The Equity Risk Premium Is Not a Number — It Is a Distribution

The Equity Risk Premium Isn't a Number — It's a Distribution

Over the 19 years from May 2007 through August 2026, the US equity risk premium — stocks over T-bills — averaged 12.15% annualized on a rolling 10-year basis. That's nearly triple the 4.33% forward-looking implied premium that Aswath Damodaran derived for the S&P 500 at the start of 2025. Even the worst 10-year window in our sample (6.46%, ending June 2017) exceeded the textbook 5% estimate. By that read, equities have been extraordinarily rewarding.

But here's the catch: 17% of individual calendar years delivered a negative equity risk premium. In 2008, the ERP was -37.7%. In 2022, it was -20.1%. A single year like that erases three years of typical premium accumulation. The equity risk premium is not a fixed 5% that you collect like a coupon. It's a wildly dispersed outcome that sometimes goes deeply negative — and the gap between what you expect and what you realize is the risk.

Educational only, not financial advice. Data: SPY and BIL daily total returns, May 2007–August 2026, via Tiingo (research-service engine).

What the Equity Risk Premium Actually Measures

The equity risk premium (ERP) is the excess return investors demand for holding stocks instead of a risk-free asset — typically T-bills or Treasury bonds. It's the answer to "why would I own equities if I could own a government guarantee?" The premium is supposed to compensate you for bearing equity risk.

Two camps estimate it differently. The historical approach looks backward: take the average excess return of stocks over T-bills across a long window. Damodaran's 1928–2018 data shows an arithmetic average of 7.93% and a geometric average of 6.11% for stocks over T-bills — with a standard error of 2.09%, meaning the estimate's noise is itself larger than many textbook ERPs (Damodaran, 2018). The implied approach looks forward: back out the expected return the market is pricing in from current stock prices and expected cash flows. As of January 2025, Damodaran's implied ERP was 4.33% — derived from an S&P 500 expected return of 8.91% minus the 10-year Treasury yield of 4.58% (Damodaran, Jan 2025). Kroll (formerly Duff & Phelps) lowered its recommended ERP to 5.0% effective June 2024 (Kroll, 2024).

Both approaches produce a single number. That's the problem. The ERP is the difference between two return streams — equities and the risk-free rate — and both are volatile. A difference of volatile quantities is more volatile than either alone.

The Mechanism: Why the ERP Swings So Hard

Think of the ERP as (stock return) − (risk-free return). Both terms move. When stocks surge and T-bill yields stay near zero — as they did from 2009 to 2021 — the realized ERP explodes upward. When stocks crash and T-bills still pay something — as in 2008 or 2022 — the ERP goes sharply negative. When T-bill yields rise fast while stocks stall — the 2023 environment, where BIL returned 4.93% — the risk-free rate eats into the premium even without a stock-market crash.

We see this clearly in the data. The risk-free rate (proxied by BIL, the 1-3 month Treasury ETF) earned 0.17% annualized from 2008 to 2015, then 4.59% annualized from 2023 to mid-2026. That's a 27× increase in the risk-free component. The ERP is not just about how much stocks return — it's about how much they return relative to a moving baseline.

This is why rolling-window estimates vary so dramatically. We computed the realized ERP across three window lengths:

Over 3-year windows, the ERP ranged from -12% to +29% — a 40-percentage-point spread. Over 10-year windows, it narrowed to a still-wide 6.5% to 17.4%. The standard deviation of the 10-year rolling ERP alone was 2.07% — nearly half the size of the entire 4.33% implied premium Damodaran calculated for 2025. The noise in the estimate is almost as large as the estimate itself.

What the Charts Show

The calendar-year view is the sharpest illustration. Three of eighteen years (2008, 2018, 2022) delivered a negative ERP. The magnitudes are brutal: -37.7% in 2008, -7.0% in 2018, -20.1% in 2022. Meanwhile, the best years (2013: +29.1%, 2019: +29.0%, 2021: +30.6%) produced premiums larger than the entire long-run historical average. The ERP doesn't smooth out across years — it clusters at the extremes.

The rolling 10-year chart tells a different story. Every single 10-year window in our sample produced a positive ERP, and every one exceeded Damodaran's 4.33% implied estimate. The minimum was 6.46%; the mean was 12.15%. If you held stocks for any 10-year period between 2007 and 2026, you collected a premium well above what textbooks predict. This is the comfort of long horizons — and the trap of believing the ERP is tame.

Why Realized Exceeds Expected — and Why That Won't Last

The gap between our realized 12.15% rolling 10-year ERP and Damodaran's 4.33% implied estimate is 7.82 percentage points. That's not measurement error. It's structural. Three forces drive it:

1. Valuation expansion. The S&P 500's P/E ratio expanded dramatically over this period. When multiples expand, price returns exceed earnings returns — mechanically inflating the realized premium. But you can only get that lift once. Forward-looking estimates like Damodaran's implied ERP bake in current valuations, not future multiple expansion.

2. Near-zero risk-free rates. From 2008 to 2021, T-bill yields were effectively zero. BIL returned 0.17% annualized from 2008 to 2015. The risk-free rate was barely a hurdle, so almost any positive stock return showed up as premium. When rates normalized in 2023-2024, BIL jumped to 4.59% annualized — and the ERP arithmetic got harder even though stocks kept rising.

3. Survivorship and selection. Our sample uses SPY, which tracks the survivors. Companies that went bankrupt or were delisted don't appear. This is the standard survivorship caveat for any index-based return series (Damodaran, 2026 ERP paper).

None of this means the ERP is "really" 12%. It means the realized premium over a specific bull-market window was 12%, and that window is not the universe of possible outcomes. The 3-year windows that went negative are equally real.

The Estimation Window Is the Bet

Here's the practical problem. If you're building a financial model, you need one ERP number for your discount rate. Which do you pick?

• Historical 1928–2018 geometric: 6.11%. Smooth, but includes a century of regimes that may not repeat.
• Implied (forward-looking): ~4.33% as of Jan 2025. Market-consistent, but assumes current valuations and growth expectations are correct.
• Recent realized (10y rolling): ~12%. Captures the current regime, but that regime included the largest valuation expansion in a generation.

The choice between these isn't a technical detail. It's a bet about which regime you're in. Use the 12% number and you're assuming the next decade looks like the last one. Use the 4.33% number and you're assuming markets are currently pricing the right premium. Use the 6.11% century-long geometric average and you're assuming mean reversion across 90 years of data. Each assumption has a name and a failure mode.

Damodaran himself flagged this in his 2026 ERP paper, arguing that the implied premium — despite requiring forward growth assumptions — is "on more defensible ground" than the earnings yield approach because it at least uses current market prices rather than a historical average that may not reflect today's risk environment (Damodaran, Jan 2025). He also noted that during the 2008 crisis, the implied ERP nearly doubled from 4.2% to over 8% in two months — exactly the kind of time-variation that a fixed historical average would miss (Damodaran, 2026).

Where This Meets Our Work

At RiskHarvest, we don't rely on a single ERP estimate. The reason is exactly what this data shows: the premium is a distribution, not a constant, and the left tail (those -37.7% and -20.1% years) is where portfolios fail. Our approach uses vol-targeted risk sizing — sizing equity exposure to a target volatility rather than a fixed weight — precisely because it dampens the left-tail years that destroy long-run compounding. When realized volatility spikes (as it does in the negative-ERP years), position sizes shrink automatically, reducing the damage. This is the mechanism that turns a 40-percentage-point 3-year ERP spread into a survivable path.

We also apply a premia-decomposition test to every portfolio holding: each asset is classified by the risk premium it actually harvests — equity, term, default, commodity — so we can verify that "diversifiers" are backed by genuinely different premia, not three repackaged versions of the same equity beta. If the realized ERP is a 40-point distribution, you want exposure to premia that don't all draw from the same bad year. The equity risk premium going -37.7% in 2008 while the term premium (long bonds) went positive is not a coincidence — it's the reason premia decomposition matters.

The One Thing to Remember

The equity risk premium is not 5%. It is not 4.33%. It is not 12%. It is a distribution that ranges from -38% to +31% in single years and from +6.5% to +17.4% across 10-year windows. The number you put in your spreadsheet is an assumption about which part of that distribution you expect to land in — and the difference between your assumption and what actually happens is, definitionally, the risk you're being paid to bear.