When the Buffer Runs Thin

Part 1 established that a four hundred basis point rise in the real risk-free rate was absorbed almost entirely by the compensation equity investors receive over it — the forward earnings yield less the ten-year real yield, which I call the spread — rather than by share prices falling. This part asks what that compression has meant, and what followed the only previous occasion on which it went further.

What the spread cannot tell you

What cannot be determined from price data is which component did the absorbing — a smaller required premium, or faster expected growth, or both.

It is worth being precise about why, since analysts do publish growth forecasts. Consensus estimates for the next year or two are available and widely followed. But the growth term in this identity is not that number. It is the rate at which earnings are expected to compound indefinitely, beyond any forecast horizon, and nobody publishes that. A market can agree entirely on next year's earnings and still disagree completely about what the decade after looks like. So anyone asserting that the equity risk premium has fallen to around one per cent is supplying a long-run growth assumption of their own and presenting the result as a measurement.

The compression tracks the policy cycle, and the last leg is recent

Two readings worth drawing out.

The bulk of the compression happened during the hiking cycle. The spread narrowed by 340 basis points in total between December 2021 and September 2026. Two hundred and thirty-three of those basis points — sixty-eight per cent of the whole move — arrived in the nineteen months between December 2021 and the final hike in July 2023.

The easing cycle re-rated the market while the spread kept narrowing. The multiple rose from 19.89x to 22.04x between the last hike and the last cut, even as the spread fell from 3.43 to 2.63. This was not a market starved of profit growth: forward earnings rose 34.6% over the same stretch. Investors were simply willing to pay more for each dollar of those earnings while their compensation over the risk-free real rate shrank. That is the clearest statement of what the last three years have been.

September is the month absorption stopped working

For thirteen months the spread held between 2.60 and 3.13 while the real yield climbed from 1.82% to 2.43%. The earnings yield rose in step and the mechanism worked.

In September it did not. The real yield rose 48.7 basis points in a single month; the forward earnings yield rose 13.9. The remaining 35 basis points went into the spread, taking it to 2.3635.

The speed of the move is, I suspect, the explanation. A 48.7 basis point rise in the real yield inside one month is the tenth largest in 318 months of this series — only nine months since 2000 have seen a faster one. And of those ten, September 2026 is the only occasion on which the spread was already compressed: every other instance left the spread at 3.56 or wider, most of them far wider.

That matters because of how absorption works. The earnings yield rises as analysts mark up their forecasts, and analysts revise over weeks, not days. When the discount rate moves faster than estimates can follow, there is nothing to absorb it with, and the gap falls straight into the spread. September's move was concentrated in its final days, which left no time at all for the usual mechanism to operate.

That is the lowest reading since May 2002

Of 319 monthly observations since January 2000, thirty have been lower than the 2.364 recorded at the end of September 2026 — the 9.4th percentile. All thirty fall in one unbroken run, January 2000 to May 2002. There is no second episode.

Today's reading is the minimum of the current regime and below the minimum of every regime except the dot-com bust.

What happened the last time, and the caveat that must travel with it

The thirty months in which the spread sat below today's level run consecutively from January 2000 to May 2002. For each of them I asked a single question: starting from that month, what had happened two years later? January 2000 is measured against January 2002, February 2000 against February 2002, and so on through May 2002 against May 2004. Thirty starting points, each with its own two-year outcome.

Peak to trough on month-end closes, the index fell from 1,517.68 in August 2000 at 23.0x to 815.28 in September 2002 at 14.6x — −46.3%, which understates the intraday extremes. Returns here are index returns and exclude dividends.

Those thirty outcomes are not thirty pieces of evidence. They are overlapping windows cut from a single stretch of history: January 2000's two-year window shares twenty-three of its twenty-four months with February 2000's. All thirty together span 4.4 calendar years, where thirty genuinely separate two-year windows would need sixty. The honest reading is that this is one episode observed thirty times over, which makes it a precedent rather than a pattern.

I use it regardless, and the reason is not statistical. The dot-com peak and its unwind is the comparison investors and commentators keep reaching for when they discuss today's market, and it is the obvious analogue for a boom built on a new technology. Measuring an analogy that is already being drawn everywhere seems to me more useful than refusing to measure it on sample-size grounds.

The finding that should unsettle people

Look again at the decomposition of that drawdown. The index fell 22.6% at the median. The multiple contributed −18.9%. Forward earnings contributed −5.6%, and rose in twelve of the thirty windows. The mean change in forward earnings across all thirty was +0.2% — flat.

The market fell by a fifth without needing earnings to fail.

I had assumed, before running this, that prolonged equity weakness would require both higher yields and slower earnings delivery — that the two would have to arrive together. The one episode on record says otherwise. The de-rating did the damage on its own, with earnings expectations broadly intact throughout.

For anyone deciding today between a long bond and staying in equities, that is the uncomfortable part. The equity case does not only depend on earnings continuing to compound. It depends on the market continuing to pay the same price for them.

Three reasons today is not 2002, and they are not trivial

The multiple is materially cheaper. Across that episode the median forward P/E was 21.3x, reaching 25.2x at the peak. Today it is 18.9x. In only two of those thirty months was the forward earnings yield higher than today's 5.28%.

The real yield is lower. The episode's median was 3.48%, its peak 4.33%. Today's is 2.92%. In all thirty of those months the real yield stood above today's.

Two recent observations point the other way. December 2024 and January 2025 carried spreads of 2.397 and 2.422, within a few basis points of today's. Twelve months later the forward multiple was 22.04x against 21.59x, and 21.89x against 22.05x — up 2.1% and down 0.7%. Neither was followed by any de-rating at all. They are the only modern readings near this level, and they do not support the parallel.

What all three amount to is that today is a milder version of the same configuration, not a different one. The spread is compressed in both periods for the same two reasons working together: shares priced above their long-run multiple, against a real yield above its long-run level. Measured across the full twenty-six years, the median forward multiple is 16.7x and the median real yield 1.37%. In 2000–02 the market ran at 21.3x against a 3.48% real yield — both well above. Today it runs at 18.9x against 2.92% — both still above, but by less on each count.

That is why today's spread of 2.364 sits above the 2000–02 median of 1.22 and a long way above its extreme of −0.36. The same machine is running; it has simply not been pushed as hard.

What the evidence supports, and what it does not

The evidence rests on two things: one historical episode, and a statistical tendency that is partly an artefact of how the spread is built. That is thinner than the volume of numbers makes it look, and saying so seems more useful than dressing it up.

The tendency first. Across all 295 overlapping twenty-four-month windows, months beginning with a low spread were typically followed by a falling multiple — a rank correlation of +0.529. Against index returns it is +0.313. Against subsequent earnings it is −0.054, indistinguishable from zero. That last figure is the one that matters: where the spread starts tells you something about the price investors will pay per dollar of earnings, and nothing at all about whether those earnings will grow or shrink. A compressed spread is a warning about valuation, not a forecast of a profits recession.

Now the artefact. The spread and the multiple share a term — one is the earnings yield minus the real yield, the other is one divided by the earnings yield. Measured on the same date they correlate −0.695, which is to say a low spread and a high multiple are largely the same observation wearing two hats. So much of "a low spread precedes a falling multiple" is really "a high multiple precedes a falling multiple" — ordinary mean reversion, requiring no theory about real yields at all.

Anyone familiar with the critiques of yield-gap measures will go further and ask why bother with the spread at all, since the sharpest of them find the earnings yield alone forecasts returns at least as well as any comparison against bonds. As forecasting, I think that is largely right.

But the spread is not being used here to forecast. It is an accounting of where a four-hundred-basis-point rate shock went, and for that the real yield is not optional, because it is the thing that moved. A multiple on its own tells you the market de-rated eleven per cent since 2021. It cannot tell you that the market simultaneously absorbed four points of rising real rates, nor where it put them. That is what the spread is for, and why the compression matters even if it forecasts nothing.

The narrower claim, which I think holds, is this. The compensation an equity investor receives for taking risk instead of a guaranteed real return — the forward earnings yield less the ten-year real yield — is at its lowest in twenty-four years, and it got there because four points of rising real yields were absorbed almost entirely by that compensation shrinking rather than by share prices falling. In September the absorption faltered for the first time in more than a year, on one of the ten fastest monthly rises in the real yield since 2000, arriving too quickly for earnings forecasts to follow. And on the only previous occasion when compensation was thinner, the damage came through the multiple while earnings held up — so continued earnings growth cannot be treated as sufficient protection.

That is the case for looking seriously at long-dated bonds — ten-year and thirty-year Treasuries, and the funds that hold them. Part 3 takes up what to buy: the duration arithmetic, whether a bond fund does what its holders believe it does, whether bonds still hedge equities at all, and how the sums change for an investor earning in Singapore dollars rather than US dollars.

Notes on the data

Source. All figures are from Bloomberg. The historical series runs monthly from January 2000 to September 2026, 319 observations.

Two reference dates. Yield levels and the curve decomposition are the close of business on 9 October 2026. All equity analysis — earnings yields, multiples, spreads, index levels and every historical comparison — uses month-end data to 30 September 2026, which is where the consistent series ends. No calculation in this article combines figures from the two dates.

Definitions.

• Forward earnings yield — the inverse of the forward price-to-earnings multiple.
• Forward earnings per share — derived as the index level divided by the forward multiple.
• The spread — the forward earnings yield less the ten-year real yield on Treasury Inflation-Protected Securities.
• Thirty-year breakeven — derived as the nominal yield less the real yield.

Returns. Index returns are price returns and exclude dividends.

Medians. Regime comparisons use medians throughout, for consistency and because the series contains genuine extremes. Means differ immaterially: across the full sample the forward multiple averages 17.0x against a median of 16.7x, and the real yield 1.25% against a median of 1.37%.

Views expressed are those of the author and do not constitute investment advice.