Unproven, not disproven: four decades cannot test the 42% gap against simpler measures
Has this valuation gap predicted ten-year returns?
In brief
What the record says, in a few lines; what would change the read beneath- After the 47 months in today's decile the next ten years' real return was below average in 33 of them.→ Exhibit 3
- Since 1995 a wider gap preceded lower ten-year returns, but 4 independent decades cannot separate it from the earnings yield less the real yield.→ Exhibit 10
- from 1995, the gap's ten-year forecasts did worse than the historical mean on 2 non-overlapping windows.→ Exhibit 8
- Since 1926, the earnings yield less the real yield cut the historical mean's ten-year forecast error by 26%, on 9 non-overlapping windows.→ Exhibit 5
- Recomputed from what was known each month, the two-stage gap reads −52.1% in Nov 2026, against the Valuation page's −41.5%.→ Exhibit 1
| What would change this view | Trigger | Today | Decided by |
|---|---|---|---|
| The gap starts adding to the earnings yield less the real yield | ten-year t above 1.96 and encompassing p below 0.05 | t 0.36, p 0.44 (Nov 2026)not crossed | Shiller's monthly update: each adds a realised month of returns |
| The gap's forecasts start beating the historical mean | ten-year out-of-sample R² above 0 | −0.38 (origins to Nov 2016)not crossed | Shiller's monthly update: each adds a realised ten-year return |
| The benchmark's long record fails | the earnings yield less the real yield, ten-year out-of-sample R² since 1926 at or below 0 | +0.26not crossed | Shiller's monthly update |
| Real-time data change the verdict | the revised-data gap passes the encompassing rule at a horizon where the real-time gap fails, or the reverse | the same verdict at every horizonnot crossed | the Philadelphia Fed's next real-time release, mid-November 2026 |
The reasoning behind each point
- What followedAfter the 47 months in today's decile the next ten years' real return was below average in 33 of them.
Today's two-stage gap, −52.1% (Nov 2026), sits in decile 1 of the 607 months since 1975 (−80.2% to −47.2%). Of its 61 months, 47 have a known ten-year real return: median +0.5% a year against 6.65% for every month since 1871. A handful of overlapping decades, not a forecast: the record is unproven, not disproven.
- EncompassingSince 1995 a wider gap preceded lower ten-year returns, but 4 independent decades cannot separate it from the earnings yield less the real yield.
With both in one regression of forward real returns, the gap's Hodrick t-statistics at 1, 3, 5, 10 years are 0.56, −0.06, 0.00, 0.36; the encompassing test of forecasts (the earnings yield less the real yield against the gap) gives p-values of 0.59, 0.70, 0.73, 0.44. The rule set before the estimates were made asks for t above 1.96 and p below 0.05 together. The one-stage gap, at ten years, has t 0.87 and p 0.001: it passes only the out-of-sample half. No horizon passes, but the test has little power: with 4 non-overlapping ten-year windows a gap as strong as its own estimate would often fail the rule, and at this horizon the out-of-sample encompassing test's small-sample correction breaks down, so its p-value is a rough guide only. The two measures move together (both are earnings over price less a rate). The revised-data history gives the same verdict (ten-year t 0.39).
- Out of sampleOut of sample from 1995, the gap's ten-year forecasts did worse than the historical mean on 2 non-overlapping windows.
Each month's forecast uses only returns already known then. On the same months the earnings yield less the real yield scores −0.18, the cyclically adjusted earnings yield −0.15, the one-stage gap −0.30. 248 monthly forecasts overlap heavily: behind them are 2 independent ten-year windows.
- Long recordSince 1926, the earnings yield less the real yield cut the historical mean's ten-year forecast error by 26%, on 9 non-overlapping windows.
Out-of-sample R² at one, three, five and ten years: +0.04, +0.06, +0.13, +0.26; the cyclically adjusted earnings yield +0.03, +0.05, +0.08, +0.15. Trained from 1881, each forecast uses only returns known when it was made. The predictor is nearly a random walk (monthly autocorrelation 0.990) and its shocks move against returns (correlation −0.50), so the in-sample slopes are biased up; the bootstrap removes it.
- TodayRecomputed from what was known each month, the two-stage gap reads −52.1% in Nov 2026, against the Valuation page's −41.5%.
The one-stage gap reads −56.5% against −27.6%. The history's rules differ from that page's on purpose, because they must work in 1975 as well as today: long-run growth is the surveys' 4.4% rather than the Committee's 4.0%; near-term growth is the earnings model's 5.5% rather than 5.8%; and the premium compared with is the median of every month's implied premium since Jun 1970, 5.55pp.
- Expected returnModel-conditional: the gap's regression implies +5.8% a year real over the next ten years.
The same arithmetic on the earnings yield less the real yield (1.64pp today), fitted since 1881, gives +4.5% with an error range of −1.2 to 10.7; its record is longer, so it is the better guide of the two, though on the gap's own months it too lost to the historical mean (out-of-sample R² −0.18), and before 1997 its real yield is the 10-year less trailing inflation, not a TIPS yield. Across the payout range the gap's figure runs from 5.6 to 6.0. Neither is a forecast: each is a fitted line read at today's value. The slope is shallow: at +0.038pp a year per point of gap, today's −52.1% takes 2.0pp off the line's zero-gap 7.7%, which leaves +5.8% only 0.9pp below the 6.65% average of every month since 1871.
Next: Range and scenarios — the range around today's prices
What premium has the market implied, month by month?
The recomputed from what was known each month since 1970, today and the full-history median marked; the gap to fair value as the other viewExhibit 1 · Implied equity risk premium since 1970, as known each month
About this exhibit
Exhibit 2 · Inputs and gaps as known each month against the Valuation page's
| Input | This history | Valuation page, 16 Nov 2026 | Why they differ |
|---|---|---|---|
| Two-stage gap to the model's value | −52.1% | −41.5% | revised data: −52.4% |
| One-stage gap to the model's value | −56.5% | −27.6% | the Valuation page's one-stage gap prices forward earnings and consensus growth |
| Long-run growth g_L | 4.40% | 4.00% | the surveys' ten-year real GDP 2.1 + CPI 2.3, against the Committee's longer-run medians |
| Near-term growth g_S, a year | +5.48% | +5.77% | the earnings model on index EPS, against the profits fan bridged to EPS |
| Long rate | 4.88% | 5.12% | Shiller's long rate (the month's average), against the Treasury 10-year par yield |
| Trailing earnings yield E0 ÷ P | 3.18% | 3.99% | earnings to Mar 2026 (six months back), against the latest trailing earnings |
| Premium the price implies (two-stage) | 2.41pp | – | the premium that returns today's price through the identity |
| Premium the gap is measured at | 5.55pp | 5.17pp | the median of every month's implied premium so far, against the yield-gap premium (today's implied premium plus the yield gap's distance to its median in months like these) |
At the Valuation page's own inputs the identity returns its published −41.504% to within rounding.
About this exhibit
What followed readings like today's?
Every month's gap against the real return of the next ten years, today's reading marked; the months in today's decile and how often the decade after fell shortExhibit 3 · Each month's valuation gap against the next ten years' real return
About this exhibit
What do the plain benchmarks say over the long record?
The earnings yield, the earnings yield less the real yield and the earnings yield less the bond yield since 1881, and their forecast record since 1926Exhibit 4 · Simple valuation measures over the long record
About this exhibit
Exhibit 5 · Forecast record of simple valuation measures since 1926
| Measure | R², 1y | R², 3y | R², 5y | R², 10y | Clark–West p, 10y | 10y slope less bias | t, Hodrick | Windows fitted · forecast |
|---|---|---|---|---|---|---|---|---|
| Earnings yield less real yield | +0.04 | +0.06 | +0.13 | +0.26 | 0.010 | +0.519 | 3.03 | 13 · 9 |
| Cyclically adjusted earnings yield | +0.03 | +0.05 | +0.09 | +0.15 | 0.016 | +0.525 | 2.38 | 13 · 9 |
| Earnings yield less bond yield | +0.01 | −0.01 | −0.03 | −0.10 | 0.220 | +0.422 | 2.07 | 14 · 9 |
About this exhibit
What does today's reading imply for the next ten years? ()
The ten-year real return the history's regressions imply at today's readings, with the error of each rule as its rangeExhibit 6 · Ten-year real return each fitted line implies at the latest reading
| Rule | Reading today | Fitted line | Implied, % a year real | Range | Kind | Its record |
|---|---|---|---|---|---|---|
| Two-stage gap, real-timefitted May 1975 – Nov 2016, 4 windows | −52.09 | 7.72 + 0.038 × reading | +5.8 | -7.7 to 6.3 | out-of-sample record | R² −0.382 windows forecast |
| Earnings yield less real yieldfitted Mar 1881 – Nov 2016, 13 windows | +1.65 | 3.57 + 0.589 × reading | +4.5 | -1.2 to 10.7 | out-of-sample record | R² +0.269 windows forecast |
| Cyclically adjusted earnings yieldfitted Jun 1881 – Nov 2016, 13 windows | +2.42 | 1.07 + 0.769 × reading | +2.9 | -1.8 to 10.6 | out-of-sample record | R² +0.159 windows forecast |
| Two-stage gap, payout k across its declared range | 5.6 to 6.0 | – | assumed values | payout k from 0.75 to 0.95, each end refitted |
the 10th and 90th percentiles of the rule's out-of-sample errors, 1995–2016 (2 non-overlapping windows) for the gap; the 10th and 90th percentiles of the rule's out-of-sample errors, 1926–2016 (9 non-overlapping windows) for the benchmarks. On its few windows the record cannot tell the gap from the earnings yield less the real yield, whose record is longer, so that row is the better guide of the two.
About this exhibit
The tests: slopes, forecasts , , the inputs and the payout assumption
Overlap-robust slopes with the small-sample bias removed, forecasts scored out of sample, whether the gap adds to the earnings yield less the real yield, what moved the gap, and each January's estimate
Exhibit 7 · How strongly each measure predicted ten-year real returns
| Measure | Slope | t, Newey–West | t, Hodrick | Bias under the null | Slope less bias | Bootstrap p | R² | Windows |
|---|---|---|---|---|---|---|---|---|
| Two-stage gap, real-timeMay 1975 – Nov 2016, per point of gap | +0.038 | 1.85 | 1.61 | +0.004 | +0.033 | 0.051 | 0.18 | 4 |
| Two-stage gap, revised dataMay 1975 – Nov 2016, per point of gap | +0.039 | 1.88 | 1.62 | +0.004 | +0.035 | 0.047 | 0.19 | 4 |
| One-stage gap, real-timeMay 1962 – Nov 2016, per point of gap | +0.063 | 3.72 | 2.13 | +0.006 | +0.057 | 0.020 | 0.18 | 5 |
| One-stage gap, revised dataMay 1962 – Nov 2016, per point of gap | +0.063 | 3.99 | 2.06 | +0.007 | +0.056 | 0.023 | 0.18 | 5 |
| Cyclically adjusted earnings yieldJun 1881 – Nov 2016, per point of gap | +0.769 | 5.11 | 2.38 | +0.244 | +0.525 | 0.071 | 0.24 | 13 |
| Earnings yield less real yieldMar 1881 – Nov 2016, per percentage point | +0.589 | 5.08 | 3.03 | +0.070 | +0.519 | 0.021 | 0.28 | 13 |
| Earnings yield less bond yieldJun 1872 – Nov 2016, per percentage point | +0.492 | 1.89 | 2.07 | +0.070 | +0.422 | 0.064 | 0.12 | 14 |
Windows: non-overlapping 10-year spans in the sample — the honest count behind each number.
About this exhibit
Exhibit 8 · Out-of-sample forecast record of each measure against the historical mean
| Measure | 1 year | 3 years | 5 years | 10 years | Clark–West p, 10y |
|---|---|---|---|---|---|
| Two-stage gap, real-time | −0.05 | −0.30 | −0.40 | −0.38 | 0.539 |
| Two-stage gap, revised data | −0.05 | −0.30 | −0.41 | −0.39 | 0.569 |
| One-stage gap, real-time | −0.05 | −0.28 | −0.37 | −0.30 | 0.416 |
| One-stage gap, revised data | −0.05 | −0.29 | −0.39 | −0.28 | 0.507 |
| Cyclically adjusted earnings yield | −0.14 | −0.39 | −0.74 | −0.15 | 0.467 |
| Earnings yield less real yield | −0.04 | −0.19 | −0.23 | −0.18 | 0.847 |
| Earnings yield less bond yield | −0.04 | −0.15 | −0.17 | −0.03 | 0.605 |
| Non-overlapping windows forecast | 39 | 11 | 6 | 2 |
About this exhibit
Exhibit 9 · Where each measure won and lost against the historical mean
About this exhibit
Exhibit 10 · Gap and earnings yield less real yield tested together, by horizon
| Horizon | Gap slope | t, Hodrick | t, Newey–West | Benchmark t, Hodrick | Weight on the gap (λ) | Encompassing p | Clark–West p | Adds? |
|---|---|---|---|---|---|---|---|---|
| 1 year50 windows in sample | +0.023 | 0.56 | 0.54 | 0.16 | -0.22 | 0.590 | 0.829 | no |
| 3 years16 windows in sample | −0.002 | -0.07 | -0.07 | 0.34 | -0.42 | 0.703 | 0.704 | no |
| 5 years9 windows in sample | ±0.000 | 0.00 | -0.01 | 0.50 | -0.66 | 0.729 | 0.858 | no |
| 10 years4 windows in sample | +0.010 | 0.36 | 1.35 | 0.94 | 0.03 | 0.436 | 0.569 | no |
About this exhibit
Exhibit 11 · Long-run and near-term growth inputs, as known each month
About this exhibit
Exhibit 12 · Implied premium against its median so far, and the long rate
About this exhibit
Exhibit 13 · Gap on real-time data less gap on revised data, since 1965
About this exhibit
Exhibit 14 · The record at the ends of the payout range
| Identity | k | Gap today | 10y slope | t, Hodrick | R², out of sample | Implied 10y return |
|---|---|---|---|---|---|---|
| two-stage | 0.75 | −53.5% | +0.029 | 1.57 | −0.50 | +6.0% |
| one-stage | 0.75 | −57.6% | +0.056 | 2.13 | −0.45 | +3.1% |
| two-stage | 0.95 | −51.1% | +0.045 | 1.63 | −0.33 | +5.6% |
| one-stage | 0.95 | −55.1% | +0.066 | 2.05 | −0.23 | +2.8% |
| two-stage | 0.85 | −52.1% | +0.038 | 1.61 | −0.38 | +5.8% |
| one-stage | 0.85 | −56.5% | +0.063 | 2.13 | −0.30 | +2.9% |
Exhibit 15 · The earnings model's parameters estimated each January
| Estimated | Quarters | g_L | φ | σ_η | ρ1 | ρ2 | σ_ν | σ_ε | Jumps | Log-likelihood |
|---|---|---|---|---|---|---|---|---|---|---|
| Jan 2026 | 262 | 4.48 | 0.95 fixed | 0.000 | 0.86 | -0.38 | 6.95 | 0.00 | Oct 2008, Apr 2009, Oct 2009, Jan 2010 | -880.8 |
| Jan 2025 | 258 | 4.23 | 0.95 fixed | 0.000 | 0.86 | -0.39 | 7.00 | 0.00 | Oct 2008, Apr 2009, Oct 2009, Jan 2010 | -869.2 |
| Jan 2024 | 254 | 4.40 | 0.95 fixed | 0.000 | 0.86 | -0.39 | 7.04 | 0.00 | Oct 2008, Apr 2009, Oct 2009, Jan 2010 | -857.2 |
| Jan 2023 | 250 | 5.23 | 0.95 fixed | 0.000 | 0.86 | -0.39 | 7.07 | 0.00 | Oct 2008, Apr 2009, Oct 2009, Jan 2010 | -844.6 |
| Jan 2022 | 246 | 4.80 | 0.95 fixed | 0.000 | 0.87 | -0.41 | 7.07 | 0.00 | Oct 2008, Apr 2009, Oct 2009, Jan 2010 | -831.2 |
| Jan 2021 | 242 | 4.12 | 0.95 fixed | 0.000 | 0.93 | -0.45 | 6.73 | 0.00 | Oct 2008, Apr 2009, Oct 2009, Jan 2010 | -805.6 |
| Jan 2020 | 238 | 4.19 | 0.95 fixed | 0.000 | 0.95 | -0.47 | 6.58 | 0.00 | Oct 2008, Apr 2009, Oct 2009, Jan 2010 | -787.3 |
| Jan 2019 | 234 | 4.36 | 0.95 fixed | 0.000 | 0.95 | -0.47 | 6.63 | 0.00 | Oct 2008, Apr 2009, Oct 2009, Jan 2010 | -775.8 |
| Jan 2018 | 230 | 4.65 | 0.95 fixed | 0.000 | 0.95 | -0.47 | 6.68 | 0.04 | Oct 2008, Apr 2009, Oct 2009, Jan 2010 | -764.4 |
| Jan 2017 | 226 | 4.49 | 0.95 fixed | 0.000 | 0.95 | -0.47 | 6.74 | 0.00 | Oct 2008, Apr 2009, Oct 2009, Jan 2010 | -752.9 |
| Jan 2016 | 222 | 4.65 | 0.95 fixed | 0.000 | 0.81 | -0.37 | 7.17 | 0.00 | Oct 2008, Oct 2009, Jan 2010 | -753.3 |
| Jan 2015 | 218 | 4.80 | 0.95 fixed | 0.000 | 0.81 | -0.37 | 7.22 | 0.00 | Oct 2008, Oct 2009, Jan 2010 | -741.3 |
Near-term growth g_S. The Earnings page's unobserved-components model on index EPS — its state space, bounds and procedure (estimate freely; if φ reaches a bound or σ_η collapses, fix φ at 0.95; quarters beyond |z| > 3 become jump dummies; re-estimate), steps one to three from the default start as on that page's trailing-EPS fit, on quarterly growth from 1960, with μ̄ = g_L ÷ 4. Re-estimated each January (and at the first month with 40 quarters, 1970) on the quarters known then — 57 estimates in the real-time run — and filtered every month with the last estimate's parameters; a new quarter whose one-step error exceeds the jump cut is treated as a jump rather than moving the state, as re-estimating would treat it (23 such quarters in the real-time run). g_S is the model's expected growth over the next 2 years from the last known quarter — the long-run state reverting to μ̄, the cycle following its AR(2) — as a compound annual rate. Months the identity cannot price. The two-stage gap needs (1 + g_L) + H(g_S − g_L) above zero and r − g_L above 0.5pp. In the real-time run 12 months fail the first (2010: the January estimate after the 2008–09 write-downs put near-term growth so far below zero that the bracket turned negative), and 0 the second; they are left out of every statistic, never filled.
About this exhibit
Exhibit 16 · The history every January
| January | Two-stage gap | revised | One-stage gap | revised | g_L | revised | g_S | Reference premium | Long rate | Earnings yield | Next ten years |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 2026surveys | −46.4 | −46.8 | −50.3 | −50.8 | 4.48 | 4.48 | +4.6 | 5.58 | 4.21 | 3.21 | – |
| 2025surveys | −49.9 | −50.2 | −54.7 | −55.2 | 4.23 | 4.23 | +5.7 | 5.62 | 4.63 | 3.29 | – |
| 2024surveys | −36.0 | −36.2 | −41.6 | −42.6 | 4.40 | 4.40 | +5.4 | 5.66 | 4.06 | 3.76 | – |
| 2023surveys | +4.1 | +2.3 | −0.8 | −2.9 | 5.23 | 5.23 | +3.9 | 5.69 | 3.53 | 4.85 | – |
| 2022surveys | +8.2 | +5.5 | +2.0 | −1.2 | 4.80 | 4.80 | +2.3 | 5.68 | 1.76 | 3.47 | – |
| 2021surveys | −12.2 | −14.3 | −24.3 | −26.8 | 4.12 | 4.12 | +4.6 | 5.70 | 1.08 | 2.62 | – |
| 2020surveys | +10.8 | +8.6 | ±0.0 | −2.7 | 4.19 | 4.19 | +3.9 | 5.67 | 1.76 | 4.13 | – |
| 2019surveys | +5.1 | +4.6 | −5.7 | −8.4 | 4.36 | 4.36 | +5.0 | 5.66 | 2.71 | 4.70 | – |
| 2018surveys | −10.8 | −12.4 | −18.7 | −20.3 | 4.65 | 4.65 | +3.8 | 5.67 | 2.58 | 3.73 | – |
| 2017surveys | −5.2 | −6.1 | −17.7 | −19.7 | 4.49 | 4.49 | +4.9 | 5.66 | 2.43 | 3.82 | – |
| 2016surveys | +39.8 | +37.3 | +21.8 | +18.4 | 4.65 | 4.65 | +3.9 | 5.62 | 2.09 | 4.95 | +11.4 |
| 2015surveys | +74.0 | +71.1 | +42.5 | +37.8 | 4.80 | 4.80 | +5.5 | 5.56 | 1.88 | 5.08 | +9.5 |
Appendix
The rules for what was known when, the anchor, the model, the identity, the statistics, sources and checks36 checks, all pass
The Valuation page prints today's gap to the model's value. This page recomputes that gap at every month since the data allow, from what was known then — long-run growth from the Philadelphia Fed's real-time data and surveys, near-term growth from the earnings model re-estimated each January — and asks whether it ever told a reader more about the returns that followed than the plain earnings yield less the real yield.
The predictive regression and its tests
| Symbol | Meaning | From | Assumed or estimated |
|---|---|---|---|
| the month's log real total return; the left side is its mean over the next h months, a year (× 12) | Shiller's real total-return index | – | |
| the measure at month t: the two-stage gap, or a benchmark | this page | – | |
| the slope; its errors Newey–West (h lags) and Hodrick (1992) | estimated | estimated | |
| Campbell–Thompson out-of-sample R² against the historical mean | estimated | estimated |
Overlapping monthly origins: the count that matters is the non-overlapping ten-year windows. The Stambaugh bias is removed by a bootstrap under the null.
Exhibit 17 · Definitions
- Model-conditionalWhat the last open section is, and is not.
Model-conditional, not a forecast: the last open section reads each regression line at today's value. The line is fitted on a record of 4 non-overlapping ten-year windows for the gap and 13 for the earnings yield less the real yield; its range is the rule's own out-of-sample error, not a probability.
- WhenWhat was known at each month.
What was known when. Each month's price is Shiller's average of daily closes. Earnings are the four-quarter total at the last quarter-end at least 6 months back — from Mar 1926; before that the figures are annual and the last December at least 6 months back is used, since the months between are interpolated towards the next December. CPI is read 2 months back. The long rate is the month's own from Apr 1953; before that Shiller's annual series is interpolated between Januaries, so the January at least 12 months back is used. The cyclically adjusted earnings yield is the average of ten years of earnings to the last known month, each in the prices of the last known CPI month, over the price. A check asserts every rule at every month.
- AnchorLong-run growth known at the time.
Long-run growth g_L known at each month. From Mar 1992: the Survey of Professional Forecasters' ten-year medians, real GDP (asked in the first-quarter survey) plus CPI inflation (every survey) — 2.1 + 2.3 = 4.4% in Nov 2026. Before that: nominal output's compound growth over the last 20 years (at least 10) in the Philadelphia Fed's real-time data as published at the time, the first from Dec 1965; before that, today's figures, and the page says so. A survey or real-time release of a quarter is used from the month after the quarter's middle month. Nominal output is GNP in releases before 1992 and GDP after, as published. CPI inflation runs a few tenths above the GDP deflator, so the survey anchor sits a little above nominal GDP growth.
- Real timeTwo runs, and what separates them.
Two runs. Real-time: every output figure as published at the time. Revised: the same quarters as published today. The surveys are never revised, so from Mar 1992 the anchor is the same in both, and from Jan 1993 so is every estimate of the earnings model; before, they differ by the revisions to twenty years of output growth (−0.13pp on average in g_L, real-time less revised). The gaps still differ a little after Mar 1992 (2.26 points on average, +0.26 in Nov 2026) because the reference premium is a median over every earlier month, including those before it.
- The modelNear-term earnings growth, re-estimated each January.
Near-term growth g_S. The Earnings page's unobserved-components model on index EPS — its state space, bounds and procedure (estimate freely; if φ reaches a bound or σ_η collapses, fix φ at 0.95; quarters beyond |z| > 3 become jump dummies; re-estimate), steps one to three from the default start as on that page's trailing-EPS fit, on quarterly growth from 1960, with μ̄ = g_L ÷ 4. Re-estimated each January (and at the first month with 40 quarters, 1970) on the quarters known then — 57 estimates in the real-time run — and filtered every month with the last estimate's parameters; a new quarter whose one-step error exceeds the jump cut is treated as a jump rather than moving the state, as re-estimating would treat it (23 such quarters in the real-time run). g_S is the model's expected growth over the next 2 years from the last known quarter — the long-run state reverting to μ̄, the cycle following its AR(2) — as a compound annual rate.
- IdentityThe Valuation page's identity, month by month.
The identity is the Valuation page's, the site's one implementation: P = k·E0·[(1 + g_L) + H(g_S − g_L)] ÷ (r − g_L), with k 0.84 and H 3 (6-year fade) as that page sets them, E0 ÷ P the trailing earnings yield known at the month, and r the long rate plus a reference premium. The one-stage gap sets g_S = g_L. At the Valuation page's own inputs the same function returns its published −41.5% — a check.
- PremiumThe premium the gap is measured at.
The reference premium. At each month the identity implies the premium that returns the price exactly; the reference is the median of that implied premium over every earlier month and this one (at least 5 years of them), and the gap is the model price at the reference premium over the price, less one. Positive: the index sits below fair value. The Valuation page uses a median over months with the same sign of the stock–bond correlation since 2005, which has no earlier history. A gap made this way is close to a transform of the earnings yield against the bond yield — which is why the encompassing test is compulsory.
- RecordSlopes, errors and the bias.
The record. For each measure and horizon h (one, three, five and ten years), forward log real total returns (Shiller's real total-return index, a year) regressed on the measure at the month. Monthly origins overlap, so the errors are Newey–West with h monthly lags and Hodrick's (1992) — one-month returns under the null times the sum of the last h readings of the predictor. The slope is biased in small samples when a persistent predictor's shocks move against returns (Stambaugh, 1999): a bootstrap under the null (1,000 samples, the predictor's own AR(1) with its shocks paired to the returns', a fixed seed) gives the bias at each horizon and a one-sided p-value. Non-overlapping windows are counted for every number.
- Out of sampleForecasts made with what was known.
Out of sample. At each month a regression on the pairs whose returns were already known then, against the historical mean of the same returns, once 120 such pairs exist; Campbell and Thompson's R² = 1 − the model's squared errors ÷ the mean's; Clark and West's test for a nested model; Goyal and Welch's cumulative difference of squared errors, which rises when the measure beats the mean. The gap and every benchmark are scored on the same months (ten-year forecasts from Apr 1995); the benchmarks also on their long record, forecasts from 1926, trained from 1881.
- EncompassingThe test the gap must pass.
The encompassing test. The gap adds something beyond the earnings yield less the real yield at a horizon only if both hold: in one regression with both, the gap's Hodrick t-statistic exceeds 1.96; and out of sample the Harvey–Leybourne–Newbold test rejects that the benchmark's forecasts encompass the gap's (p below 0.05). The rule was set before the estimates. The earnings yield less the real yield is the earnings yield less the long rate plus trailing ten-year CPI inflation — a real yield that needs no index-linked bonds and so runs from 1881.
- GapsMonths the identity cannot price.
Months the identity cannot price. The two-stage gap needs (1 + g_L) + H(g_S − g_L) above zero and r − g_L above 0.5pp. In the real-time run 12 months fail the first (2010: the January estimate after the 2008–09 write-downs put near-term growth so far below zero that the bracket turned negative), and 0 the second; they are left out of every statistic, never filled.
- LimitsWhat the record can bear.
What the record can bear. The gap exists from 1975 (two-stage) and 1962 (one-stage): 4 and 5 non-overlapping ten-year windows. The benchmarks run from 1881: 13. Payout policy changed over the century (lower dividend payout after 1950, buybacks after 1990), which moves the right k; the detail fold shows the record at k = 0.75 and 0.95.
- DatesData dates.
Data dates: Shiller's workbook read on 17 Nov 2026 (prices to Nov 2026, earnings to Sep 2026); the Philadelphia Fed's files read on 17 Nov 2026 (the last real-time release Nov 2026, surveys to Nov 2026); BEA's national accounts read on 16 Nov 2026; the Valuation page's inputs 16 Nov 2026; made from data read on 17 Nov 2026.
- LicenceOfficial data, Shiller's index series, the rest calculated on this site.
The Philadelphia Fed's data and BEA's accounts are free to use. Shiller's series is shown with its citation. The histories, the model and the statistics are computed here.
How the numbers are made
shiller ──▶ price, trailing earnings, dividends, CPI and the long rate, each read at its lag (earnings six months, CPI two) ──▶ monthly inputs known at t rtdsm ──▶ nominal output's compound growth over twenty years in the release current at t, and the same quarters in today's release ──▶ the anchor before 1992, real-time and revised spf ──▶ ten-year real GDP median + ten-year CPI median, the latest surveys known at t ──▶ the anchor from 1992 monthly inputs known at t + the anchor before 1992, real-time and revised + the anchor from 1992 ──▶ the trend-and-cycle model: the Earnings page's model on index EPS, re-estimated each January, filtered each month; its expected growth over two years ──▶ near-term growth g_S monthly inputs known at t + near-term growth g_S + valuation ──▶ the two-stage fair-value model: the implied premium each month, its expanding median, the gap at that median (k and H as Valuation) ──▶ the gap histories the gap histories + monthly inputs known at t ──▶ long-horizon regressions: long-horizon regressions, Hodrick and Newey–West errors, the bootstrap under the null, expanding-window forecasts, Clark–West and the encompassing test ──▶ the record the record + valuation ──▶ the ten-year real return at today's readings, with each rule's out-of-sample error as its range (model-conditional) ──▶ the expected return bea ──▶ the real-time set's latest column against BEA's current national accounts, quarter by quarter ──▶ a cross-publisher check
Sources
- licensedRobert J. Shiller, U.S. stock market data 1871–present (ie_data.xls): monthly S&P composite price, 12-month trailing reported earnings and dividends, CPI, the long-term government bond yield and the real total-return index
Robert J. Shiller, Irrational Exuberance data (shillerdata.com); index data are S&P Dow Jones Indices property. · https://shillerdata.com/
- publicFederal Reserve Bank of Philadelphia, Real-Time Data Set for Macroeconomists: nominal and real output (NOUTPUT, ROUTPUT), quarterly observations as published at each quarter since 1965 (QvQd workbooks)
Source: Federal Reserve Bank of Philadelphia, Real-Time Data Set for Macroeconomists · https://www.philadelphiafed.org/surveys-and-data/real-time-data-research/real-time-data-set-for-macroeconomists
- publicFederal Reserve Bank of Philadelphia, Survey of Professional Forecasters: median ten-year real GDP growth (RGDP10, first-quarter surveys since 1992) and ten-year CPI inflation (CPI10, since 1991 Q4) forecasts
Source: Federal Reserve Bank of Philadelphia, Survey of Professional Forecasters · https://www.philadelphiafed.org/surveys-and-data/real-time-data-research/survey-of-professional-forecasters
- publicU.S. Bureau of Economic Analysis, NIPA Tables 1.1.5 and 1.1.6 (nominal and real GDP), quarterly, current release (NipaDataQ.txt) — the cross-publisher check of the real-time set's latest column
Source: U.S. Bureau of Economic Analysis · https://apps.bea.gov/national/Release/TXT/NipaDataQ.txt
- computed hereThe Valuation page's published gaps and identity inputs, read through the site's manifest as reference copies
Computed here on the sources that page lists.
- computed hereComputed here: the payout identity month by month, the Earnings page's unobserved-components model re-estimated on what was known each January, long-horizon regressions with Newey–West and Hodrick errors, a bootstrap under the null, and expanding-window forecasts
Computed here.
Pages this one copies from
Reproduction checks
| Check | Published | Computed | Tolerance | Result |
|---|---|---|---|---|
| The identity's settings are the Valuation page's: payout k 0.8448 (its measured payout, cash returned ÷ as-reported earnings as it prints them), fade 6 years (H = 3), near-term horizon 2 years and floor 0.5pp, read from its specification, equal what its snapshot publishes (settings that differ) | 0 | 0 | 0 | pass |
| The identity (the site's one implementation) at the Valuation page's published inputs — k 0.84, E0/P 0.03984559, g_S 5.768%, g_L 4.0%, r 10.290%, H 3 — returns its published two-stage gap −41.504% (abs difference; the tolerance, 2.2e-05, is the copies' rounding carried through the identity) | -41.504 | -41.504 | 0 | pass |
| Shiller's real total-return index recomputed month by month as the prior value × (real price + real dividend ÷ 12) ÷ prior real price, Feb 1871 – Sep 2026 (1868 months; max abs relative difference) | 0 | 0 | 0 | pass |
| The real-time set's latest nominal output (as published Nov 2026) is BEA's national accounts as first fetched after it (2026-11-16 fetch) rounded to $0.1bn, in every quarter both carry but its latest, Jan 1947 – Apr 2026 (318 quarters; quarters that differ) | 0 | 0 | 0 | pass |
| The real-time set's latest real output (as published Nov 2026) is BEA's national accounts as first fetched after it (2026-11-16 fetch) rounded to $0.1bn, in every quarter both carry but its latest, Jan 1947 – Apr 2026 (318 quarters; quarters that differ) | 0 | 0 | 0 | pass |
| Calendar safety, earnings: in every one of the 1,871 months the earnings read are at least 6 months old and sit at a quarter-end month from Mar 1926 (a December before, when the figures are annual) (months that break it) | 0 | 0 | 0 | pass |
| Calendar safety, CPI and the long rate: CPI at least 2 months old; the long rate of the month itself from Apr 1953 and a January at least 12 months back before it, in every month (months that break either) | 0 | 0 | 0 | pass |
| Calendar safety, the anchor: every real-time release read was known at its month (the month after its quarter's middle month) and ends before its own quarter; today's release is used only before the first real-time one (Dec 1965); every survey read was known at its month (851 months; months that break it) | 0 | 0 | 0 | pass |
| The survey anchor is the two medians added, real GDP ten years + CPI ten years, in every month from Mar 1992 (max abs difference, pp) | 0 | 0 | 0 | pass |
| Where no real-time release applies (before the first, and in the survey years) the two runs read the same anchor (max abs difference, pp) | 0 | 0 | 0 | pass |
| The model's settings are the Earnings page's: the jump cut (|z| > 3) and the bounds on φ read from its specification equal what its snapshot publishes (settings that differ) | 0 | 0 | 0 | pass |
| Every estimation's log-likelihood is reproduced by the independent Kalman filter (114 estimations, both runs; max abs difference) | 0 | 0 | 0 | pass |
| The monthly filter (with the jump rule for quarters after an estimation) is the core filter line for line at every estimation date (max abs difference in the filtered state) | 0 | 0 | 0 | pass |
| Calendar safety, the model: every quarter of earnings growth estimated or filtered at a month ended at least 6 months before it (months and estimations that break it) | 0 | 0 | 0 | pass |
| The trailing earnings the identity divides by are the last quarter the model has seen, at every month (months that differ) | 0 | 0 | 0 | pass |
| At every month the premium the identity implies returns the price exactly (both runs, both identities; max abs difference in model price ÷ price) | 0 | 0 | 0 | pass |
| The one-stage gap is the payout model written out, k(1 + g)·E/P ÷ (r − g) − 1, at every month (max abs difference, %) | 0 | 0 | 0 | pass |
| Least squares and Newey–West(h) errors for every measure and horizon (28 regressions) equal statsmodels' OLS with HAC at h lags, no small-sample correction (max abs difference) | 0 | 0 | 0 | pass |
| Hodrick's errors: the construction at one month with the regression's own residuals equals statsmodels' White (HC0) errors for every measure (7 regressions; max abs difference) | 0 | 0 | 0 | pass |
| The bootstrap's actual slope is the regression's slope on the same months, every measure and horizon (max abs difference) | 0 | 0 | 0 | pass |
| Campbell–Thompson's out-of-sample R² equals the Goyal–Welch cumulative squared-error difference at the last origin divided by the historical mean's squared errors, for every measure and horizon on both records (40; max abs difference) | 0 | 0 | 0 | pass |
| Every range this page prints carries its kind and is well formed (4 ranges: three out-of-sample error ranges, one payout assumption range; problems) | 0 | 0 | 0 | pass |
| From Jan 1993, the first estimate after the surveys give the anchor, both runs estimate the same model on the same quarters: near-term growth is the same in every month (407 months; max abs difference, pp) | 0 | 0 | 0 | pass |
| Write-down months: Shiller's trailing reported EPS last stood at its high to date in Jun 2007 (84.92) and fell to 6.86 by Mar 2009 (−91.9%) before the first month the identity cannot price, Jan 2010 (recomputed from the workbook: the peak is a running maximum, the trough the minimum after it; abs difference, %) | -91.922 | -91.922 | 0 | pass |
| What followed: today's gap (−52.1%) sits in decile 1 of the 607 monthly readings (−80.2% to −47.2%); 61 months share it and 47 have a known ten-year return (months outside the decile's edges) | 0 | 0 | 0 | pass |
| What followed: 33 of the 47 returns sit below the average ten-year real return of every month with one (6.65% a year, 1,751 months since 1871), recounted (abs difference) | 33 | 33 | 0 | pass |
| What followed: the scatter's pairs are every month with both a gap and a known ten-year return (months that differ) | 0 | 0 | 0 | pass |
| Register: the Valuation page's gap copied here (−41.504%) is the two-stage gap that page publishes (−41.504%) (abs difference, %) | -41.504 | -41.504 | 0.001 | pass |
| Damodaran's monthly series copied from the Valuation page: 217 months, 2008-09 – 2026-10, its last value his ERP (T12m) 3.70% (abs difference) | 3.7 | 3.7 | 0 | pass |
| The Valuation page: the source page is published with all its checks passing (dated 17 Nov 2026, 278 checks) | 1 | 1 | 0 | pass |
| The Valuation page: the site still serves the page this one copied (dated 17 Nov 2026) | 1 | 1 | 0 | pass |
| The Earnings page: the source page is published with all its checks passing (dated 17 Nov 2026, 92 checks) | 1 | 1 | 0 | pass |
| The Earnings page: the site still serves the page this one copied (dated 17 Nov 2026) | 1 | 1 | 0 | pass |
| The Valuation page: every copied observation equals the source page, matched by date (217 values in 1 arrays; mismatches) | 0 | 0 | 0 | pass |
| The Valuation page: every copied single value equals the source page (19 values; mismatches) | 0 | 0 | 0 | pass |
| The Earnings page: every copied single value equals the source page (3 values; mismatches) | 0 | 0 | 0 | pass |
- Robert J. Shiller, U.S. stock market data 1871–present (ie_data.xls): monthly S&P composite price, 12-month trailing reported earnings and dividends, CPI, the long-term government bond yield and the real total-return index — Robert J. Shiller, Irrational Exuberance data (shillerdata.com); index data are S&P Dow Jones Indices property.
- Federal Reserve Bank of Philadelphia, Real-Time Data Set for Macroeconomists: nominal and real output (NOUTPUT, ROUTPUT), quarterly observations as published at each quarter since 1965 (QvQd workbooks) — Source: Federal Reserve Bank of Philadelphia, Real-Time Data Set for Macroeconomists
- Federal Reserve Bank of Philadelphia, Survey of Professional Forecasters: median ten-year real GDP growth (RGDP10, first-quarter surveys since 1992) and ten-year CPI inflation (CPI10, since 1991 Q4) forecasts — Source: Federal Reserve Bank of Philadelphia, Survey of Professional Forecasters
- U.S. Bureau of Economic Analysis, NIPA Tables 1.1.5 and 1.1.6 (nominal and real GDP), quarterly, current release (NipaDataQ.txt) — the cross-publisher check of the real-time set's latest column — Source: U.S. Bureau of Economic Analysis
- The Valuation page's published gaps and identity inputs, read through the site's manifest as reference copies — Computed here on the sources that page lists.
- Computed here: the payout identity month by month, the Earnings page's unobserved-components model re-estimated on what was known each January, long-horizon regressions with Newey–West and Hodrick errors, a bootstrap under the null, and expanding-window forecasts — Computed here.