Valuation Metrics Cheatsheet for the AI Bubble Debate: CAPE, Forward P/E, ERP

Every few weeks this autumn a new headline has said the market is "more expensive than ever" or "only as expensive as 1999". Both camps usually quote a number: a CAPE near 40, a forward P/E above 22, a risk premium that has shrunk, market cap far above GDP. The numbers are often correct. The conclusions drawn from them often are not, because each metric answers a different question.

The AI trade makes the debate louder. A handful of very large companies carry a large share of index earnings and an even larger share of index expectations. If those expectations are right, today's multiples will look reasonable in hindsight. If they are wrong, the same multiples will look like 2000. No single ratio can tell you which, but knowing what each ratio measures keeps you from arguing past each other.

This cheatsheet covers the five metrics that show up most in that debate: what each one is, how to compute it, what it has historically predicted, what it has not, and where it stood around November 2025 according to dated sources.

Educational content, not financial advice.

1. The one-table summary

The core distinction: valuation metrics are weak short-term timing tools and better (but noisy) long-horizon return estimators. Almost every misuse in the AI bubble debate comes from reading a 10-year signal as a 6-month one.

2. Shiller CAPE

What it is

The cyclically adjusted price-to-earnings ratio divides the real (inflation-adjusted) price of the S&P 500 by the average of real earnings over the previous ten years. Averaging over a decade smooths out recessions and booms, so a single bad or great year does not swing the ratio.

It comes from John Campbell and Robert Shiller, whose 1998 paper "Valuation Ratios and the Long-Run Stock Market Outlook" (Journal of Portfolio Management) showed that price-earnings and dividend-price ratios were poor at forecasting future earnings or dividend growth, but useful at forecasting future stock price changes over long horizons. Shiller popularized the measure in Irrational Exuberance and maintains a public dataset going back to 1871.

What it predicts

• Long-horizon returns, roughly. High CAPE readings have historically been followed by lower average returns over the next ten years, and low readings by higher ones. The relationship is statistically meaningful but the scatter is wide.
• Expectations, not outcomes. Cliff Asness (AQR, "An Old Friend: The Stock Market's Shiller P/E", November 2012) found that, based on history, the 2012 CAPE level of about 22 implied a real return of no more than about 1% a year for the following decade. His caveat is the useful part: "This type of analysis is not much help to market timers, but it can help to set reasonable expectations." The decade after 2012 turned out far better than that central estimate, a reminder that the dispersion around the line is huge.

What it does not predict

• Timing. CAPE can stay high for years. It was above its long-run average for most of the period since the 1990s.
• Crashes on schedule. John Campbell, interviewed by NPR on November 13, 2025: "This isn't something that's going to tell you, oh, there's going to be a crash tomorrow."

Known criticisms

• Accounting changes. Jeremy Siegel ("The Shiller CAPE Ratio: A New Look", Financial Analysts Journal, 2016) argued that changes in GAAP earnings rules, especially write-down rules adopted in the early 2000s, depress reported earnings in downturns and make recent CAPE readings look more expensive than they are. He suggested using national accounts (NIPA) profits as an alternative.
• Payout policy. More buybacks and fewer dividends mean faster per-share earnings growth, which a backward-looking 10-year average understates.
• Composition. The index today has more high-margin, asset-light businesses than in 1950 or 1980. A higher "normal" CAPE is plausible; how much higher is the debate.

Where it stood

In the same NPR piece (November 13, 2025), Campbell said: "The CAPE ratio today is pretty close to 40. And that's higher than it's been at any other time besides the turn of the millennium." He added that "high values of this ratio are associated with low subsequent 10-year returns." That is the right framing: a long-horizon expectation, not a sell signal.

3. Forward P/E

What it is

Price divided by the consensus analyst estimate of earnings per share for the next 12 months. It is the metric strategists and the financial press use most, because it reflects what the market is paying for the profits it expects next.

Where it stood

FactSet's Earnings Insight of November 21, 2025 reported:

FactSet also noted that from September 30 to November 21 the index price fell 2.2% while the forward EPS estimate rose 3.2%. That is the mechanism behind every P/E move: the multiple can fall because prices drop, because estimates rise, or both.

What it does and does not tell you

• Useful for: comparing sectors and markets on a like-for-like basis, and seeing how much of a rally is earnings revision versus multiple expansion.
• Not useful for: long-run return estimates. It uses one year of earnings, and that year is a forecast. Consensus estimates are revised constantly, and they are least reliable at turning points, exactly when you need them most.
• The AI-specific catch: in a concentrated market, the index forward P/E is a cap-weighted blend. A 27.5 multiple on technology and 15.3 on financials average to 21.5. The "market" multiple says little about either.

4. Equity risk premium (Damodaran)

What it is

The equity risk premium (ERP) is the extra return investors demand for owning stocks instead of a risk-free government bond. There are two ways to estimate it:

Aswath Damodaran (NYU Stern) publishes an implied ERP for the S&P 500 at the start of each month and a detailed update each January. His method: start from the index level, project cash returned to shareholders (dividends plus buybacks) using analyst earnings forecasts for the next few years, let growth converge to the risk-free rate, and solve for the internal rate of return. The ERP is that expected return minus the 10-year Treasury yield.

Where it stood

• January 1, 2025 (Damodaran, "Data Update 2 for 2025", January 17, 2025): an implied expected return of 8.91% and an implied ERP of 4.33%, after "netting out the ten-year bond rate on Jan 1, 2025, of 4.58%". In the same post he wrote: "I don't fall into the bubble crowd, since the ERP is not 2%, as it was at the end of 1999."
• October 1, 2025 (Damodaran Online home page): implied ERP of 3.77% using his trailing 12-month cash flows with adjusted payout, with a 10-year Treasury rate of 4.16%.

So the premium narrowed through 2025, but it remained well above the roughly 2% he estimates for the end of 1999.

What it does and does not tell you

• Useful for: pricing stocks relative to bonds. A low ERP means stocks are priced for less compensation per unit of risk. It also feeds directly into discount rates for company valuations.
• Strength vs. CAPE: it is forward-looking and accounts for interest rates, which CAPE ignores.
• Weakness: it is only as good as the cash flow forecasts. If AI-driven earnings growth disappoints, today's ERP was overstated. Damodaran's own response to the "you need assumptions" critique: "the earnings yield approach makes assumptions about both as well."
• Not a timer. A low ERP can persist; it tells you expected compensation is thinner, not when prices will adjust.

5. Buffett indicator (market cap to GDP)

What it is

Total US stock market capitalization divided by US GDP. It gets its name from Warren Buffett's Fortune essay with Carol Loomis (December 10, 2001), where he showed the ratio of the value of US publicly traded securities to GNP over about 80 years and wrote that it "is probably the best single measure of where valuations stand at any given moment." He also pointed out that the ratio had reached an unprecedented level around the dot-com peak.

What it does and does not tell you

• Useful for: a simple, assumption-free view of how big the stock market is relative to the economy that has to support its profits.
• Weaknesses:
• US-listed companies earn a large and growing share of their revenue abroad, but GDP is domestic.
• Corporate profits as a share of GDP are not constant. Higher margins can justify a higher ratio.
• Market cap includes only listed companies. Shifts between private and public markets change the ratio without changing the economy.
• Where it stood: commonly cited versions of the ratio were reported at or near record highs during 2025. Different data sources (Wilshire vs. other total-market indexes, GDP vs. GNP, nominal timing) give different exact numbers, so compare a series only with itself.

6. Earnings yield vs. bond yield (the "Fed model")

What it is

Earnings yield is the inverse of the P/E: E/P. A forward P/E of 21.5 is a forward earnings yield of about 4.65%. The "Fed model" compares this with the 10-year Treasury yield: if stocks yield more than bonds, they are "cheap", and vice versa.

The problem

Cliff Asness ("Fight the Fed Model", Journal of Portfolio Management, 2003) made two points that still hold:

1. It mixes real and nominal. Earnings tend to rise with inflation over time, so the earnings yield is closer to a real yield. The Treasury yield is nominal. Comparing them is apples to oranges.
2. Descriptive, not predictive. The relationship does describe how investors set P/Es (low rates have tended to coincide with high multiples), but historically it had no power to forecast long-term real stock returns, while the plain P/E did.

A better comparison

If you want a stocks-versus-bonds comparison, compare like with like: CAPE earnings yield (1/CAPE, a real quantity) against the real yield on inflation-protected Treasuries (TIPS). Or use the implied ERP, which does the comparison inside a cash flow model.

7. What each metric is for

8. A small Python calculator

This computes the metrics from inputs you supply. The numbers in the example are the dated readings quoted above, or round illustrations where noted; they are not current data.

from dataclasses import dataclass
from statistics import mean


@dataclass
class Market:
    price: float               # index level
    forward_eps: float         # next 12 months, consensus
    trailing_eps: float        # last 12 months, reported
    real_eps_10y: list[float]  # last 10 years of inflation-adjusted annual EPS
    nominal_10y: float         # 10-year Treasury yield, decimal
    real_10y: float            # 10-year TIPS yield, decimal


def metrics(m: Market) -> dict[str, float]:
    cape = m.price / mean(m.real_eps_10y)
    fwd_pe = m.price / m.forward_eps
    return {
        "cape": round(cape, 1),
        "forward_pe": round(fwd_pe, 1),
        "trailing_pe": round(m.price / m.trailing_eps, 1),
        "forward_earnings_yield_%": round(100 / fwd_pe, 2),
        "cape_earnings_yield_%": round(100 / cape, 2),
        # Fed model: mixes nominal and real, shown only to compare
        "fed_model_gap_%": round(100 / fwd_pe - 100 * m.nominal_10y, 2),
        # Like-for-like: real earnings yield minus real bond yield
        "cape_yield_minus_real_yield_%": round(100 / cape - 100 * m.real_10y, 2),
    }


def buffett_indicator(total_market_cap: float, gdp: float) -> float:
    return round(100 * total_market_cap / gdp, 1)


# Illustration: index level and EPS chosen so forward P/E = 21.5 (FactSet, Nov 21, 2025)
# and CAPE = 40 (Campbell, Nov 13, 2025). Treasury 4.16% (Damodaran, Oct 1, 2025).
# The TIPS yield is an illustrative round number.
m = Market(
    price=6_600,
    forward_eps=6_600 / 21.5,
    trailing_eps=6_600 / 26.9,
    real_eps_10y=[165.0] * 10,
    nominal_10y=0.0416,
    real_10y=0.018,
)
print(metrics(m))
# {'cape': 40.0, 'forward_pe': 21.5, 'trailing_pe': 26.9,
#  'forward_earnings_yield_%': 4.65, 'cape_earnings_yield_%': 2.5,
#  'fed_model_gap_%': 0.49, 'cape_yield_minus_real_yield_%': 0.7}

Reading the output:

• The Fed model gap is small but positive: forward earnings yield slightly above the nominal Treasury. On that view stocks look "fairly priced".
• The CAPE yield minus real yield is also small. A 2.5% real earnings yield against an illustrative 1.8% real bond yield leaves little extra compensation for equity risk, on a backward-looking earnings base.
• Same market, two framings, two different stories. That is the whole AI bubble debate in two lines of arithmetic.

To use real data: Shiller's dataset provides monthly real prices and earnings back to 1871, FactSet and S&P publish forward EPS, and Damodaran posts his implied ERP monthly. Replace the illustrative inputs, and keep the date next to every number.

9. Common mistakes in the debate

• Quoting CAPE as a timing signal. It was above average for most of the last three decades. A high CAPE says lower long-run expected returns, not "sell now".
• Using forward P/E as if forecasts were facts. If AI earnings estimates are revised down, the forward P/E rises without the price moving.
• Comparing a nominal bond yield with a real earnings yield. Use real yields, or use the implied ERP.
• Ignoring concentration. An index-level multiple is a weighted average of very different businesses. Look at the distribution, not just the mean.
• Mixing data sources. Operating vs. GAAP earnings, Wilshire vs. other market cap series, calendar vs. forward 12 months. Each choice moves the number.
• Forgetting what a high valuation needs to be justified. Every multiple embeds an assumption about growth and risk. Write that assumption down and ask whether you believe it.

Why this matters

• Different metrics answer different questions. Most disagreements about "bubble or not" are people answering different questions with different ratios.
• Valuation is about expected returns, not dates. High multiples lower the long-run return you should plan for. They say very little about the next quarter.
• Rates matter, but compare correctly. Real with real, nominal with nominal, or use a model that handles both.
• Concentration changes the reading. In late 2025, the index multiple reflects a small group of AI-exposed giants as much as the broad economy.
• Every number needs a date and a source. Valuation readings move every day; an undated "CAPE is 40" is half a fact.

Final Thoughts

In late November 2025 the metrics broadly agree on one thing: US large caps are expensive by historical standards. CAPE is close to 40, the forward P/E is above its 5- and 10-year averages, and the implied equity risk premium has narrowed during the year. They disagree on how expensive, mainly because they make different assumptions about earnings, accounting and interest rates.

None of them says what happens next month. What they do support is a planning stance: assume lower long-run returns from today's starting point than from cheaper ones, understand which assumptions your portfolio depends on, and treat any single valuation number as a starting point for questions rather than an answer.

Sources

• John Y. Campbell and Robert J. Shiller, "Valuation Ratios and the Long-Run Stock Market Outlook", Journal of Portfolio Management 24(2), 1998, and the 2001 NBER update: P/E and dividend ratios forecast long-run price changes, not earnings or dividend growth.
• Robert Shiller, online data: monthly S&P price, earnings and CAPE series since 1871.
• NPR, Is an AI bubble brewing? Shiller PE ratio nears levels seen before dot-com crash (November 13, 2025): John Campbell quotes on CAPE "pretty close to 40", low subsequent 10-year returns, and no crash timing.
• FactSet, Earnings Insight, November 21, 2025: forward P/E 21.5, 5-year average 20.0, 10-year average 18.7, sector P/Es, trailing P/E 26.9.
• Aswath Damodaran, Data Update 2 for 2025: The Party Continued for US Equities (January 17, 2025): implied ERP 4.33%, expected return 8.91%, 10-year rate 4.58%, the 2% ERP at the end of 1999.
• Aswath Damodaran, Damodaran Online home page, as posted in October 2025: implied ERP of 3.77% on October 1, 2025, with a 4.16% Treasury rate.
• Warren Buffett and Carol Loomis, "Warren Buffett on the Stock Market", Fortune, December 10, 2001 (see the Buffett indicator summary): market value to GNP as "probably the best single measure of where valuations stand at any given moment."
• Cliff Asness, Fight the Fed Model (Journal of Portfolio Management, 2003): real vs. nominal flaw; descriptive, not predictive.
• Cliff Asness, An Old Friend: The Stock Market's Shiller P/E (AQR, November 2012): CAPE and 10-year returns, "not much help to market timers".
• Jeremy J. Siegel, The Shiller CAPE Ratio: A New Look (Financial Analysts Journal 72(3), 2016): accounting changes bias CAPE upward; NIPA earnings as an alternative.