Current regime and pricing: which engines are cheap today, and which are expensive

Question. Every other page in this repository asks whether an engine works on average. This one asks a different question: of the return engines this portfolio can reach — US beta, international beta, emerging beta, a value tilt, momentum, trend, credit, duration, cash and gold — which are cheaply priced right now and which are expensive, and does “cheap” predict anything at all?

Decision it informs. Whether any weight in the current construction should move because of the level of a market variable in August 2026, rather than because of its long-run average. It does not set the equity share, does not forecast a market, and does not supersede any page below.

Out of scope. The US CAPE level itself and the US/international split, which valuation and the allocation owns; sleeve selection, which alternative sleeves owns; and any moving-average or trend timing rule, which timing rules owns.

as of 2026-08-23. Every measured figure regenerates from studies/_current_pricing_tables.py, run with uv run python -m portfolio_edge.studies._current_pricing_tables; the arithmetic is pinned in research/tests/unit/test_studies_current_pricing.py. Everything measured here is exploratory: no specification was frozen before the numbers were seen, no experiment is registered, and nothing below may support a promoted claim.

---

Conclusion

1. The two things this portfolio does not own are the two things that are cheap, and the two things it does own are the two that are dear. Long real yields sit at the top of their record — the thirty-year TIPS real yield of 2.95% is above 99.5% of its own history since 2010 — while US equity and corporate credit are both near the expensive extreme of theirs. That is an uncomfortable arrangement and it is not on its own a reason to trade, for the reason in point 4.
2. Credit is the clearest extreme on the longest record available. The Moody’s Baa−Aaa quality spread — what an investor is paid for holding weaker corporate credit rather than the strongest — is 0.43 pp, wider in 98.2% of the 1,201 months since 1926 and wider in every one of the 439 months since 1990. Investment-grade OAS is 82 bp and high yield 275 bp. It has never been cheaper to be a borrower in this record.
3. The predictable part of the opportunity set is the part this portfolio has excluded. Of twenty-seven conditioning relations tested, the only ones that survive both a Hodrick 1B statistic and a positive out-of-sample record predict bond and credit returns, not equity returns. Nothing tested here predicts the equity premium out of sample.
4. Nothing tested here predicts equity returns, and one relation shows exactly how a predictor fools you. The term spread’s five-year equity regression reports a Hodrick t of 2.32 and an in-sample R**2 of 0.122 — and an out-of-sample R**2 of −0.175. That is the same signature the CAPE regression carries. A large t on an overlapping long-horizon regression is not evidence until it has beaten a rolling mean on data its coefficients never saw.
5. The real cash rate has no conditioning record, and the premise that it does is wrong. Across six responses and two horizons its best out-of-sample R**2 is +0.007. Its level is also less remarkable than it feels: a +0.19 pp ex-post real bill sits at the 45th percentile of a century, not the top. It is only unusual against the post-2008 window, where the real bill averaged −1.06 pp and was positive in 64 of 211 months.
6. The value spread is wide and has narrowed sharply. Cheap US stocks trade at 7.5x the book-to-market of expensive ones once the ratio is moved to current prices, against a hundred-year median of 4.95x — the 81st percentile. Eighteen months ago it was near the 91st. It is a real state variable and its predictive content is unresolved: the fitted effect is 2.6 pp/yr per standard deviation against a design that could only have detected 7.2.
7. Gold is the most expensive thing on this page relative to its own history, at the 98.5th percentile of the real price since 1975 and above both the January 1980 and August 2011 real peaks.
8. Momentum and trend are not priced at all. Neither has a valuation state variable, no source in this repository publishes one, and inventing one would be a search over definitions. unpriced is the honest label.
9. What to do: nothing to the allocation. Two negative decisions get firmer — do not open a credit sleeve at this spread, and do not open a gold sleeve at this price — and one contingent statement is added: if a defensive allocation is ever opened, the inflation-linked leg is the part that is cheap. See §5.

---

1. Where every priced input sits

Percentile means the share of that series’ own history strictly below today’s reading. It is arithmetic on a sample. It is not a forecast, and §2 is where the forecasting claims live.

1.1 US equity

Owned by valuation and the allocation, which holds the derivation, the four disagreeing CAPE readings and the buyback and concentration adjustments. Repeated here only so the comparison in §3 has a row.

The two excess-yield rows are the same construction with a different real rate and they disagree by nearly a percentage point. The forward multiple disagrees with both. That tension is unresolved and this repository cannot settle it.

1.2 International

Siblis Research, read 2026-08-22, all three inside one methodology: a 1.70x US premium over developed ex-US and 1.85x over EM.

No percentile is available for that spread and this page will not invent one. The only long cross-country valuation panel in the cache — Jordà–Schularick–Taylor R6 — carries a dividend yield, not a CAPE, and ends in 2020; its US-minus-panel log dividend-yield spread reads −0.259 in 2020, the 19th percentile of 150 years. That describes 2020. Siblis publishes no history this code can read. The relative-valuation question is therefore measured in level and unmeasured in percentile, and valuation and the allocation already records that the cross-country relation is undetectable after 1990.

1.3 Real rates, the nominal curve, and cash

FRED, refetched 2026-08-23; each percentile is taken over that series’ own record, which for the TIPS series is short.

Three readings.

Nominal yields are ordinary; real yields are extreme. A 4.69% ten-year is the 43rd percentile of sixty-four years — unremarkable. The same maturity’s real yield is the 94th percentile of everything the TIPS market has ever printed, and the thirty-year real yield of 2.95% has been equalled or exceeded on 19 of 4,128 days since 2010. The difference is that inflation compensation is priced near normal (2.34% breakeven, 69th percentile) while the real component is not. The percentile is short-windowed and the level is not: 2.95% real is also the best contractual real yield the US Treasury has offered in this record, and the July 2026 ten-year TIPS auction cleared at a real yield of 2.438%, the highest since October 2008 (tipswatch, as of 2026-07-23).

The curve is positive but flat by its own standards — the 10y−3m slope of +0.82 pp is the 31st percentile. The long end is the exception: 30y−10y at +0.54 pp is the 66th percentile, so the extra term compensation available today is concentrated beyond ten years.

No term-premium estimate is supported by this cache. The Adrian–Crump–Moench and Kim–Wright decompositions are not in it, and a slope is not a term premium: it mixes expected future short rates with the premium and no arithmetic here separates them. unresolved, and honestly so.

1.4 Cash in real terms

The two conventions disagree by 1.2 pp because realised inflation over the last year (3.54%) is well above what the bond market prices for the next ten (2.34%). Both are reported because neither is definitive: the ex-post measure is the one with a century of history, and the ex-ante measure is the one an investor actually chooses against.

Against the post-2008 record cash is well paid; against the century it is ordinary. Since 2009 the ex-post real bill averaged −1.06 pp and was positive in 64 of 211 months; today’s +0.19 pp is the 74th percentile of that window and the 45th of the full one. The sentence “a positive real cash rate changes the whole opportunity set” is true about the last fifteen years and false about the last hundred.

1.5 Credit

A data-contract finding first, because it decides what can be said. ICE BofA’s option-adjusted spreads on FRED are truncated to three years at source, exactly as the total-return siblings are: measured 2026-08-23, BAMLC0A0CM and BAMLH0A0HYM2 return 786 and 787 daily rows beginning 2023-08-22. No percentile of a current OAS against its own history can be computed from FRED, and the “35th percentile” in that row means only “35th percentile of the last three years”, which is not a useful statement. The Moody’s series, which begin in 1919, are the long-history substitute.

The two long measures disagree, and the disagreement is the finding. The quality spread — Baa yield less Aaa yield, both long corporate — is at the 1.8th percentile of a century and has not been this narrow in any month since 1990. The default spread — Baa less the ten-year Treasury — is at the 46th percentile of the same century. They differ because Baa−10y is not a clean credit spread: Moody’s Baa is a long-maturity corporate yield, so subtracting a ten-year Treasury leaves a term premium mixed into the credit premium, which is the same confound alternative sleeves documents for the unhedged corporate leg. On the shorter windows where the two can be compared the story converges: Baa−10y is at the 8th percentile since 2010 and the daily series at the 14th since 1986.

Read together with the 82 bp OAS, all four measures say the same thing: an investor is being paid unusually little for corporate credit risk right now.

1.6 The value spread

The valuation gap between cheap and expensive stocks — how much more book value a dollar buys in the cheap half of the market than in the expensive half. Ken French’s own value-weighted BE/ME averages for the six size × book-to-market portfolios, expressed as a log ratio and averaged across the big-cap and small-cap pairs.

Hundred-year median: 4.95x.

The stale-price problem is the reason two rows exist. French’s BE/ME is fixed at formation: book from the prior fiscal year over market equity at the prior December. The file is built from the 202606 CRSP database and its last row is 2026-06, but that row belongs to the June-2025 formation, whose market equity is dated December 2024 — so the raw spread is twenty months old on prices. Scaling each side by its own cumulative return since then restores the price leg: big-cap value returned +62.3% against big-cap growth’s +24.1% over that stretch, which narrows the spread by 0.27 log units. Using total rather than price returns biases this downward, because the cheap side pays the larger dividend, so 7.54x is a floor rather than a point estimate.

The reading: wide, and narrower than it was. At the 81st percentile the cheap half of the market is still priced at a larger discount than in four fifths of the last century. But the spread has spent eighteen months closing while the tilt delivered, and it is no longer near the 2022 extreme.

1.7 Gold

For scale: the January 1980 peak was $2,786 and the August 2011 peak $2,669 in the same money, against a period median of $1,170. The 2026-08-21 fix of $4,582 is higher still. Gold has no cash flow and therefore no valuation ratio; the real price is the only state variable available, and it is near the top of its record.

1.8 What is not priced

Momentum and trend have no valuation state variable. French publishes no BE/ME for the momentum portfolios, AQR’s TSMOM series carries no valuation measure, and there is no accepted one in the literature this repository has read. Constructing one would be a search over definitions with no pre-registration, which is how a spurious conditioning rule gets made. They are unpriced, and that is a property of the engines, not a gap in the data.

One arithmetic note that is not a forecast: a fully collateralised futures programme earns the bill rate on its collateral, so a 3.71% bill adds roughly 3.6 pp/yr to a managed-futures sleeve’s total return that a 0.05% bill did not. For the stacked wrapper this nets to approximately nothing, because the same rate is what the wrapper pays to finance its equity exposure. The level of rates is close to neutral for the construction actually held.

---

2. What each conditioning variable has actually predicted

Monthly, Goyal–Welch 1926-07…2025-12 (the workbook’s return columns are not spliced forward). Slopes are reported as pp/yr of response per one standard deviation of the predictor, so predictors in different units compare. MDE is the smallest slope this design would have detected, at 5% two-sided size and 80% power — an estimate below it is one the study could not tell apart from zero, in either direction. t is Hodrick 1B, which does not let the number of estimated quantities grow with the horizon; the Newey-West statistic is kept beside it because the gap between them is the point.

Twenty-four macro rows were run; the fifteen above are the ones that carry a reading. The full table prints from the module.

And the value spread, on annual data where a one-year horizon needs no overlap correction at all. Predictor: the spread at the June formation date. Response: the subsequent Fama–French HML return, July to June.

Five readings, and they are the substance of this page.

Credit spreads predict credit returns and nothing else. The default spread’s relationship to subsequent corporate-over-government excess return is the most reliable thing in either table: positive out-of-sample at both horizons, R**2 of 0.256 in sample at five years. Against equity returns the same predictor is inert — t of 0.29, an out-of-sample R**2 of −0.82. A credit spread is a signal about credit, not about risk appetite generally.

The term spread predicts bond returns, with the highest statistic on the page. Hodrick t of 3.14 and an out-of-sample R**2 of +0.127 at one year. It is also the row that demonstrates the failure mode: the same predictor against equity at five years reports t = 2.32 in sample and loses to a rolling mean out of sample by 17.5%. A statistic and a record are different claims and this table shows them disagreeing inside one predictor.

The real cash rate is noise. Three responses at two horizons, best out-of-sample R**2 +0.007, best Hodrick t 1.75, and against equity the sign is negative and unresolved. It collects a suggestive grade on one cell only because the grader’s out-of-sample gate is “positive”, and +0.007 is a rounding error against a rolling mean rather than a result. The hypothesis that a real cash rate is a well-evidenced conditioning variable is not supported on this panel. It is a fact about the price of cash, not a signal about anything else.

The value spread is the most interesting unresolved here. Its Hodrick statistics are weak — 1.01, 0.71, 0.93 — and every point estimate is far below its own MDE. But its out-of-sample R**2 is positive at every horizon and every start, which is exactly what CAPE fails to do. Directionally right, economically plausible, and not resolvable on a hundred years of annual data. Anyone sizing a tilt on it is sizing on faith.

Nothing here should be read as a promotion, and the multiple-testing arithmetic says so. Twenty-seven relations were tested. At a 5% two-sided size about 1.4 spurious rejections are expected by chance, and two rows reach |t| ≥ 3.0 — the Harvey–Liu–Zhu hurdle this repository uses elsewhere. Both are bond-return relations, they share a response, and their predictors are related. Two survivors out of twenty-seven, concentrated in one response, is weak-to-moderate evidence and not a discovery.

---

3. The ranking, with the two claims kept apart

Column one is a measurement. Column two is not a claim this page makes. “Cheap by percentile” says where a price sits inside its own history. “Expected to deliver” is a forecast, and §2 shows that the forecasting record behind almost all of these is poor. The right-hand column therefore records only whether a conditioning relation with an out-of-sample record exists for that engine — not what it forecasts.

Credit is last and it is the only entry where “expensive” and “has a conditioning record” coincide. For every other engine, either the price is not extreme or the conditioning evidence is absent. That coincidence is what makes §5’s one firm conclusion possible.

---

4. Three ways this ranking could mislead

A short window makes an extreme look more extreme. The TIPS real yield’s 99.5th percentile is taken over sixteen years, all of them inside the era of quantitative easing. Before 2003 there is no TIPS market to measure. The level — a 2.95% contractual real yield — carries more information than the percentile does, and the percentile should not be read as “once in two hundred years.”

A percentile is not a probability. Nineteen of 1,748 months have equalled today’s CAPE and eighteen of them are one episode; fewer than 2% of the 1,201 months since 1926 have carried a narrower quality spread, and they are not spread evenly across the century. A count of past occurrences drawn from two or three episodes is a description, not a frequency, and valuation and the allocation makes the same point about the CAPE drawdown evidence.

Revised history flatters every backward-looking rule here. FRED, Goyal–Welch and Ken French all rebuild their full history on each release and publish no vintage archive. Every percentile and every out-of-sample R**2 above uses a revised series that no historical investor held. That biases the conditional results in their own favour, and most of them still fail.

---

5. What should change because of the level

The bar is high, and it should be. This session already established that conditioning the equity share on CAPE loses before costs, that the one valuation rule with a real gross edge is eaten by tax, and that timing rules on the equity sleeve fail deflation. So the default answer to “should this level move a weight” is no.

The equity allocation: no change. Nothing in §2 predicts equity returns out of sample. The six-line construction — stacked equity + managed futures ~25%, US total market ~25%, US value ~15%, international core ~25%, international momentum 5%, emerging value 5% — is unaffected by anything measured here. The existing advice in valuation and the allocation stands unchanged: widen the drawdown assumption, direct new contributions rather than realised gains toward the international side, and do not run a valuation-conditional weighting rule in a taxable account.

Credit: do not open the sleeve at this spread. This is the one place where the level changes a decision, and it strengthens a caution alternative sleeves already raised. Three things now line up:

• duration-hedged credit is the largest candidate addition on that page, at a measured +2.13%/yr over 1,062 months at 4.12% volatility;
• that premium was earned across the full range of spread regimes, and the current gross spread is 82 bp with the century-long quality spread at its 1.8th percentile and the tightest reading in 439 months;
• credit is the one engine here whose conditioning variable has a positive out-of-sample record — a wider default spread has bought more subsequent credit excess return, OOS R**2 +0.107 at one year and +0.154 at five.

The consequence is not that the engine is refuted. It is that the one place the evidence says entry price matters is also the place the entry price is worst, and a 14–24 bp wrapper plus expected default loss is being charged against 82 bp of gross spread. Defer, size to the spread, or build in over time. A specific, falsifiable reopening condition: revisit when the Moody’s Baa−Aaa quality spread returns to its century median of 0.90 pp, which is a level, not a forecast, and can be checked from the same series in one command.

Gold: do not open the sleeve at this price. Alternative sleeves already found gold return-dominated at examined weights on a +1.75%/yr measured excess return with a −91% drawdown. The real price at the 98.5th percentile of fifty-one years, above both prior real peaks, does not make that case better.

The value tilt: keep it, do not enlarge it. The spread is genuinely wide and the direction of the evidence is favourable. It is also unresolved at a factor of nearly three below its own detection floor, and it has already closed by 0.27 log units in eighteen months. The 15% US value and 5% emerging value weights need no adjustment from this page.

If a defensive allocation is ever opened, the inflation-linked leg is the cheap part. The portfolio holds no bonds, and the recommendation is explicit that the defensive allocation should be decided by the withdrawal path and the ability to persist, not by a return forecast. That does not change. What changes is the price of the instrument if that decision is ever made: a 30-year TIPS at 2.95% real is a contractual line in the certainty class this repository reserves for statutes, and it is the best such line in the record. Long-horizon, contributions of 5–15%/yr, no near-term withdrawal need — this investor has no current use for it. The note is for the review, not for today.

Cash: no change, and the framing was wrong. A +0.19 pp ex-post real bill is the 45th percentile of a century. It is unusual only against the post-2008 window. And it predicts nothing.

---

6. Provenance, and what could not be measured

Every FRED series is refetched on each run and its sha256 and retrieval timestamp print at the top of the module’s output. The Goyal–Welch workbook is sha256:1e4b6527…, retrieved 2026-08-17, data through 2025-12. Ken French’s 6_Portfolios_2x3 is sha256:06108313…, refetched 2026-08-23 and byte-identical, built from the 202606 CRSP database. The LBMA PM fix was refetched 2026-08-23.

A hash identifies the bytes used. It does not make a revised series point-in-time, and none of these sources publishes a vintage archive this code can read.

Four things could not be measured and are recorded as gaps, not as excuses.

• A percentile for any current OAS. ICE BofA’s spread series on FRED are capped at three years at source, the same cap the total-return siblings carry. The Moody’s quality and default spreads are the long-history substitutes and they are not the same instrument.
• A percentile for the US-versus-international valuation gap. The only long panel in the cache ends in 2020 and measures a dividend yield, not a CAPE.
• A term-premium decomposition. Neither the ACM nor the Kim–Wright series is in the cache, and a curve slope is not a term premium.
• A value spread outside the US. French’s developed-ex-US and emerging portfolio files carry returns, firm counts and market caps but no BE/ME block, so the same construction cannot be run for the international core or the emerging value sleeve. That is the single most useful missing input on this page, because it is where two of the portfolio’s six lines live.

The most informative next test, if this page is ever revisited: obtain a BE/ME or book-to-price series for developed ex-US and emerging markets and repeat §1.6 and §2 on them. Everything else here is either measured or shown to be unmeasurable from committed sources.