Year: 2026
Type: article
Abstract: This study examines how forecast accuracy varies across horizons for 85 Los Angeles County cities, the county, and California. The data are monthly Zillow Home Value Index values from January 2000 through August 2026. Holt's linear trend, damped exponential smoothing (ETS), autoregressive integrated moving average (ARIMA), and an equally weighted ARIMA and ETS combination are compared with naïve and drift benchmarks. Specifications are selected using data through 2016. Expanding window forecasts are evaluated one, three, six, and twelve months ahead over a 2017–2019 preliminary evaluation and a test period from January 2020 through August 2026. Errors are measured in dollars within locations and by mean absolute scaled error (MASE) across cities. At one month, ARIMA reduces MAE relative to drift by a median of 71.1% across cities. Its advantage declines as the horizon increases. At twelve months, drift has the lowest median error across cities, and ARIMA's MAE is a median of 9.8% higher. Drift also leads in every price and volatility tercile at this horizon, but rankings vary across subperiods. Residual correlation and convergence failures limit claims about ARIMA's fit. Accuracy estimates are retrospective because the data come from a single revised vintage. Naïve and drift benchmarks should remain part of model comparisons at every horizon.
Author: Zheyan Yu
Institution: Pasadena City College
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7235161819
