed as a percentage of today's price, never as an index level; beside it, the model's value at the expected inputs (g_S 5.77%, ERP 5.170pp) and the column's 10-year, as a percentage of today's price — the one hold-path gap the board and the Valuation page print. Model-conditional: what the identity implies under stated inputs. It is one row across the 10-year columns, not a grid: the site estimates no elasticity of index earnings or of the premium to oil, credit or the dollar, so those are printed as not modelled rather than given a number (the earlier route's oil elasticity was a value the reader typed, with no preset the site could stand behind). The Scenarios page's branch table runs the same identity for every named scenario, and every input is copied from the Valuation page and re-compared on each update.
- ReportThe September report's construction.
The September review made its target matrix from daily betas of the S&P 500 on USO and TLT since the Strait closed (−0.10, +0.31), a 156-week mapping of TLT's weekly return to the 10-year change (−0.14% per bp, R² 0.91) and the grid's Brent mid-points against $104.61 and cell yields against 4.96%. Its script is not in the repo; the construction is reproduced from its printed description with log changes for the oil mid-points and 7,591.70 × exp(r) for the level, and the cell yields are backed out of the printed base-row levels. The issue prints 'above $120' without a mid-point; 125 is this page's assumption. At 120 the printed escalation level reproduces within 0.1% under every convention tried, so the September review's script most likely used the lower bound.
- OverlayThe Hormuz columns.
The Hormuz log's three branches with a Brent range — hollow deal, attrition, escalation — are marked on the grid at their mid-point changes against $104.61 (the log's revision of 2026-09-14); the genuine deal has no Brent level and is not placed. The marks carry no odds: the set is paths only, and the log's own judgement odds stay in the Hormuz page's revision record. The September review's cell yields (4.70–5.35%) are not in the repo, so the yield range is shaded rather than the nine cells placed.
- LevelsIndex levels.
Index levels on the factor route are 7,591.70 × SPY_t ÷ SPY_11Sep × exp(r) — the September review's close scaled by the ETF's move and the scenario return, computed here beside every cell of the index (a check ties them to the scalar cell); on the valuation route the page prints percentages of today's close only — an index level under a path would read as a target. A check ties the forward P/E × forward EPS convention to the 11 September close. The grid prints percentages because a percentage is what the regressions estimate.
- LicenceVendor prices; official yields and dollar.
ETF prices are Yahoo Finance data, shown with their attribution; Treasury yields and the Board's dollar index are free to use. The regressions, the grid and the reproduction of the September review's construction are computed here.
Exhibit 23 · The fair 10-year of each policy path: 4.07% on a hold, against 5.12% today
| Today's ACM 10-year term premium, +0.73pp | 2015–19 ACM average, −0.29pp | 2023 ACM peak, +0.46pp | Today +50bp, +1.23pp | |
|---|---|---|---|---|
| Hold at IORB (average 3.34%) | 4.07% | 3.05% | 3.80% | 4.57% |
| One further +25bp, then hold (average 3.39%) | 4.12% | 3.10% | 3.85% | 4.62% |
| +50bp, then hold (average 3.44%) | 4.17% | 3.15% | 3.90% | 4.67% |
| Cuts of 25bp a quarter from 2027 Q1 to 3.00 (average 3.21%) | 3.94% | 2.93% | 3.67% | 4.44% |
| Cuts of 25bp a quarter from 2027 Q3 to 3.00 (average 3.26%) | 3.98% | 2.97% | 3.72% | 4.48% |
About this exhibit
How the numbers are made
yahoo + treasury ──▶ daily returns (%) and yield changes (bp) on the sessions every series printed ──▶ the two return sets the regime model's fit + the two return sets ──▶ the cached core-markets fit filtered to today; GARCH(1,1)-t and a scalar DCC fitted on the six yield and ETF series ──▶ the two engines' states the two engines' states ──▶ the volatility model's simulated range: 20,000 paths from today's state at 65, 130 and 260 sessions, zero drift by default ──▶ the cones the cones ──▶ the share of paths in which both legs of each event hold ──▶ the joint events the two engines' states + the two return sets ──▶ the band drawn at every quarter-end since 2015 against the outcome that followed — refitted on the data to each quarter-end, and on the engine's full-sample parameters; Kupiec's and Christoffersen's tests ──▶ the coverage records
Sources
- licensedYahoo Finance chart API, daily adjusted closes (SPY QQQ SMH TLT IEF TIP GLD USO UUP BTC-USD)
Prices: Yahoo Finance. No price level is shown on this page. · https://query2.finance.yahoo.com/v8/finance/chart/
- publicU.S. Department of the Treasury, daily par yield curve rates (2- and 10-year)
Source: U.S. Department of the Treasury · https://home.treasury.gov/resource-center/data-chart-center/interest-rates/daily-treasury-rates.csv/2026/all?type=daily_treasury_yield_curve&field_tdr_date_value=2026&page&_format=csv
- computed hereSite regime engine (DCC-GARCH(1,1)-t), the ten-core-markets fit of 17 Sep 2026 reused without re-estimation; a second fit on yields and four ETFs; Monte Carlo cones and coverage backtest computed here
Model-conditional distributions computed by the site's engine; not forecasts.
Reproduction checks
| Check | Published | Computed | Tolerance | Result |
|---|---|---|---|---|
| Filtered states: standardised residuals from the fitted GARCH parameters (arch's recursion with the parameters fixed, no re-estimation) reproduce the cached ten-core-markets engine's Q̄ = Z′Z/T (max abs difference) | 0 | 0 | 0 | pass |
| … and the DCC recursion from Q̄ reproduces the cached engine's last conditional correlation matrix R_T (max abs difference) | 0 | 0 | 0 | pass |
| Start state: the simulation's SPY variance h_(T+1) = ω + α ε_T² + β h_T by hand from the fitted parameters and the 13 Nov 2026 residual (abs difference) | 2.1413 | 2.1413 | 0 | pass |
| Yields engine: DCC a + b < 1 (a = 0.0213, b = 0.9647; stationarity of the correlation recursion) | 1 | 0.986 | 0 | pass |
| Yields engine: GARCH α + β ≤ 1 for every series (max) | 1 | 1 | 0 | pass |
| Simulator identity (spine): a = b = 0, ν → ∞, constant variance (α = β = 0, ω = h_(T+1)) — the variance of the simulated 65-session sum equals 65 × h for every series (100,000 paths; max relative deviation, %) | 0 | 0.8571 | 2 | pass |
| Simulator identity (spine): a = b = 0, ν → ∞ with the fitted ω, α, β — the 65-session variance equals Σ E[h_t] from E[h_(t+1)] = ω + (α + β) E[h_t] (max relative deviation, %) | 0 | 1.0333 | 2 | pass |
| Simulator identity (yields): a = b = 0, ν → ∞, constant variance (α = β = 0, ω = h_(T+1)) — the variance of the simulated 65-session sum equals 65 × h for every series (100,000 paths; max relative deviation, %) | 0 | 0.4132 | 2 | pass |
| Simulator identity (yields): a = b = 0, ν → ∞ with the fitted ω, α, β — the 65-session variance equals Σ E[h_t] from E[h_(t+1)] = ω + (α + β) E[h_t] (max relative deviation, %) | 0 | 0.7116 | 2 | pass |
| Standardised Student-t innovations have unit variance (ν = 5.44, SPY's fit; 2,000,000 draws; relative deviation, %) | 0 | 0.15 | 2 | pass |
| Quantile arrays are monotone (q05 ≤ q10 ≤ … ≤ q95) at every recorded step, horizon, asset, engine and drift setting (violations) | 0 | 0 | 0 | pass |
| Joint probabilities never exceed either marginal (violations) | 0 | 0 | 0 | pass |
| Backtest outcomes (ten core markets): the realised 65-session SPY return at every quarter-end equals the adjusted-close ratio from the Yahoo raw file (max abs difference, pp) | 0 | 0 | 0 | pass |
| Backtest outcomes (yields): the realised 65-session 10-year change at every quarter-end equals the Treasury par-yield difference in bp (max abs difference) | 0 | 0 | 0 | pass |
| Coverage is nested: share inside the 50% band ≤ 80% band ≤ 90% band for every asset (violations) | 0 | 0 | 0 | pass |
| Coverage backtest counts every quarter-end from 2015 Q1 whose 65-session outcome is complete (ten core markets 46, yields 46; difference) | 0 | 0 | 0 | pass |
| Refitted record (S&P 500): the same 46 windows as the engine's record — its realised 65-session outcome equals that record's at every quarter-end (max abs difference; quarter-ends missing from either counted as 1) | 0 | 0 | 0 | pass |
| Refitted record (S&P 500): out of sample — every one of the 47 fits ends on or before its quarter-end (an expanding window from 2 Mar 2007; fits reaching past it) | 0 | 0 | 0 | pass |
| Refitted record (S&P 500): the tests' miss counts are the windows listed outside each band, and each coverage rate is (n − misses) ÷ n (50/80/90%: 31, 10, 3 misses of 46; sum of abs differences) | 0 | 0 | 0 | pass |
| Refitted record (10-year Treasury yield): the same 46 windows as the engine's record — its realised 65-session outcome equals that record's at every quarter-end (max abs difference; quarter-ends missing from either counted as 1) | 0 | 0 | 0 | pass |
| Refitted record (10-year Treasury yield): out of sample — every one of the 47 fits ends on or before its quarter-end (an expanding window from 2 Mar 2007; fits reaching past it) | 0 | 0 | 0 | pass |
| Refitted record (10-year Treasury yield): the tests' miss counts are the windows listed outside each band, and each coverage rate is (n − misses) ÷ n (50/80/90%: 24, 10, 7 misses of 46; sum of abs differences) | 0 | 0 | 0 | pass |
| Refitted record (WTI crude (USO)): the same 46 windows as the engine's record — its realised 65-session outcome equals that record's at every quarter-end (max abs difference; quarter-ends missing from either counted as 1) | 0 | 0 | 0 | pass |
| Refitted record (WTI crude (USO)): out of sample — every one of the 47 fits ends on or before its quarter-end (an expanding window from 2 Mar 2007; fits reaching past it) | 0 | 0 | 0 | pass |
| Refitted record (WTI crude (USO)): the tests' miss counts are the windows listed outside each band, and each coverage rate is (n − misses) ÷ n (50/80/90%: 23, 12, 5 misses of 46; sum of abs differences) | 0 | 0 | 0 | pass |
| Refitted record: coverage is nested, 50% band ≤ 80% band ≤ 90% band, for every asset (violations) | 0 | 0 | 0 | pass |
| Published coverage: every test's coverage equals 100 less its miss rate (max abs difference, pp) | 0 | 0 | 0 | pass |
| Published persistence: a + b of each DCC and α + β of every GARCH equal the sums of the published parameters (max abs difference) | 0 | 0 | 0 | pass |
| Falsifier: the federal holiday table the yields engine's session count needs is on the calendars fetched on or before 2026-11-17 (1 = found) | 1 | 1 | 0 | pass |
| Range kinds: each range this page prints names its kind — the cones simulated shocks, the refitted coverage record backtest, the engine's full-sample-parameter record past misses, the sensitivity table model (4 ranges; problems) | 0 | 0 | 0 | pass |
- Yahoo Finance chart API, daily adjusted closes (SPY QQQ SMH TLT IEF TIP GLD USO UUP BTC-USD) — Prices: Yahoo Finance. No price level is shown on this page.
- U.S. Department of the Treasury, daily par yield curve rates (2- and 10-year) — Source: U.S. Department of the Treasury
- Site regime engine (DCC-GARCH(1,1)-t), the ten-core-markets fit of 17 Sep 2026 reused without re-estimation; a second fit on yields and four ETFs; Monte Carlo cones and coverage backtest computed here — Model-conditional distributions computed by the site's engine; not forecasts.