Fisher Hypothesis

Definition

The Fisher hypothesis states that expected nominal asset returns move one-for-one with expected inflation, leaving expected real returns independent of expected inflation:

$E_t[R_{t+N}^{(i)}] = r^* + E_t[\pi_{t+N}]$

Assets representing claims to real payments (stocks) should therefore hedge inflation, while assets representing claims to nominal payments (bonds) should not. The hypothesis is an implication of classical economic theory and is closely related to the Fisher equation for nominal interest rates.

Key Ideas

The short-horizon failure: most empirical studies, from Fama-Schwert (1977) onward, find that short-horizon stock returns are negatively correlated with inflation — a stark contradiction of the Fisher hypothesis. This is sometimes attributed to a "proxy effect" (stock returns proxy for real activity, which is negatively correlated with inflation).
Boudoukh-Richardson (BR, 1993) long-horizon claim: using generalized method of moments (GMM) with Newey-West standard errors on time-overlapping returns, BR find US and UK stocks at the 5-year horizon are positively correlated with inflation — Fisher model improves with horizon. Taken as evidence that the hypothesis holds in the long run.
Overlapping data problem: when the return horizon $N > 1$ , the multi-period returns share $N-1$ observations, inducing a moving-average $\mathrm{MA}(N-1)$ component in regression residuals. Newey-West and GMM corrections become unreliable when the overlap fraction $N/T$ is large (Richardson-Stock 1989; Hodrick 1992). This limits BR to $N \leq 5$ years.
Vector autoregression (VAR) approach (Engsted-Tanggaard 2000): compute multi-period expected returns directly from the one-period VAR companion form — no overlap, no Newey-West correction needed, horizon can be extended to $N=10$ years.
Country heterogeneity: US and Danish results diverge sharply, showing that the Fisher relationship is highly sensitive to country-specific monetary regimes and return dynamics.

How It Works

Testing Framework

The basic Fisher regression (ex post form): $R_{t \to t+N} = \alpha + \beta \pi_{t \to t+N} + \epsilon_t$ Under the hypothesis $\beta = 1$ . Problems: (i) errors-in-variables bias (actual inflation = expected + surprise), biasing $\hat\beta$ toward zero asymptotically; (ii) time-overlapping residuals when $N > 1$ .

GMM Approach (Boudoukh-Richardson 1993)

Use a vector of instruments $Z_t$ known at time $t$ (past returns, inflation, dividend yield, interest rate changes). The orthogonality conditions:

$E[(R_{t \to t+N} - \alpha - \beta \pi_{t \to t+N}) \otimes Z_t] = 0$

yield consistent estimates. A J-test of overidentifying restrictions tests whether instruments are uncorrelated with the ex ante real return. Limitation: standard errors remain unreliable at large N due to overlap.

VAR Approach (Engsted-Tanggaard 2000)

Estimate a first-order VAR in mean-corrected variables: $X_t = AX_{t-1} + w_t$ , where $X_t = [R^s_t,\, R^b_t,\, \pi_t,\, \delta_t,\, \Delta i_t]'$ (stock return, bond return, inflation, log dividend-price ratio, change in short rate).

Multi-period expected returns and inflation are linear in the current state:

$\hat{R}^{(i)}_{t,N} \equiv E\!\left(\sum_{j=1}^N R^{(i)}_{t+j} \,\middle|\, H_t\right) = \mathbf{e}_i' \left(\sum_{j=1}^N A^j\right) X_t$

$\hat\pi_{t,N} \equiv E\!\left(\sum_{j=1}^N \pi_{t+j} \,\middle|\, H_t\right) = \mathbf{e}_\pi' \left(\sum_{j=1}^N A^j\right) X_t$

These involve no time-overlapping observations. Fisher hypothesis evaluation: report Corr( $\hat{R}^{(i)}_{t,N}$ , $\hat\pi_{t,N}$ ) and $\sigma(\hat{R}^{(i)}_{t,N})/\sigma(\hat\pi_{t,N})$ ; both should equal unity if Fisher holds exactly.

Unexpected inflation: $\hat{u}_{t,N} = \pi_{t,t+N} - \hat\pi_{t,N}$ . Inflation hedging assessed by regressing ex post returns on $\hat{u}_{t,N}$ .

Main Results (Engsted-Tanggaard 2000)

Correlation between expected N-period returns and expected N-period inflation (standard errors from 1000 bootstrap replications):

Asset	Country	N=1	N=5	N=10
Stocks	US	0.47	0.52	0.38
Bonds	US	0.004	0.70	0.73
Stocks	Denmark	0.33	0.78	0.89
Bonds	Denmark	0.29	0.13	0.06

US stocks contradict Boudoukh-Richardson: Fisher model does not improve with horizon; correlation weakens at 10 years.
US bonds strongly support Fisher at long horizons.
Danish stocks provide near-perfect Fisher at 10 years (corr $\approx 0.89$ , SD ratio $\approx 1$ ).
Danish bonds: Fisher fails and worsens with horizon.
No asset in either country provides a meaningful hedge against unexpected inflation ( $R^2 \approx 0$ in hedging regressions).

Structural VAR (SVAR) Fisher Decomposition (St-Amant 1996)

St-Amant (1996) exploits the long-run Fisher relationship as a structural identification restriction rather than testing it. The identifying assumption is that the ex ante real interest rate is stationary — meaning ex ante real rate shocks have no permanent effect on nominal rates — while inflation expectation shocks do have permanent effects. This is the Blanchard-Quah (1989) long-run restriction strategy applied to a bivariate system $(\Delta i_t, r_t)'$ where $r_t = i_{t,k} - \pi_t$ .

Findings for U.S. 1-year and 10-year bond rates (Feb 1957 – Jun 1995):

At short horizons, real rate shocks dominate the 10-year rate variance (~75% at 1 month) but split roughly 50/50 for the 1-year rate.
By 48 months, both rates are ~80% explained by inflation expectations.
Inflation expectation shocks build gradually (slow regime adjustment); real rate shocks die out within ~2 years.
1970s–early 1980s rate rise: almost entirely inflation expectations.
1994–95 rate fluctuations: almost entirely ex ante real rate movements, consistent with Michigan survey comparison (inflation expectations stable throughout).

The approach has three advantages over survey-based methods: it is based on market prices, requires no constant real rate assumption, and can be applied to countries without indexed bond markets.

Why It Matters

The Fisher hypothesis is a cornerstone of monetary economics — if it holds, inflation is a veil over real returns and monetary policy has no real long-run effects on asset markets.
Conflicting empirical results (short-horizon failure vs. long-horizon partial success) make the Fisher hypothesis one of the unsettled empirical regularities in finance.
Country-level heterogeneity in the Fisher relationship complicates global asset allocation: the same asset class (stocks) can be a Fisher hedge in one country but not another at the same horizon.
The overlapping-data critique applies broadly to long-horizon tests (stock predictability, uncovered interest parity (UIP), term structure) — the VAR companion-form approach is a general remedy.

Open Questions

Why do US stocks fail the Fisher model at long horizons when Danish stocks succeed? Potential explanations: different dividend-price dynamics, different monetary regimes, different inflation processes, or a sample artifact.
The proxy effect (Fama 1977) and the Mundell-Tobin effect offer competing explanations for the short-horizon negative stock-inflation correlation; neither is definitively settled.
Whether the Fisher hypothesis holds better in low-inflation vs. high-inflation regimes remains an open empirical question.
The paper does not test real versus nominal returns separately; real return variation is conflated with inflation variation in the analysis.