Macro ReviewHow to read
Outlook · Synthesis · Range record and stress tests

The scenario model accounts for 36% of the S&P 500's median worst week.

What did each asset do in the market's most extreme weeks, and would the scenario model have said so?

Weekly returns to 13 Nov 2026Updated 16 Nov 2026
S&P 500, median of its ten worst weeks−6.7%what happened · 10 of 672 weeks, 2014–2026
The scenario model's median, same weeks−2.4% · betas on the 156 weeks before each
Share of the median week the model accounts for36%no market factor in the model
Weeks the model gets the sign right8of 10 · S&P 500 worst weeks

In brief

Each asset's median week in the four kinds of extreme week since 2014, and the scenario model's prediction for the same weeks
  • The S&P 500's ten worst weeks had a median of −6.7%; the model said −2.4%.→ Exhibit 1
  • In oil's ten largest weekly gains, energy stocks' median week was +5.3%; the model said +5.1%.→ Exhibit 2
  • In the ten largest weekly rises in the 10-year, long Treasuries' median week was −5.2%; the model said −5.2%.→ Exhibit 2
  • In the S&P 500's ten worst weeks since 2014, long Treasuries' median week was +1.6%.→ Exhibit 3
What would change this view
What would change this viewTriggerTodayDecided by
The model misses the index's worst weeksthe model's median below 25% of the S&P 500's median worst week36% (13 Nov 2026)not crossedthe next week among the ten worst
The reasoning behind each point
  • The index
    The S&P 500's ten worst weeks had a median of −6.7%; the model said −2.4%.

    The model has no market factor: it carries the index through oil, the 10-year, credit and the dollar, so a week the market falls on its own is mostly its miss. The signs agree in 8 of 10 weeks; the mean miss is 5.9 points.

  • Energy
    In oil's ten largest weekly gains, energy stocks' median week was +5.3%; the model said +5.1%.

    The signs agree in 9 of 10 weeks; the mean miss is 4.4 points.

  • Treasuries
    In the ten largest weekly rises in the 10-year, long Treasuries' median week was −5.2%; the model said −5.2%.

    The signs agree in 10 of 10 weeks; the mean miss is 0.8 points.

  • The sets
    In the S&P 500's ten worst weeks since 2014, long Treasuries' median week was +1.6%.

    Each set is the ten most extreme of 672 complete weeks from 3 Jan 2014 to 13 Nov 2026; 4 weeks fall in more than one set.

Next: Earnings outlook — the index's own statements carried forward

Source: Yahoo Finance; U.S. Treasury; Federal Reserve Board (H.10); Macro Review calculationscross.stress-read

Model against episode

For each stress week, the scenario model's prediction from the factors' actual moves that week, with betas estimated only on the 156 weeks before it, beside what each asset did

Exhibit 1 · S&P 500 return in its worst weeks, actual and predicted

Weekly log return, %, the 10 worst of 672 complete weeks, 3 Jan 2014 – 13 Nov 2026; dark: what happened; blue: the scenario model's prediction
About
About this exhibit
Bar chart of the S&P 500's weekly log return in each of its worst complete weeks since 2014 (dark: what happened), beside the scenario model's prediction for the same week from that week's oil, 10-year, credit and dollar moves (blue)
S&P 500: its 10 worst weeks, and what the scenario model said
Weekly log return in each of the 10 worst of 672 complete weeks, 3 Jan 2014 – 13 Nov 2026 (dark: what happened), and the scenario model's prediction for the same week (blue: model-conditional), from the factors' moves that week with betas on the 156 weeks before it. Median −6.7% against −2.4%; the signs agree in 8 of 10.
Source: Yahoo Finance; U.S. Treasury; Federal Reserve Board (H.10); Macro Review calculationscross.stress-3

Exhibit 2 · Each asset's median return in stress weeks, actual against predicted

Median weekly log return, %, over the 10 weeks of the selected set, 3 Jan 2014 – 13 Nov 2026; bars: what happened; diamonds: the model's prediction
About
About this exhibit
Chart of each asset's median weekly log return in the selected set of stress weeks (S&P 500 down by default; 10-year up, oil up and dollar up on a toggle): bars show what happened, diamonds the scenario model's median prediction
S&P 500 down: each asset's median week, and what the scenario model said
Median weekly log return of each asset in the S&P 500's ten worst weeks of 672 complete weeks, 3 Jan 2014 – 13 Nov 2026 (bars: what happened), and the scenario model's median prediction for the same weeks (diamonds: model-conditional) — betas on oil, the 10-year per +10bp, HYG − LQD and the broad dollar over the 156 complete weeks before each week (at least 104), times the factors' moves that week. The model has no market factor.
Source: Yahoo Finance; U.S. Treasury; Federal Reserve Board (H.10); Macro Review calculationscross.stress-1

What each asset did in each set of extreme weeks

Each asset's median or mean week in the S&P 500's ten worst weeks, the 10-year's ten largest rises and oil's and the dollar's ten largest gains since 2014, beside all weeks — what happened, not a forecast

Exhibit 3 · Each asset's typical return in each kind of extreme week, and overall

Median weekly log return, %, in the 10 weeks of each kind, of 672 weeks 3 Jan 2014 – 13 Nov 2026; last column: all weeks
About
Asset10-year yield: largest weekly risesOil (USO): largest weekly gainsBroad dollar: largest weekly gainsS&P 500: worst weeksAll weeks
Sectors
TechnologyXLK+0.4%+1.0%−1.7%−7.6%+0.4%
Communication servicesXLC+1.1%*−0.9%*−3.2%*−8.2%*+0.3%
Consumer discretionaryXLY+0.7%−1.4%−1.1%−7.9%+0.3%
Consumer staplesXLP+1.4%±0.0%−2.1%−4.7%+0.3%
Health careXLV+0.8%+0.7%−0.9%−6.7%+0.3%
FinancialsXLF+2.2%−0.9%−5.1%−7.5%+0.4%
IndustrialsXLI+2.0%±0.0%−3.2%−6.8%+0.4%
EnergyXLE+3.3%+5.3%−4.2%−12.6%+0.5%
MaterialsXLB+1.9%−0.2%−4.2%−8.3%+0.3%
UtilitiesXLU+0.8%−0.2%−3.6%−4.6%+0.4%
Real estateXLRE−0.1%*±0.0%*−4.4%*−5.7%+0.3%
Industries
SemiconductorsSMH+2.1%+1.8%−2.2%−9.5%+0.7%
Oil & gas E&PXOP+3.7%+7.3%−4.9%−13.6%+0.3%
Regional banksKRE+1.8%−1.8%−5.6%−9.2%+0.3%
HomebuildersXHB−0.6%−0.6%−3.2%−8.8%+0.2%
TransportsIYT+2.9%±0.0%−3.5%−7.4%+0.3%
RetailXRT+1.7%+0.1%−3.9%−7.9%+0.3%
REITsVNQ−0.7%−0.7%−3.4%−5.9%+0.3%
Indexes
S&P 500 equal weightRSP+1.4%−0.5%−3.2%−7.2%+0.3%
Russell 2000IWM+1.3%−0.2%−4.2%−8.6%+0.3%
Nasdaq-100QQQ+1.2%+0.1%−1.0%−7.8%+0.5%
Style spreads
Growth − valueGRW-VAL−1.4%+0.2%+0.5%−0.4%+0.1%
Momentum − S&P 500MOM−0.4%−0.9%+0.1%−0.5%+0.1%
Quality − S&P 500QUAL−0.3%+0.2%−0.1%±0.0%±0.0%
Small − largeSMALL+0.1%−0.2%−1.9%−1.5%−0.1%
Equal − cap weightBREADTH+0.2%±0.0%−1.4%−0.2%±0.0%
Low vol − S&P 500LOWVOL−1.0%−0.6%−0.4%+1.3%−0.1%
Cyclicals − defensivesCYC-DEF+1.2%+0.7%−1.3%−1.6%±0.0%
AI complex − defensivesAI-DEF+0.6%+2.1%+0.9%−1.3%+0.4%
Reference assets
S&P 500SPY+1.4%+0.1%−1.7%−6.7%+0.4%
Treasuries 20y+TLT−5.2%−0.6%−1.1%+1.6%+0.1%
GoldGLD−0.5%−0.1%−3.4%−1.5%+0.2%
US dollar bullish (UUP)UUP−0.3%+0.6%+2.2%±0.0%+0.1%
High yieldHYG−0.8%±0.0%−2.2%−2.1%+0.1%
About this exhibit
Table of each asset's median weekly log return (mean on a toggle) in four sets of extreme weeks since 2014: the 10-year's largest rises, oil's (USO) largest gains, the broad dollar's largest gains and the S&P 500's worst weeks; a final column gives all weeks
Tail weeks since 2014: how each asset behaved
Median weekly log return of each asset in the 10 weeks with the largest 10-year rise, the 10 largest USO gains, the 10 largest broad-dollar gains and the 10 worst S&P 500 weeks, of 672 complete weeks 3 Jan 2014 – 13 Nov 2026; last column = all weeks. * = asset has fewer than 10 of the weeks (shorter history). One other week ties the 10th-largest yield rise; the earlier one is kept.
Source: Regressions computed here on Yahoo Finance ETF prices (licensed), Treasury par yields and the Board's H.10 (public).cross.factors-10

The in-sample record beside the refitted one

Share of quarter-ends inside each band with parameters fitted once on the whole sample, against the refitted record on Range and scenarios

Exhibit 4 · How often each market ended inside its bands, in-sample record

Share of 46 quarter-ends, 2015 Q1 to 2026 Q2, inside the 25th–75th, 10th–90th and 5th–95th bands, %; whole-sample parameters; fitted mean in grey; shaded off nominal
Range: past missesAbout
AssetEngine50% band80% band90% bandAbove / below 90%Windows outside the 90% band (by the quarter-end they start from)
S&P 500spine33 6174 8393 910 / 32018 Q3, 2019 Q4, 2024 Q4
Nasdaq-100spine46 5078 8396 910 / 22018 Q3, 2019 Q4
Semiconductorsspine39 5474 8587 893 / 32018 Q3, 2019 Q4, 2023 Q4, 2024 Q4, 2025 Q1, 2026 Q1
Treasuries 20y+spine50 5074 7689 871 / 42016 Q3, 2019 Q4, 2020 Q4, 2021 Q4, 2023 Q2
Treasuries 7–10yspine46 4870 7483 805 / 32015 Q4, 2016 Q3, 2018 Q3, 2019 Q4, 2020 Q4, 2021 Q4, 2023 Q3, 2024 Q2
TIPSspine43 5274 7691 913 / 12015 Q4, 2019 Q4, 2021 Q4, 2024 Q2
Goldspine39 4376 8596 981 / 12015 Q4, 2016 Q3
WTI crude (USO)spine52 4678 8387 893 / 32015 Q2, 2018 Q3, 2019 Q4, 2021 Q4, 2025 Q4, 2026 Q2
US dollar bullish (UUP)spine52 4676 7087 876 / 02016 Q3, 2018 Q1, 2019 Q4, 2022 Q1, 2022 Q2, 2024 Q3
Bitcoin (USD)spine46 3961 6574 8010 / 22015 Q3, 2016 Q1, 2016 Q3, 2017 Q1, 2017 Q3, 2018 Q3, 2019 Q1, 2020 Q3, 2020 Q4, 2022 Q1, 2022 Q4, 2023 Q4
2-year Treasury yieldyields46 4665 6783 805 / 32016 Q3, 2017 Q3, 2017 Q4, 2019 Q4, 2021 Q3, 2021 Q4, 2023 Q3, 2024 Q2
10-year Treasury yieldyields48 4876 7485 856 / 12016 Q3, 2019 Q4, 2020 Q4, 2021 Q4, 2023 Q2, 2024 Q3, 2026 Q2
S&P 500yields30 6574 8391 910 / 42018 Q3, 2019 Q4, 2022 Q1, 2024 Q4
WTI crude (USO)yields48 4378 7887 892 / 42015 Q2, 2015 Q3, 2018 Q3, 2019 Q4, 2021 Q4, 2025 Q4
Goldyields33 4878 8591 983 / 12015 Q4, 2016 Q3, 2025 Q2, 2025 Q3
US dollar bullish (UUP)yields48 4676 7087 896 / 02016 Q3, 2018 Q1, 2019 Q4, 2022 Q1, 2022 Q2, 2024 Q3
About this exhibit
Table of each market and engine: share of quarter-ends inside the 50%, 80% and 90% bands (zero drift, fitted mean in grey), misses above and below the 90% band, and the quarters missed; shaded when far from nominal
The cones' own record, every market: full-sample parameters
Share of quarter-ends (46 from 2015 Q1 to 2026 Q2) at which the realised 65-session outcome fell inside the band simulated from the filtered state on that date, with parameters fitted on the whole sample — in-sample parameters, not an out-of-sample record; zero drift, with the fitted mean in grey. Shaded when more than 6pp from nominal.
Source: Yahoo Finance; U.S. Treasury; Macro Review calculationscross.cones-5

The two in full

Their fitted parameters, and how far the 13-week edges move when the drift, the tails or the correlation change

Exhibit 5 · Volatility model settings for each core market

Daily returns in %, 3,058 sessions, 18 Sep 2014 to 13 Nov 2026; mean, variance weights, tail thickness (ν) and starting annual volatility
About
Seriesμωαβα+βνstart σ (ann.)
S&P 500SPY+0.0990.02440.1750.8180.99335.423.2%
Nasdaq-100QQQ+0.1380.02990.1430.8550.99825.437.1%
SemiconductorsSMH+0.1610.05410.0930.8990.99236.264.9%
Treasuries 20y+TLT+0.0070.01010.0590.9290.987714.612.4%
Treasuries 7–10yIEF+0.0060.00160.0620.9290.991011.66.9%
TIPSTIP+0.0140.00220.0820.8970.97938.44.9%
GoldGLD+0.0370.01160.0530.9380.99075.526.3%
WTI crude (USO)USO+0.0860.15150.0950.8800.97477.146.1%
US dollar bullish (UUP)UUP+0.0190.00330.0470.9360.98309.56.2%
Bitcoin (USD)BTC+0.1260.45500.1240.8761.00003.154.8%

Bold rows sit at the α + β = 1 boundary of the estimator: their expected variance grows with the horizon instead of settling, so their long cones are the widest. Start σ = √(h at the first simulated session × 252).

About this exhibit
Table of fitted parameters for each core market (mean, variance intercept, shock and persistence weights, their sum, tail degrees of freedom) and each market's starting annualised volatility
Ten core markets engine: GARCH(1,1)-t by series
Daily units: returns in % (yields: changes in bp) · 3,058 sessions 18 Sept 2014 – 13 Nov 2026 · DCC a = 0.0180, b = 0.9660, a + b = 0.9839 · cached fit of the regime page, not re-estimated
Source: Yahoo Finance; U.S. Treasury; Macro Review calculationscross.cones-7

Exhibit 6 · Volatility model settings for the 2-year and 10-year yields and other markets

Daily changes in bp for yields, returns in % otherwise; 3,213 sessions, 3 Jan 2014 to 16 Nov 2026; starting annual volatility in the last column
About
Seriesμωαβα+βνstart σ (ann.)
2-year Treasury yieldy2+0.1230.06310.0710.9291.00006.0117.2bp
10-year Treasury yieldy10+0.0440.22820.0470.9440.991813.193.2bp
S&P 500SPY+0.1000.02560.1800.8120.99205.426.2%
WTI crude (USO)USO+0.0790.06970.0930.8990.99206.651.5%
GoldGLD+0.0350.01080.0500.9410.99125.427.5%
US dollar bullish (UUP)UUP+0.0190.00220.0480.9410.98918.96.4%

Bold rows sit at the α + β = 1 boundary of the estimator: their expected variance grows with the horizon instead of settling, so their long cones are the widest. Start σ = √(h at the first simulated session × 252).

About this exhibit
Table of fitted parameters for the 2- and 10-year yields, the S&P 500, WTI crude, gold and the dollar, with each series' starting annualised volatility
Yields engine: GARCH(1,1)-t by series
Daily units: returns in % (yields: changes in bp) · 3,213 sessions 3 Jan 2014 – 16 Nov 2026 · DCC a = 0.0213, b = 0.9647, a + b = 0.9860 · fitted here
Source: Yahoo Finance; U.S. Treasury; Macro Review calculationscross.cones-9

Exhibit 7 · Each core market's 13-week range under alternative model settings

%: 5th to 95th percentile of the 65-session cone from 13 Nov 2026; default (fat tails, changing correlation, zero drift), fitted mean, normal tails, constant correlation
Range: other models
SeriesDefaultFitted meanNormal tailsConstant corr.
S&P 500−15.8 … +19.0%−10.2 … +26.9%−17.5 … +21.5%−16.2 … +18.4%
Nasdaq-100−24.8 … +32.7%−17.7 … +45.1%−26.3 … +36.6%−25.4 … +32.3%
Semiconductors−38.2 … +62.9%−31.4 … +80.9%−39.4 … +64.9%−39.2 … +60.9%
Treasuries 20y+−10.1 … +11.3%−9.7 … +11.8%−10.3 … +11.2%−10.2 … +11.2%
Treasuries 7–10y−5.5 … +5.8%−5.2 … +6.1%−5.5 … +5.7%−5.4 … +5.8%
TIPS−4.0 … +4.2%−3.2 … +5.1%−4.1 … +4.2%−4.0 … +4.3%
Gold−18.0 … +22.5%−16.0 … +25.4%−18.4 … +22.7%−18.4 … +21.8%
WTI crude−29.3 … +40.1%−25.2 … +48.2%−29.6 … +42.4%−29.6 … +41.8%
US dollar bullish−5.3 … +5.5%−4.1 … +6.8%−5.3 … +5.5%−5.2 … +5.6%
Bitcoin−41.4 … +74.3%−36.4 … +89.2%−48.3 … +92.2%−39.6 … +69.6%

Read across a row: how much of the cone's width is the specification rather than the data. A wide gap between Default and Normal tails is the Student-t degrees of freedom doing the work; a wide gap to Constant corr. means the dynamic correlation, not the variances, is setting the joint spread; Fitted mean shifts the whole interval rather than widening it.

About this exhibit
Table of each of the ten core markets' 13-week 5th–95th range under four settings: default, fitted mean, normal tails and constant correlation
Sensitivity of the 13-week edges (ten core markets engine)
5th to 95th percentile of the 65-session cone: default (Student-t, DCC, zero drift) · with the fitted mean · normal innovations at the same variances · constant correlation Q̄ (a = b = 0).
Source: Yahoo Finance; U.S. Treasury; Macro Review calculationscross.cones-8

Exhibit 8 · The yields model's 13-week ranges under alternative model settings

bp for yields and % otherwise: 5th to 95th percentile of the 65-session cone from 13 Nov 2026; default, fitted mean, normal tails, constant correlation
Range: other models
SeriesDefaultFitted meanNormal tailsConstant corr.
2-year yield−95 … +96bp−87 … +104bp−99 … +98bp−95 … +96bp
10-year yield−75 … +75bp−73 … +78bp−76 … +75bp−77 … +76bp
S&P 500−17.0 … +21.2%−11.4 … +29.4%−18.5 … +22.7%−16.6 … +20.7%
WTI crude−33.5 … +49.3%−30.0 … +57.2%−34.2 … +51.0%−33.0 … +49.3%
Gold−19.1 … +23.3%−17.3 … +26.1%−19.3 … +23.8%−18.8 … +23.0%
US dollar bullish−5.3 … +5.7%−4.1 … +7.0%−5.3 … +5.5%−5.4 … +5.6%

Read across a row: how much of the cone's width is the specification rather than the data. A wide gap between Default and Normal tails is the Student-t degrees of freedom doing the work; a wide gap to Constant corr. means the dynamic correlation, not the variances, is setting the joint spread; Fitted mean shifts the whole interval rather than widening it.

About this exhibit
Table of the 13-week 5th–95th range for the 2- and 10-year yields and four markets under four settings: default, fitted mean, normal tails and constant correlation
Sensitivity of the 13-week edges (yields engine)
5th to 95th percentile of the 65-session cone: default (Student-t, DCC, zero drift) · with the fitted mean · normal innovations at the same variances · constant correlation Q̄ (a = b = 0).
Source: Yahoo Finance; U.S. Treasury; Macro Review calculationscross.cones-10

The engine, and the distribution-free comparison

Parameters, seeds, and the that assumes no distribution — but carries the drift of the days it resamples

Exhibit 9 · Each factor's range on the model against resampled history

% or bp, 10th and 90th percentiles over the selected horizon: the model (zero drift) beside resampled blocks of days since 1 Mar 2018
FactorωαβνEngine 10thEngine 90thBootstrap 10thBootstrap 90thBootstrap drift
US total market%0.03460.15480.82626.2−23.87+24.66−13.26+30.74+9.9
10-year yieldbp0.00290.05850.933211.6−113+115−85+137+26
Credit%0.00210.14240.85486.6−8.22+8.61−6.33+13.11+4.1
Oil%0.18140.09970.87026.2−49.36+48.64−56.29+58.56+4.0
Dollar%0.00470.06250.910811.5−8.32+8.23−3.11+12.36+4.6
Breakeven inflationbp0.00110.07950.90977.4−58+58−59+62+1
Emerging − developed%0.02120.07360.89248.7−17.28+17.54−16.21+11.06−2.8
Gold%0.01510.06570.92435.4−27.04+26.90−7.85+32.98+11.9
Short volatility%0.57640.28700.66323.5−59.79+61.30−36.89+51.70+9.5

1 year ahead; the GARCH parameters are in the engine's own units (basis points over ten for a yield factor), the quantiles in percent or basis points. The bootstrap assumes no distribution, but it is not drift-free: it resamples the days as they happened, so each band is centred on that factor's average move since 1 Mar 2018 over the horizon — the last column (+9.9% for the market over 1 year) — where the engine carries none. The gap between the two columns is that drift plus the engine's model. Residuals: 58 assets, each resampled from its own history by the stationary block bootstrap (blocks of 21 sessions on average, seed 20260929); the alternative gives each its own GARCH, the 11 sector residuals drawn together with today's correlation from Correlations & regime (tiers12), every other asset's independently. Betas: independent normal, per path, seed 20260928.

About this exhibit
Table of the nine factors' fitted volatility parameters and their 10th and 90th percentiles from the model at the selected horizon, beside the same percentiles from resampled blocks of historical days and their drift
The factor model, and the comparison that assumes no distribution
GARCH(1,1)-t per raw factor with a scalar DCC (a 0.0180, b 0.9641), fitted on 2,191 sessions from 1 Mar 2018 — 2018 rather than 2014, because the short-volatility proxy restarts then and a DCC needs every series on every day. Beside each factor's simulated band, the same horizon from a stationary block bootstrap of historical days (blocks of 21 sessions, 4,000 paths), which assumes no distribution at all.
Source: Yahoo Finance; U.S. Treasury; Macro Review calculationscross.projections-6

Appendix

What happened and what the model said, how each prediction is made, and every source and check164 checks, all pass
About this page

What happened, then what the model said. The stress weeks are the 10 most extreme of 672 complete weeks since 3 Jan 2014 on four series — the S&P 500's worst weeks, the 10-year's largest rises, oil's and the dollar's largest gains — and the table beneath the record is each asset's median week in each set. The record beneath asks whether the scenario model, which carries every asset through oil, the 10-year, credit and the dollar, would have said so: each prediction uses betas estimated only on the weeks before it and the factors' actual moves in the week, so it tests the mapping from the factors to the assets, not a forecast of the factors. It is model-conditional, and descriptions of 10 weeks are not a distribution.

Exhibit 10 · How the stress weeks and the model's record are made

  • Record
    What happened, and what the model said.

    What happened is observed and the model's prediction is model-conditional. The stress weeks are the weekly five-factor model's sets: the 10 most extreme of 672 complete weeks from 3 Jan 2014 to 13 Nov 2026 on each selecting series. For each week, each asset's weekly log return is set against a constant, oil (USO), the 10-year change per +10bp, HYG − LQD and the broad dollar over the 156 complete weeks before it (at least 104); the prediction is those betas times the factors' moves in the week itself. The record tests the mapping from the factors to the assets, not a forecast of the factors. A fund without 104 weeks of history before a week is not predicted for it, and its count says so. The record is published under the author's decision of 8 Oct 2026 (D-2026-10-08-RA15).

  • Weeks
    How the stress weeks are chosen.

    Tail tables: of the 672 complete weeks from 3 Jan 2014 to 13 Nov 2026, the 10 with the largest weekly rise in the 10-year par yield, the 10 largest USO weekly log returns, the 10 largest broad-dollar weekly log changes and the 10 worst S&P 500 weeks; for each set the median and mean weekly log return of every asset on this page and of TLT, GLD, UUP and HYG, beside the same statistics over all weeks. Ten weeks are a description, not a distribution: one 2020 week can move a mean.

  • Overlap
    Weeks in more than one set.

    4 weeks fall in more than one set: 11 Nov 2016, 13 Mar 2020, 20 Mar 2020, 3 Apr 2020.

Source: Yahoo Finance; U.S. Treasury; Federal Reserve Board (H.10); Macro Review calculationscross.stress-2

Exhibit 11 · How the projections are made

About
  • 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 factors

    The nine raw factor proxies of Factor analysis's equity lens, simulated as they trade. The engine is fitted on the 2,191 sessions from 1 Mar 2018 — 2018 rather than 2014, because the short-volatility proxy restarts then and a DCC needs a full history for every factor.

  • Residualising

    Each simulated path is residualised with the slopes the Return attribution page fixed at the month-end before this update (2026-10-30, applying to 2026-11), in the same tier order, so the betas meet the factors they were estimated on. The regressions' constants are left out: each is the in-sample mean of what the parents did not explain, and subtracted from draws with zero drift it would give the residualised factor a drift of its own (for gold, about −24% a year). The slopes are held over the horizon; in the data they are refitted every month-end, and holding them is an assumption this page makes and prints.

  • 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.

  • The comparison

    A block bootstrap of historical raw factor days, blocks of 21 sessions, 4,000 paths. It assumes no distribution, but it does carry the drift of the days it resamples — each factor's average move since 1 Mar 2018, printed beside it — where the engine carries none; the gap between the two is that drift plus the engine's model.

  • The second route

    A second road to the same asset: Δln P = Δln E + Δln PE, with the multiple moved by the Valuation page's own two-stage identity and applied to today's observed multiple, so the route does not close the valuation gap. The rates draw is this page's own, the earnings draw is inverse-CDF sampling of the Earnings page's default fan — its NIPA profits grid bridged to EPS by that page's bridge (−0.92 log points a year), as the Valuation page bridges it — and the three are tied by the Valuation page's Gaussian copula over all quarters. The premium's four-quarter changes are drawn centred on their mean, so on average the multiple carries no drift of its own; the pool is skewed, so the multiple leg's median is not zero, and it is printed beside the route. The earnings leg keeps the Earnings page's median, which is the growth that page expects. At zero shocks it returns today's price exactly. It is not calibrated: the backtest scores the factor route.

  • The backtest

    The 63-session band at 18 quarter-ends from 31 Mar 2022 to 30 Jun 2026 for 12 assets, 2,000 paths each, made at each quarter-end the way the page makes it today and scored at every step of the ladder: fixed betas with a GARCH-t residual; (a) betas drawn; (b) the residual bootstrapped; (c) the width calibrated. The filtered state, the betas and their errors, the residual's variance and its bootstrap pool are point in time. The GARCH and DCC parameters are the full sample's, which flatters coverage; beside them, the same ladder with the engine and the residual GARCH refitted at each year-end on data to that date (6 refits), with the full sample's k applied unchanged. The pooled interval comes from a block bootstrap over quarter-ends (8 quarters a block, 2,000 replications) because the assets are correlated and a binomial interval on the pooled count would be far too narrow. The backtest published before 24 September left the residual out of the band it scored and mapped the four factor-proxy ETFs through betas on residualised factors; it now scores the band the page draws.

  • 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.

Source: Yahoo Finance; U.S. Treasury; Macro Review calculationscross.projections-9

Before you read

Macro Review is a personal research site, still being built.

Nothing on it is investment advice, a recommendation or a solicitation.

Check every figure for yourself. Each names its publisher — confirm it there before you rely on it.

Figures can be revised, delayed or wrong.

Cones, fans, scenarios, fair values and “what followed” tables are conditional on the model each page names: they describe, they do not predict.

No warranty is given; use is at your own risk.