以動態機率模型預測台股熊市之發生

Abstract

有別於過去國內文獻在預測台股熊市時普遍使用的靜態機率模型，本文係利用動態機率模型輔以多種不同屬性的預測變數，以期提升對台股熊市預測的準確性。在Nyberg (2010)及Kauppi and Saikkonen(2008)的研究結果已證實動態機率模型不論是在樣本內或樣本外，其預測能力優於靜態機率模型。本文同樣利用動態機率模型對台股熊市進行實證，最後也驗證了此一論點，惟動態模型勝出的程度會因單變數或多變數模型的採用而有不同。再者，以往文獻慣以使用國內總體經濟指標（例如利率、匯率、通膨率、貨幣與財政政策指標等）作為熊市機率的預測變數，本文則是額外探討了一些股市活動指標和國際總經變數與商品價格指標，並從中發現美國長短天期利差與M&A件數在短期（1個月）到長期（12個月）均具有顯著的台股熊市預測力。而大盤報酬率、股市本益比變動、股市股利殖利率變動、美股道瓊工業指數變動、投資投機等級債券利差、MSCI新興市場指數變動等變數對短天期（3個月）內發生之熊市具預測力。另外，分析師評等則是在較長天期（6到12個月）具有預測力。本文利用以上結果所建構出的動態多變數機率模型成功提升了對台股熊市的解釋能力。最後，本文也運用機率模型設計出一套簡易的市場擇時交易策略，並在歷史數據實證下得出優於大盤的平均月報酬。但若要顯著優於大盤，則運用預測模型得出的熊市機率來調整無風險性資產與股票間的比重配置會比較有機會實現高於大盤的報酬率。不過，動態模型相較於靜態模型，在此交易績效實證下並沒有帶來顯著較高的報酬率。

Different from the most commonly used static model for bear market forecasting in Taiwan, this paper employed dynamic probit model with multivariate predictive variables in order to enhance the predictability of bear markets of the TAIEX. Empirical research in Nyberg (2010) and Kauppi and Saikkonen (2008) have already proved that the predictive power of dynamic probit models is higher than static probit model, regardless of in-sample or out-of-sample results. After our examinations targeting Taiwan stock market, we reached the same conclusions as they did, though the extent of superiority of dynamic models depends on whether a univariate or multivariate model is adopted. Moreover, besides those domestic macroeconomic variables, which have been fully discussed in market forecasting thesis, this paper considered more diversified variables such as stock market activity related indicators and international macroeconomic & other commodity price indicators. We found that US term spread and the number of M&A are good predictors in 1 month to 12 months forecasting scope, while TAIEX historical return, the change of TAIEX PER, the change of TAIEX dividend yield, the change of Dow Jones index, yield spread between investment grade-high yield bonds, and the change of MSCI EM index proved to be significant in short term bear market forecasting. As for half-to-one year forecasting period, ratings from equity research analysts would be a better indicator. Our design of dynamic probit models embedded with multivariate predictive indicators successfully improve the explanatory power for bear market forecasting if comparing with traditional static univariate predicting models. Finally, we developed a market timing trading strategy base on our optimal predicting models and tested it with historical data, revealing an average monthly return that beat the passive buy-and-hold strategy. The results could be more pronounced once a bear market probability weighted asset allocation rule is followed. However, we should be noted that dynamic models failed to outperform static models in providing higher return under this trading test.