學習式向量量化演算法預測房市泡沫

Abstract

本研究目的為建立一個房市泡沫的預警系統，使用長時間的總體經濟歷史資料，並使用學習向量量化演算法來建構預測的模型，以提供政策制定者或投資人作為參考的依據，若預測的投票率有逐步上升或是過高時，便能考慮在泡沫發生前，提早實施總體審慎政策。 首先，本文使用18個已開發國家自1870年起的總體經濟歷史資料，並使用雙重泡沫認定法來認定泡沫發生，分別為房價快速上升與房價向下修正，而本文主要以房價快速上升所認定的泡沫作為被解釋變數，解釋變數則是選用29個總體經濟指標的年資料。第二，在重複一百次模型訓練後，得到預測泡沫發生的投票率，並使用向後特徵淘汰法與向前特徵選取法來挑選解釋變數，得到最適的17個解釋變數。第三，評估LVQ模型的模型表現，以及比較傳統的Logit模型，實證發現LVQ的模型表現較Logit模型優秀。最後，以台灣近十年的房市為例，使用18個國家的數據來建構預測模型，得到台灣發生泡沫的預測投票率，其中在2020年的投票率為近十年中最高，顯示此時的房市相較其他時期熱絡。

This research establishes an early warning system for housing bubbles, using long-term macroeconomic history data and applying the learning vector quantitative algorithm to construct a prediction model as a reference for politicians, policymakers, and investors. If the predicted vote rate gradually increases or is too high, then a macroprudential policy can be implemented early before the bubble appears. This paper first presents macroeconomic historical data of 18 developed countries since 1870 and sets up the double bubble identification method to identify the occurrence of bubbles - namely, the rapid rise in housing prices and their subsequent downward revision. Bubbles represent the dependent variables, and the independent variables are annual data of 29 macroeconomic indicators. Second, after repeating the model training 100 times, we are able to obtain the vote rate of predicting a bubble’s occurrence. Backward feature elimination and forward feature selection help select the independent variables, in which we list the 17 most suitable independent variables. Third, we evaluate the performance of the LVQ model and that of the traditional Logit model and empirically find that the former is better than the latter. Finally, taking the housing market in Taiwan in the past ten years as an example and utilizing data from 18 countries to construct a forecast model to obtain the predicted vote rate of a domestic bubble, results show that the vote rate of a bubble in the year 2020 is the highest over the past ten years, indicating that the housing market at that time was relatively hotter than in other periods.