PaperlayerView exact revision on arXiv

Source-linked AI summary

Forecasting the U.S. Real House Price Index

Vasilios Plakandaras, Rangan Gupta, Periklis Gogas, Theophilos Papadimitriou

arXiv:1707.04868v1q-fin.CP

TL;DR

Forecasting sharp U.S. house-price downturns matters for economic stability. The paper combines EEMD and SVR for forecasting, and reports that EEMD-AR-SVR outperforms competing models and detects the 2006–09 downturn early.

  • Problem

    Forecasting sharp U.S. house-price downturns matters for economic stability.

    Source: §1 · pp. 2, 4

  • Method

    The paper combines Ensemble Empirical Mode Decomposition with Support Vector Regression to construct house-price forecasting models.

    Source: §1 · p. 4

  • Results

    EEMD-AR-SVR outperforms all competing models in in-sample and out-of-sample forecasting and reports early accurate detection of the 2006–09 price downturn.

    Source: §4 · p. 22

  • Takeaways & Limitations

    The authors argue that the proposed forecasting methodology can be used effectively as an early-warning tool for house-price downturns.

    Source: §4 · p. 22

    The EEMD process may lose essential information about the underlying phenomenon.

    Source: §2.1 · p. 5

Abstract

from arXiv · show

The 2006 sudden and immense downturn in U.S. House Prices sparked the 2007 global financial crisis and revived the interest about forecasting such imminent threats for economic stability. In this paper we propose a novel hybrid forecasting methodology that combines the Ensemble Empirical Mode Decomposition (EEMD) from the field of signal processing with the Support Vector Regression (SVR) methodology that originates from machine learning. We test the forecasting ability of the proposed model against a Random Walk (RW) model, a Bayesian Autoregressive and a Bayesian Vector Autoregressive model. The proposed methodology outperforms all the competing models with half the error of the RW model with and without drift in out-of-sample forecasting. Finally, we argue that this new methodology can be used as an early warning system for forecasting sudden house prices drops with direct policy implications.

1. Introduction and literature review

The paper situates house-price forecasting as relevant to economic activity and policy, then proposes a hybrid signal-processing and machine-learning approach. It extends earlier forecasting work with a longer sample, additional macroeconomic variables, and multi-period forecasts aimed at early warning of the 2006–07 downturn.

1. Introduction and literature review

  • House-price fluctuations are presented as indicators of economic activity and inflation, making accurate forecasts potentially useful to market participants and monetary policy authorities.The introduction connects housing construction with a large share of GDP and house prices with the economy’s overall wealth.Source: §1 · p. 2
  • Earlier Spatial BVAR studies forecast the direction of real house prices but significantly underestimated the decline, which the authors attributed to excluding fundamentals.The studies forecast the downturn over 2007:01 to 2008:01 using monthly real house-price growth rates.Source: §1 · p. 3
  • Prior work found in-sample nonlinearity in house-price growth, but virtually no evidence favored nonlinear models in out-of-sample point, interval, or density forecasts.This discussion provides context for testing a different forecasting methodology.Source: §1 · p. 4
  • The proposed EEMD-SVR approach is compared with Bayesian autoregressive and vector autoregressive models for one-year-ahead forecasting, avoiding the endogeneity issues associated with Bayesian predictive regressions.The paper uses Ensemble Empirical Mode Decomposition from signal processing and Support Vector Regression from machine learning.Source: §1 · p. 4
  • Compared with previous studies, the study uses annual observations from 1890 to 2012 and 11 macroeconomic variables, extending evaluation to multi-period-ahead forecasting as a house-price early-warning system focused on the 2006–2007 downturn.Source: §1 · p. 4

2. Methodology

The methodology combines EEMD decomposition and component aggregation with Elastic Net variable selection and SVR forecasting, while specifying VAR and Bayesian benchmark models. It describes how these methods address time-series components and smoothing, regression, and regressor selection.

2.1 Overview

  • The proposed three-step forecasting method applies EEMD smoothing, selects variables with Elastic Net, and fits a Support Vector Regression model.Source: §2.1 · p. 5
  • EEMD separates inputs into low-frequency and fluctuating components, which are forecast separately and summed, with 4-fold cross-validation used during training.Source: §2.1 · pp. 5–6

2.2 VAR and BVAR Specifications

  • The Bayesian VAR and BAR variants use a Minnesota prior and are specified in levels, with the paper stating that this Bayesian approach does not require stationarity.Source: §2.2 · pp. 7–8

2.3 Ensemble Empirical Mode Decomposition and time series smoothing

  • EEMD decomposes a time series into intrinsic mode functions through iteration until the residual has no local extrema, selecting added Gaussian-noise amplitude by a maximized relative RMSE criterion.Source: §2.3 · p. 8
  • The authors extend prior trend-extraction work by evaluating all IMFs to construct representations of both short-run dynamics and a long-run trend function.Source: §2.3 · p. 9
  • The smoothed series sums IMFs from a selected cutoff onward with the final EEMD residual; the example sums the 3rd, 4th, and 5th IMFs.Selecting the cutoff separates short- from long-term dynamics.Source: §2.3 · p. 10

2.4 Support Vector Regression (SVR)

  • SVR fits a regression function using an error-tolerance band, penalizing deviations beyond ε while balancing fit and generalization through ε and C.The model is trained on most of the dataset and tested on a held-out subset to evaluate generalization.Source: §2.4 · pp. 11–12
  • Kernel functions enable nonlinear SVR by representing projections through dot products in the original space; simulations use linear, RBF, sigmoid, and polynomial kernels.The kernel trick avoids explicitly computing each data point’s mapping.Source: §2.4 · p. 13

2.5 Elastic Net

  • Elastic Net combines LASSO and ridge regression with a regularization parameter ranging from zero to one, spanning the two as special cases.It can adapt to data characteristics and select regressors for regression models.Source: §2.5 · pp. 13–14

3 The Dataset

The dataset pairs a long annual U.S. housing-price history with macroeconomic predictors, using 1890–1988 for in-sample estimation and 1989–2012 out of sample.

3 The Dataset

  • The annual sample spans 1890–2012, with 1890–1988 in sample and 1989–2012 out of sample.The authors use the longer history to assess forecasting over prolonged periods and changes in house-price trends.Source: §3 · p. 14
  • The authors argue that a short sample may capture only episodic trends, whereas longer periods reveal forecasting performance across prolonged periods and house-price trend changes.The out-of-sample period begins in 1989 to cover a period of sharp housing-market activity.Source: §3 · p. 14
  • The dataset contains 11 annual U.S. macroeconomic variables, including real house prices, fiscal policy, GDP per capita, unemployment, interest rates, inflation, population, construction costs, stock prices, and oil prices.The variables were selected based on the literature.Source: §3 · pp. 14–15
  • Real house prices reached historically high levels during 2006–2009, followed by a sharp drop of unprecedented magnitude.Figure 5 plots the Schiller housing price index (1890=100); shaded areas mark NBER recession periods.Source: §3 · p. 15, Figure 5 · p. 16

3. Empirical Results

The empirical results assess EEMD-AR-SVR’s out-of-sample forecasting performance for house-price downturns across forecasting horizons and discuss the potential significance of anticipating price declines as indicators of economic contraction. They also consider how such forecasts could inform monetary policy adjustments.

3. Empirical Results

  • Out of sample, EEMD-AR-SVR outperformed all models on MAPE and DS, while EEMD-EN-SVR’s MAPE doubled and was worse than all except BVAR.The table reports one-period-ahead static forecasts.Source: Table 1 · p. 18, §3 · p. 18
  • In sample, EEMD-AR-SVR had the highest accuracy by MAPE, closely followed by EEMD-EN-SVR, which was most accurate by DS.The comparison uses the two reported criteria, MAPE and Directional Symmetry (DS).Source: Table 1 · p. 18, §3 · p. 18
  • BAR and BVAR forecast less accurately than the Random Walk in- and out-of-sample; EEMD-AR-SVR’s outperformance of RW using lagged prices led the authors to reject weak-form efficiency.The authors state that historical values of the real house price index alone adequately forecast it under EEMD-AR-SVR.Source: §3 · pp. 18–19
  • EEMD-AR-SVR alone among five models timely forecast the 2006–2009 downturn; the other three models captured it a year late, while EEMD-EN-SVR performed poorly, possibly from overfitting.The paper also says EEMD-AR-SVR tracks actual prices closely.Source: Figure 6 · p. 19, §3 · pp. 19–20
  • Across horizons, EEMD-AR-SVR beat RW on both criteria in- and out-of-sample for years 1–4, and out-of-sample through year 6; DS outperformance extended to year 9, but RW won both criteria in year 10.In-sample MAPE also favored EEMD-AR-SVR at all ten horizons.Source: Table 2 · p. 20, §3 · p. 21
  • EEMD-AR-SVR forecast the housing-price collapse up to two years ahead, nearly two years before it occurred.The passage associates the forecast with the significant 2006–2009 fall in U.S. housing prices.Source: Figure 8 · p. 21, §3 · p. 22
  • The 2006 house-price drop signaled an economic contraction followed by the 2007–2008 recession.The passage describes house-price movements as an indicator of the business-cycle phase, referring to Figure 7 and Leamer (2007).Source: §3 · p. 22
  • Falling house prices in 1989–1990 marked a contraction before the brief 1991 recession.Source: §3 · p. 22
  • Anticipating house-price deterioration may indicate an approaching contraction and support monetary-policy adjustment.Source: §3 · p. 22

4. Conclusion

The conclusion presents EEMD-AR-SVR as a competitive framework for real U.S. house-price forecasting, including single- and multi-period horizons and early detection of the 2006–2009 downturn. It argues that the methodology may also serve as a policy instrument for determining the economy’s business-cycle phase.

  • The study examines eleven predictors for the real U.S. House Prices index and compares BAR and BVAR techniques with EEMD-SVR.Source: §4 · p. 22
  • EEMD-AR-SVR outperforms all competing models in both in-sample and out-of-sample forecasting.Source: §4 · p. 22
  • The framework covers single- and multi-period-ahead forecasting and reports early accurate detection of the 2006–2009 price downturn.Source: §4 · p. 22
  • The authors argue the methodology can be used effectively as a policy instrument for determining the economy’s business-cycle phase.Source: §4 · p. 22
Loading 1707.04868v1…