The trailing price-to-earnings ratio has one fatal property at index level: its denominator is a single year of profits, and corporate profits are cyclical. At a cycle peak the denominator is fat, the ratio looks modest, and the market looks safe at exactly the wrong moment. At a trough the denominator collapses, the ratio explodes, and the market looks lethal at exactly the moment it is cheap.
The cyclically adjusted ratio replaces the denominator with an average of ten years of earnings, each restated into current money. Ten years is a compromise, chosen because it is long enough to span a business cycle and short enough that the businesses in the index have not changed beyond recognition. The construction has four steps and no judgement anywhere in it except the choice of earnings definition.
How is the CAPE ratio calculated?
- Take the index's earnings per index unit for each of the last ten fiscal years, on one consistent definition.
- Take a consumer price index for the same ten periods, on one consistent base.
- Restate each year's earnings into the current period's money by multiplying by (current CPI ÷ that year's CPI).
- Average the ten restated figures and divide the current index level by that average.
Advertisement
Worked Example 1 runs all four steps on an illustrative Indian series and Worked Example 2 on an illustrative United States series. Two properties of the output are worth naming before you see the arithmetic. The ratio will normally sit above the trailing multiple in an economy whose real earnings have been growing, because a growing series has a mean below its latest value. On the illustrative Indian series the ratio is 32.00 against a trailing 23.21, a factor of 1.3786, and the whole of that gap is real earnings growth of 7.48 percent a year compounded over nine years.
The inflation adjustment is not decoration. Skip it, average the nominal earnings, and the same Indian series gives 37.87 rather than 32.00. That is 5.87 ratio points, an 18.35 percent overstatement, produced by nothing but a missing deflator over a period in which the price level rose 52.00 percent. In a high-inflation market the naive version does not merely mismeasure the level; it manufactures a trend, because the deflator's effect grows with the inflation you failed to remove.
Where the inputs come from, and what to date-stamp. For India, index-level earnings per unit and the trailing multiple are published by the index provider on its own factsheets and daily statistics pages, and the consumer price index comes from the national statistical office. For the United States, the long-run reconstruction with as-reported earnings and a consistent deflator is published as a downloadable data set by Robert Shiller at Yale, and the index-level earnings are published by the index provider. Every reading you take from any of these carries the date you took it, because all of them are revised.
What the ratio predicts, what it does not, and the objections that bite
The relationship is real. Start from a low cyclically adjusted multiple and the next decade's real return has generally been good; start from a high one and it has generally been poor. The mechanism is not mysterious. A high multiple is a low earnings yield, and a low earnings yield is a small starting cash claim per unit of price, so the return has to come from growth or from a further re-rating.
The relationship is also weak, and the weakness is the part that gets discarded when the finding is repeated. Use a fitted line of the form ten-year annualised real return = 15.00 − 0.32 × ratio, with a residual standard deviation of 3.90 percentage points and a coefficient of determination of 0.36. Those coefficients are illustrative and are shaped to resemble the published relationship; rebuild them from the Shiller data set before you quote any of them. Take them as given and read what they say.
| Starting ratio | Fitted 10y real return | 90% band | Cumulative real multiple over 10 years |
|---|---|---|---|
| 15.00 | 10.20% a year | 3.78% to 16.62% | 1.450× to 4.651× |
| 20.00 | 8.60% | 2.18% to 15.02% | 1.241× to 4.051× |
| 25.00 | 7.00% | 0.58% to 13.42% | 1.060× to 3.521× |
| 30.00 | 5.40% | −1.02% to 11.82% | 0.903× to 3.055× |
| 32.00 | 4.76% | −1.66% to 11.18% | 0.846× to 2.885× |
| 34.48 | 3.97% | −2.45% to 10.38% | 0.780× to 2.685× |
| 40.00 | 2.20% | −4.22% to 8.62% | 0.650× to 2.285× |
How reliable is the CAPE ratio as a forecast?
The band is 12.83 percentage points wide at every starting point, because it is set by the residual and the residual does not shrink when you are more confident. At a ratio of 32.00 the honest statement is that a decade from now the real value of an index rupee is somewhere between 0.85 times and 2.88 times what it is today, and the point estimate sits at 1.59 times. Anyone who converts that into a decision about next quarter has misread the instrument.
Two numbers make the weakness concrete. A coefficient of determination of 0.36 is a correlation of −0.60 and leaves 64.0 percent of the variation in subsequent ten-year returns unexplained. And one residual standard deviation of 3.90 points, divided by the slope of 0.32, equals 12.19 ratio points. So the noise in a single ten-year outcome is worth more than the entire distance between a ratio of 20 and a ratio of 32.
Then there is a statistical problem that inflates every published t-statistic in this literature. Ten-year forward windows overlap. Across 1,740 months from 1881 to 2026 there are 1,620 overlapping ten-year windows but only 14 non-overlapping decades, a ratio of 115.71. Treat the overlapping windows as independent and the standard errors are understated by roughly 10.76 times. A naive t-statistic of 8.0 becomes about 0.74. The correct treatment uses an autocorrelation-consistent covariance estimator rather than this crude division, and it lands in the same place: the effective sample behind the whole relationship is about a dozen independent observations, which is why the slope is credible and its precision is not.
Do write-offs and buybacks distort the CAPE ratio?
Two accounting objections attack the denominator directly, and both are computable.
The first is the write-off. As-reported earnings absorb goodwill impairments and restructuring charges in the year they are taken, and a single catastrophic year drags the ten-year mean down for a decade, pushing the ratio up with no change in price. Nine normal years at 100 and one year at −50 give a mean of 85 against a clean 100, so the ratio reads 17.65 percent higher than it should. On the illustrative United States series, a 2020 whose real earnings were 200 units lower would lift the ratio from 34.48 to 38.37 by itself.
The second is buybacks. A per-share earnings series grows faster than aggregate earnings whenever the share count is shrinking, and the ten-year mean then sits further below the latest reading than the underlying earnings power justifies. Model it directly: real aggregate earnings growing at 3.00 percent a year with a net buyback of 2.00 percent a year. Without the buyback, current earnings per share sit 1.1382 times the ten-year mean. With it, 1.2383 times. The ratio therefore reads 8.80 percent higher at the same price and the same aggregate earnings. That is not a small correction in a market where the buyback yield exceeds the dividend yield.
Neither objection kills the gauge. Both mean the same practical thing: read the ratio against its own history in its own market, on one earnings definition you never change, and treat cross-country and cross-era comparisons of the level as illegitimate.