Statement
In US real estate price indices (residential repeat-sales, commercial transaction-based, and to a lesser degree REIT indices), log price changes exhibit positive serial correlation at short to medium horizons (1 to ~30 months) and negative serial correlation (mean reversion) at longer horizons (>3 years). This momentum-then-reversal pattern is one of the most consistently reported empirical regularities in the real estate forecasting literature.
Evidence summary
The Ghysels-Plazzi-Torous-Valkanov (2013) Handbook chapter is the source
that consolidates this stylized fact. Section 3.1 surveys roughly ten
studies dating back to Gau (1984), with the strongest recent statements
in Case-Shiller (1989), Hill-Sirmans-Knight (1999), Schindler (2011), and
Gu (2002). The chapter then runs its own long-horizon regression
r_{t+1:t+T} = α(T) + β(T) r_{t-T+1:t} + ε on the Census, FHFA, TBI, and
REIT indices for 1991:Q2-2010:Q4 (Figure 3) and finds:
- Census median: negative short-run autocorrelation, no significant long-horizon pattern. INCONSISTENT with the claim.
- FHFA repeat-sales: positive and significant β(T) for T up to ~30 months; turns negative but insignificant at longer horizons. CONSISTENT with the claim.
- TBI commercial: not serially correlated at very short horizon; positive and significant for 6-18 months; weak negative reversal at long horizons. CONSISTENT with the claim.
- REIT: positive but insignificant at 1-month; negative drift in β(T) significant after ~30 months. CONSISTENT with the long-run-reversal half of the claim.
The chapter explicitly cautions that “part of [the positive short-run serial correlation] is due to the construction of the index, as discussed in Section 2.1, and partly to market inefficiencies in real estate markets.” That is, the FHFA/TBI positive β(T) estimates conflate two mechanisms:
- Genuine momentum: slow incorporation of information by infrequent transactors and high search costs.
- Construction artifact: repeat-sales weighting and hedonic smoothing induce positive autocorrelation even under a random-walk true price process (Webb 1981; Goetzmann 1992).
The chapter cannot fully separate these in the aggregate data and explicitly does not claim the predictability is exploitable.
Conditions and scope
- Index type: holds for repeat-sales and hedonic indices, not for median-price indices, and is weakest in stock-market-based REIT indices (which behave like small-cap equity).
- Frequency: documented at monthly and quarterly frequencies. At annual or longer frequencies the short-run momentum signal weakens.
- Geography: cross-sectional dispersion is large. Gu (2002) on the CMHPI for all US states finds the sign and magnitude of persistence varies geographically and over time. Crawford-Fratantoni (2003) find past returns explain 6% of variance in Ohio but 75% in California.
- Sample period: most of the cited evidence is from 1970-2010, covering at most one full boom-bust cycle. The chapter does not test whether the pattern holds in subperiods.
- Tradability: the chapter is explicit that even when β(T) is statistically significant, the magnitude is generally not enough to cover ~6% transaction costs. Linneman (1986), Gau (1985), Rayburn-Devaney-Evans (1987), McIntosh-Henderson (1989) all reach this conclusion. Schindler (2011) finds an exception in bubble-prone California markets, but these are sample-conditional.
- Out-of-sample: regime-switching models that nominally capture the pattern in-sample are beaten by ARIMA out-of-sample (Crawford-Fratantoni 2003), suggesting overfitting.
Counter-evidence
- Median-price indices (Census, NAR) show NEGATIVE short-run autocorrelation, not positive. The chapter does not resolve whether this is because median indices remove the construction artifact or because composition shifts dominate the signal.
- Out-of-sample failure: ARIMA beats regime-switching specifications (Crawford-Fratantoni 2003); McIntosh-Henderson (1989) find ARMA models on Dallas-Fort Worth office data have higher MSPE than the unconditional mean.
- Construction artifact: Webb (1981) and Goetzmann (1992) show that repeat-sales returns can have asymptotic autocorrelation approaching -0.5 even under a random-walk true price; the sign of the artifact depends on the timing of sales. So the observed positive autocorrelation in some indices and negative in others may both be construction-driven.
- The largest “tradable” momentum is in bubble markets (LA, Las Vegas, San Diego, San Francisco — Schindler 2011), which raises the concern that the in-sample predictability would not survive an out-of-sample test that included a different bubble or no bubble.
Linked ideas
- other-commercial-real-estate-boom-bust — Duca-Ling (2015) error-correction framework implies short-run cap rate predictability via adjustment to equilibrium, with 40-72% quarterly error correction speed.
- expected-returns-expected-growth-rents-commercial — Plazzi-Torous-Valkanov (2010) establish that cap rates predict returns for apartments, industrial, and retail (but not offices), complementing the momentum/reversal finding with a structural present-value explanation.
Open questions
- Can the construction artifact be separated from the genuine market-microstructure momentum at the index level? The chapter suggests this requires either disaggregated transaction data (which is mostly proprietary) or a structural model of the index-construction error process.
- Is the long-run reversal robust to sample period? Most of the evidence is dominated by the 2000-2010 boom-bust.
- Does the pattern survive a properly-evaluated out-of-sample test? The chapter notes this is the central unresolved question and that most of the literature is in-sample.
- For commercial real estate (the focus of our project), does the short-run momentum hold AFTER controlling for the four property-type-specific cash-flow dynamics (apartments, industrial, office, retail)? The chapter does not run this test directly.